Downloaded from www.annualreviews.org. Guest (guest) IP: 74.109.18.71 On: Sat, 17 May 2025 11:55:57
Annual Review of Immunology
Systems Human Immunology
and AI: Immune Setpoint and
Immune Health
Yona Lei1 and John S. Tsang1,2,3
1
Yale Center for Systems and Engineering Immunology and Department of Immunobiology,
Yale University School of Medicine, New Haven, Connecticut, USA; email: john.tsang@yale.edu
2
Department of Biomedical Engineering, Yale University, New Haven, Connecticut, USA
3
Chan Zuckerberg Biohub NY, New Haven, Connecticut, USA
Annu. Rev. Immunol. 2025. 43:693–722
Keywords
The Annual Review of Immunology is online at
immunol.annualreviews.org
systems immunology, immune variation, immune setpoint, machine
learning, artificial intelligence, AI, immune health, human immunology
https://doi.org/10.1146/annurev-immunol-090122042631
Copyright © 2025 by the author(s). This work is
licensed under a Creative Commons Attribution 4.0
International License, which permits unrestricted
use, distribution, and reproduction in any medium,
provided the original author and source are credited.
See credit lines of images or other third-party
material in this article for license information.
Abstract
The immune system, critical for human health and implicated in many
diseases, defends against pathogens, monitors physiological stress, and maintains tissue and organismal homeostasis. It exhibits substantial variability
both within and across individuals and populations. Recent technological and conceptual progress in systems human immunology has provided
predictive insights that link personal immune states to intervention responses and disease susceptibilities. Artificial intelligence (AI), particularly
machine learning (ML), has emerged as a powerful tool for analyzing complex immune data sets, revealing hidden patterns across biological scales, and
enabling predictive models for individualistic immune responses and potentially personalized interventions. This review highlights recent advances in
deciphering human immune variation and predicting outcomes, particularly
through the concepts of immune setpoint, immune health, and use of the
immune system as a window for measuring health. We also provide a brief
history of AI; review ML modeling approaches, including their applications
in systems human immunology; and explore the potential of AI to develop
predictive models and personal immune state embeddings to detect early
signs of disease, forecast responses to interventions, and guide personalized
health strategies.
693
1. INTRODUCTION
Downloaded from www.annualreviews.org. Guest (guest) IP: 74.109.18.71 On: Sat, 17 May 2025 11:55:57
The immune system plays critical roles in the health and function of all organs throughout the
life span; its dysfunction has been implicated in most if not all human diseases. In addition to defending against pathogens and tumors, the immune system acts as a sensor of physiologic stress,
monitoring and integrating both internal and environmental signals to mount appropriate responses to maintain homeostasis (1–3). The immune system thus provides an informative window
into an individual’s health and physiologic state.
Technological advances over the past two decades, including single-cell multiomics, proteomics, and immune cell phenotyping, have opened new avenues for interrogating the immune
system at unprecedented levels of depth and resolution (4). When applied to the most readily
accessible human tissues, such as blood, skin, and airway, these technologies provide a wealth of information about an individual’s immune state in health and disease (1). Variability is a fundamental
hallmark of human immunity. Diverse factors sculpt each individual’s unique immunological states
(5), driving variable responses to pathogens, vaccines, and treatments to result in a wide spectrum
of health outcomes. This heterogeneity can pose challenges; for instance, vaccines may not be
universally effective (6). However, population variation is essential to empower the development
of predictive models of immune system behavior (5, 7), which can enable the identification of intervention targets (7, 8). Thus, understanding and harnessing human immune variation are central
to human immunology and are crucial for developing and optimizing personalized interventions
and outcomes.
Modern artificial intelligence (AI) and particularly machine learning (ML) have recently
emerged as powerful tools for deciphering complex biological systems (9, 10). As high-throughput
technologies continue to yield a deluge of immunological data across biological scales, ML
approaches offer the capability to extract biologically meaningful relationships, reduce data dimensions, build predictive models, and define biomarkers and potential mechanistic determinants
of immune statuses and health outcomes (11). The ability of ML models to compress highdimensional data such as those capturing immune system states (with many configurations) into
smaller representations (or embeddings) is key to deciphering and harnessing human immune
variation and building powerful predictive models. Not only can these compressed representations
enhance predictive capacity, but also they can facilitate the discovery and mechanistic modeling of
biological players and patterns not immediately apparent in higher dimensions (11, 12). Thus,
AI may provide both a unifying conceptual paradigm—decoding molecular and cellular logic
within and across biological scales (Figure 1)—and powerful tools to enable systems human immunology, whose mission is not merely to transform ever-larger data sets into biological insights
and provide tools for precision medicine but also to achieve a quantitative, mechanistic understanding of the human immune system together with its interaction with physiology as a holistic
system.
In this review, we explore the landscape of systems human immunology and AI, focusing on
recent advances in investigating human immune variation and the concepts of immune setpoints
and immune health, which are critical anchors linking personal immune states and functional
outcomes. We then provide a brief history of AI/ML and an overview of predictive modeling
approaches. Finally, we offer our perspective on how emerging technologies and AI may drive
progress toward the systems human immunology vision, including the exciting prospect of leveraging the immune system as a sensor of health by developing predictive models of immune
setpoints and immune health. These models could detect signs of disease long before symptom onset (13–15), predict responses to interventions (7, 16, 17), and chart future health trajectories. Such
694
Lei • Tsang
Multiscale nature
of human immunity
X
(of any dimension)
Mapping
function F
Y
(of any dimension)
Molecule
(DNA, proteins, RNA)
Downloaded from www.annualreviews.org. Guest (guest) IP: 74.109.18.71 On: Sat, 17 May 2025 11:55:57
Cell
(type, state, composition)
Tissue
(structure, state, function)
Structure
predictor
(e.g., protein sequence)
(structure, function,
gene expression)
Organ
Cell–cell
interaction
predictor
System
Individual/organism
Population
(e.g., gene expression)
(cell dynamics,
trajectory, responses)
Figure 1
Artificial intelligence (AI) framework for decoding multiscale immune system complexity. Given any input data, machine learning (ML)
algorithms can automatically learn a mapping function that transforms input X of any dimension into output Y of any dimension. ML
offers a unifying framework for uncovering and mapping complex input–output relationships in the immune system. The human
immune system orchestrates defenses and maintains homeostasis across multiple levels of biological organization, ranging from
molecular and cellular interactions to tissue and organismal responses and, ultimately, to population-level phenomena. Information
processed at one level of the immune system can be integrated to shape responses in and across other levels and bodily systems. ML
approaches have been developed to map input–output relationships and extract biological insights both within and across these scales.
For example, at the molecular level, ML algorithms have been developed to map protein sequence (input) to three-dimensional
structure (output), assess variant effects on protein function, and predict gene expression patterns from sequence data. More generally,
such an ML framework can be employed to effectively map various input features to functions, including the mapping of gene
expression programs to cell types, dynamics, and functional responses at the cellular level, as well as the mapping of personal immune
states to functional responses at the organismal level. Thus, AI/ML-driven approaches facilitate the integration and analysis of diverse
data types, enabling predictions of immune outcomes across different scales. Such a comprehensive framework fosters a more holistic,
predictive, and mechanistic understanding of human immunity and holds the potential to advance precision medicine.
predictive capacity could empower precise interventions to prevent diseases, enhance responses,
and guide lifestyle choices.
2. HUMAN IMMUNE VARIATION
Human immunity exhibits extensive variability, manifesting both across individuals and within a
single person over time (5, 18). Examples and consequences of public health importance abound,
including experimental malaria vaccines that showed high protective efficacy when evaluated in
the United States but whose efficacy dropped significantly when tested in Africa (19). Reduction
in immunogenicity has also been reported for other vaccines, including bacille Calmette-Guérin
(BCG) (20, 21) and yellow fever (22), although these observations do not suggest that individuals
in Africa have less robust immune responses; for example, Africa as a continent appeared to have
fared better in COVID-19 outcomes in comparison to the United States and Europe (23). While
these differences could be attributed to many factors, variation in individual immune statuses
and responses plays a crucial role. More broadly, variability is pervasive in diverse contexts,
including responses to cancer immunotherapy (24, 25) and other therapeutics (26, 27), as well as
the magnitude and frequency of disease activities (or flares) observed in autoimmune conditions
such as lupus (28, 29).
www.annualreviews.org • Systems Human Immunology and AI
695
2.1. Intrinsic Variation and Heritability
Downloaded from www.annualreviews.org. Guest (guest) IP: 74.109.18.71 On: Sat, 17 May 2025 11:55:57
Immune variation is driven by a complex interplay among heritable and nonheritable factors,
which together shape molecular, cellular, and tissue states in an individual to influence physiological functions and responses (18, 30) (Figure 2a). Genome-wide association and twin studies
have revealed genetic drivers of immune parameters such as circulating immune cell frequencies and their disease associations (31, 32). The extent of genetic contribution depends on the
immune parameter and type of response. Twin studies suggest that antibody responses to early
childhood vaccines are more heritable than those in adults (33, 34); presumably, as genetically
identical twins age, their immune systems diverge due to nonheritable contributions from cumulative environmental exposures. Indeed, immune cell frequencies, circulating cytokine levels,
and antibody responses to seasonal influenza vaccination are largely nonheritable in adults (35),
although certain cell phenotypes seem highly heritable (36). Certain changes in these immune
parameters likely involve the epigenome, which exhibits plasticity in response to environmental stimuli. Epigenetic modifications can alter gene expression programs in immune cells, thus
influencing cellular phenotypes and the regulatory networks that regulate immune responses. In
general, it can be challenging to ascertain whether epigenetic mechanisms are causal for or merely
correlated with responses, although another twin study showed that 70% of the chromatin modification variations can be explained by nonheritable factors (37), suggesting that environmental
exposures together with other variables (e.g., sex) can mediate their effects on immune functions
through epigenetic mechanisms. Thus, nonheritable factors are key determinants of a person’s
immune state and function.
Variables intrinsic to a person, such as sex, can shape immune states and functions. Sex differences have been observed in susceptibility to infection and autoimmune disease, vaccination
responses, and treatment outcomes, which could be consequences of sex dimorphic innate and
adaptive immune parameters and responses (38, 39). For example, females tend to have elevated
early inflammatory responses and subsequently increased immunogenicity to vaccination (39). At
the cellular level, the frequency and activity of immune cells can differ between the sexes. Males
typically have higher natural killer cell frequencies (40), whereas females tend to have more B cells
that perhaps explain their greater antibody responses to vaccination regardless of age (38, 40).
Men with elevated testosterone levels exhibit an immunosuppressive transcriptional signature enriched for genes involved in lipid biosynthesis and have blunted influenza vaccine responses (41).
The health consequences of these sex differences, including increased prevalence of autoimmunity
and organ rejection posttransplantation in females, can be profound (38).
2.2. Development, Aging, and Cumulative Exposures
The immune system undergoes extensive development and environment-driven transformation
throughout an individual’s life span (42, 43). Thus, many immune parameters and functions are
linked to age. The immune system of infants and children is different from that of adults, although much remains to be learned regarding immune variation within and across individuals
as a function of age. Early life exposures sensed by the immune system are thought to be key
determinants of health trajectory later in life, including susceptibility to immune-mediated diseases and infections (43). Interestingly, a longitudinal, system-level analysis of peripheral blood
mononuclear cells from preterm and term human newborns revealed that, despite large initial
differences, their immune states converged toward a shared developmental trajectory at 3 months
of age (44). However, this robust, stereotypical immune system development during early life can
be hampered by gut bacterial dysbiosis, highlighting the importance of the microbiome in shaping
the development of the immune system and immune variation. The observation that the newborn
696
Lei • Tsang
a
Sex
Age
Genetics
Intrinsic
Personal
immune states
Variable immune responses due to
differences in personal immune states
Parameter
Perturbation
Geography
Downloaded from www.annualreviews.org. Guest (guest) IP: 74.109.18.71 On: Sat, 17 May 2025 11:55:57
Composite
Microbiome
Intersubject
variation
Response 1
Response 2
Within-subject
variation
Vaccination
Infection
Organ/tissue
networks
Time
Extrinsic
b
Cellular
networks
Immune profiling
and multiomics
Molecular
networks
AI
Low
c
Healthy
Response
High
Subclinical/
Symptomatic
presymptomatic
disease
(hidden pathology)
Aging
CVDs,
musculoskeletal disorders,
neurodegeneration
Autoimmunity, cancer, etc.
Figure 2
Integrating multimodal immune profiling and artificial intelligence (AI) to investigate human immune variation and predict immune
responses and immune health. (a) Personal immune states, or setpoints, are dynamically shaped by intrinsic (e.g., genetics, sex), extrinsic
(e.g., microbiome, vaccination, infection), and composite (e.g., age, geographical location) factors, which often reflect contributions
from multiple sources (e.g., age reflects development, senescence, and exposure). Given differences in these diverse factors, immune
states and responses vary widely across individuals. (b) Immune states can be assessed through multimodal profiling of accessible clinical
samples from sources such as blood, skin, and nasopharyngeal tissues. The resulting large, high-dimensional data capture the statuses of
immune components and networks spanning tissue, cellular, and molecular levels. Given the large number of variables (e.g., more than
a trillion immune cells in the human body; 93) and the interactions among them, the number of configurations (or immune states) is
vast. AI, including deep learning, can decipher the complex correlation structure of such data and compress this vast configuration space
into lower-dimensional embeddings that capture essential relationships among key immune entities. Such representations of personal
immune states will not only elucidate the potential mechanistic drivers of immune variation but also enhance our ability to make
accurate, personalized predictions of immune responses to perturbations—such as vaccination, infection, disease, and therapeutics—as
well as health trajectories. (c) Predictors or setpoints of immune responses can serve as a proxy for measurements of immune system
health, enabling us to chart individuals along a spectrum of immune health (color bar). As shown recently (15), such immune health
metrics could track healthy aging and capture subtle signals (e.g., deviation from healthy homeostasis) to report on “hidden” pathology
in clinically healthy individuals before overt symptoms manifest. Such early detection is particularly critical for combating chronic
diseases, which often progress silently for years. The use of the immune system as a sensor for identifying physiological deviations and
disease processes in their incipient and early stages (progression is shown by the thick arrows, going from healthy to less healthy to
overtly pathological) opens a window for early intervention that could alter disease trajectories and improve long-term health
outcomes. Abbreviation: CVD, cardiovascular disease.
www.annualreviews.org • Systems Human Immunology and AI
697
Downloaded from www.annualreviews.org. Guest (guest) IP: 74.109.18.71 On: Sat, 17 May 2025 11:55:57
immune state varies widely but could converge later in time suggests that the immune system is a
versatile sensory system that can buffer variability in inputs to robustly generate the desired output, potentially through feedback mechanisms. A genetically hardwired system would presumably
be evolutionarily less advantageous because it would be less adaptable to diverse immunological
challenges. However, given the large number of immune parameters and the much larger combinatorial space of immune states, this early convergence in the frequency and phenotype of certain
peripheral immune cell populations does not imply that variation across individuals is minimized.
We have yet to fully unravel the complexities of pediatric immune development and how individuality is established, maintained, and shaped over time. Indeed, poorly understood phenomena
important for public health, such as early-life imprinting of antigen-specific immunity (45, 46),
variable durability of childhood vaccine–induced immunity (47), and antigen-agnostic effects of
live vaccines such as BCG (48, 49), underscore the value of achieving a quantitative and predictive
understanding of early life and pediatric immune variation.
After reproductive age, a process often called immunosenescence drives immune aging and the
related concept of inflammaging (50–53). Immunosenescence is characterized by systemic chronic,
low-grade inflammation (51, 53), impaired T cell activation and metabolism (54), decreased B cell
frequencies, and defects in class-switch recombination and somatic hypermutation (55). These
features are linked to the decreased magnitude and quality of antibody responses in the elderly
(56, 57), increased infection susceptibility and mortality, and heightened risk of diseases such as
cancer (58–60). In a comprehensive longitudinal immune profiling of older versus younger adults,
Alpert et al. (61) showed that not only do older individuals have distinct and more rapid dynamical changes as they age but also their immune state predicts the future rate of change of certain
immune aging variables. These analyses thus revealed an aging trajectory that allowed an immune
age metric to be derived, and this metric could predict all-cause mortality in an independent cohort. The concept of an immune aging metric (62) is reminiscent of the desire to pinpoint an
individual’s “biological” instead of chronological age to assess health and predict future trajectory (63–67). In fact, our recent multimodal analyses of 22 monogenic immune-mediated diseases
(15) suggest that the immune changes associated with healthy aging are concordant with those observed in diverse human pathologies. Thus, deviations from health in diverse conditions, including
aging, have shared systemic immune consequences (see Section 3.3, below).
The impact of age on immunity further emphasizes that life histories and environmental
exposures (or variables extrinsic to the individual), including infections, the microbiome, and variables associated with geographic location, cause immune variation. Systems immune profiling of
670 healthy individuals aged 2 to 86 years identified cohabitation as the largest contributor of
interindividual variation in immune cell composition, with 50% less variation between individuals with a shared environment (68). This study found that sex-based differences at the molecular
level were more pronounced than those at the cellular level (68, 69), suggesting a complex interplay
whereby variation in molecular parameters can be buffered by compensatory or feedback mechanisms that result in lower variability at the cellular level, although the functional implications are
not yet clear.
Chronic or latent infection status can also modulate immune states and responses (5). For
instance, cytomegalovirus seropositivity is associated with enhanced antibody responses to influenza vaccination in young, healthy individuals, but it has no effect on older individuals who
exhibit decreased responses independent of their cytomegalovirus status (70). The role of the
microbiome has also come to light (71, 72). A trial administered broad-spectrum antibiotics to
healthy adults followed by seasonal influenza vaccination and systems immunology analyses (73).
Surprisingly, despite extensive microbiome loss, these individuals appeared capable of mounting
normal responses without diminishing antibody outputs. Mildly impaired antibody responses were
698
Lei • Tsang
Downloaded from www.annualreviews.org. Guest (guest) IP: 74.109.18.71 On: Sat, 17 May 2025 11:55:57
observed in individuals with low baseline levels of neutralizing antibodies. Furthermore, antibiotic
treatment resulted in enhanced proinflammatory signatures in circulating immune cells before
vaccination, resembling the early transcriptional responses to influenza vaccination observed earlier in the elderly (73, 74). While it remains to be determined whether preexisting antibodies are
causal, these observations highlight the potential vulnerability of populations with lower preexisting immunity, including infants and older adults, to mount inadequate vaccine responses after
microbiome perturbation.
2.3. Geography as a Composite Variable
Geographic location is a complex, composite variable in that it reflects contributions from myriad
factors including ancestry/genetics, diet, and pathogen load (75). To date, systems human immunology studies investigating immune variation have been conducted largely in adults from a
few high-income countries such as the United States. Thus, their findings may not generalize
to other regions, a limitation that is particularly relevant given that infectious diseases disproportionately affect individuals in low- and middle-income countries (76). To address this issue,
Hill et al. (77) conducted longitudinal immune phenotyping of peripheral blood mononuclear
cells from children participating in a large malaria vaccine (RTS,S) trial conducted in Tanzania
and Mozambique. A comparison between African and Dutch children revealed that while certain
immune cells follow a similar developmental trajectory, the dynamic changes in immune cell frequency can vary between countries, even those in close geographic proximity, such as Tanzania
and Mozambique. Tanzanian children had fewer circulating T follicular helper cells and plasmablasts at baseline, which were linked to lower IgG titers following RTS,S vaccination. Anemia
emerged as a potential driver of these immune differences, with higher prevalence in Tanzanian
children. Anemia was associated with reduced B cell responses, including lower levels of IgG memory B cells and plasmablasts following vaccination, implying that iron deficiency negatively affects
vaccine responsiveness (77). One limitation of this study is that both the sample and data of the
Dutch cohort were generated independently of the RTS,S malaria vaccine trial in Africa, making
it difficult to attribute the observed differences solely to geographic variation. More broadly, the
RTS,S vaccine trial was conducted in seven African countries, and vaccine efficacy varied across
study sites (78–80). Similarly, attenuated live malaria sporozoite vaccines have variable efficacy
across populations, particularly between Africa (low protection) and the United States/Germany
(high protection) (19, 81, 82), underscoring the importance of understanding human immune variation across populations to pursue goals such as reversing vaccine hyporesponsiveness (75). Much
larger cohorts spanning variables such as geographic region, age, and sex, together with coordinated human immune profiling and AI model building, are needed to unravel the determinants
and circuitry underlying human immune variation around the globe (83).
3. IMMUNE SETPOINT AND IMMUNE HEALTH
Immune variations span different types and manifest at diverse spatial and timescales. As discussed above, together they can be conceptualized as intrinsic (e.g., genetics, sex), extrinsic (e.g.,
exposure, season), or composite variables that together shape personal immune states (e.g., molecular, cellular, and tissue statuses and their interactions), which in turn determine physiologic and
functional outputs (e.g., response to vaccination, infection, disease) (Figure 2a). Linking personal
immune states to functional outputs is the concept of immune setpoints, which correspond to immune states of an individual that determine (and thus predict) specific functional output quality
and quantity. The setpoint concept originated from control theory, where a setpoint is a variable whose target (or homeostatic) value is maintained by a control system. An example is the
www.annualreviews.org • Systems Human Immunology and AI
699
Downloaded from www.annualreviews.org. Guest (guest) IP: 74.109.18.71 On: Sat, 17 May 2025 11:55:57
set ambient temperature of a room regulated by an air-conditioning system. A setpoint can be
multidimensional (e.g., temperature, pressure), and as exemplified by temperature, the setpoint
shapes the environment in which activities and processes occur. As defined, different individuals
in the human population can have different setpoint states at a given point in time, thus providing
different molecular, cellular, and tissue environments within individuals in which responses are
orchestrated to generate functional outputs (e.g., antibody and cytotoxic T cell responses to challenges) (5). A personal immune setpoint can be continuously shaped by genetic, environmental,
and physiological signals (5, 28, 84–86).
The complexity of the individual immune setpoint is further shaped by the remarkable diversity of T and B cell receptor (TCR and BCR) repertoires. Monozygotic twins, despite sharing
identical genetics, exhibit distinct antigen receptor repertoires due to the stochastic nature of
V(D)J recombination and differences in antigen-driven selection processes (87). This receptor diversity contributes significantly to variations in immune responses between individuals, including
differential responses not only to vaccines and pathogens but potentially to other diseases such
as autoimmunity and neurodegeneration (88–90). The repertoire continues to evolve throughout
life, influenced by aging and cumulative antigen exposure, further modulating the immune setpoint (91, 92). Thus, the theoretical configuration space of human immune states is vast, given
that there are approximately two trillion immune cells in the human body (93) and each cell
can exist in diverse functional states, including variable antigen recognition receptors in B and
T lymphocytes—not to mention interactions among these cells and with other cell types throughout the body (Figure 2a,b). However, the actual state space is likely much smaller due to functional
and evolutionary constraints, and the setpoint state space should have even lower dimensionality
given that the variables that shape and predict specific responses are even more constrained.
3.1. Poised and Natural Adjuvanted Setpoint
Baseline setpoint predictors have been reported for vaccination responses (7, 28, 94–98), therapy outcome (24, 99), and disease activity or progression (13, 28). An emerging concept in the
context of vaccine responses is that baseline setpoints can resemble innate immune signatures induced by vaccine adjuvants. Individuals possessing such “natural adjuvanted” baseline setpoints
tend to have more robust responses to vaccination. Our recent multimodal single-cell analysis
of two influenza vaccines—one formulated with the adjuvant AS03 and one without an adjuvant
(85)—demonstrated this concept, revealing that high responders to the unadjuvanted vaccine exhibit baseline states mirroring those specifically elicited by AS03 1 day after vaccination in myeloid
dendritic cells and CD14+ monocytes. Further investigation found that CD14+ monocytes from
high responders at baseline displayed elevated phosphosignaling responses to in vitro lipopolysaccharide (LPS) stimulation, reflecting an increased cell-intrinsic response capacity. Together, these
observations suggest that the naturally adjuvanted setpoint may drive more robust vaccine responses by enhancing innate immune cell sensing capacity. While other baseline predictors of
vaccine responses have not been compared explicitly with early immune responses elicited by
adjuvants, these baseline setpoints also tend to mark inflammation and innate immune cell activation, suggesting that a more “poised” immune state is a common feature of baseline setpoints.
However, the details can differ, and the underlying mechanism likely involves not only a single
cell type or cell-intrinsic process but also a network of interacting cell states. For example, we
found a baseline setpoint involving activation of plasmacytoid dendritic cells as well as both B
and T lymphocytes; this baseline setpoint predicted not only antibody responses to influenza and
yellow fever vaccination but also plasma cell–associated flares in lupus (28). Despite the low abundance of plasmacytoid dendritic cells in blood, this setpoint can be measured via as few as 10 blood
700
Lei • Tsang
Downloaded from www.annualreviews.org. Guest (guest) IP: 74.109.18.71 On: Sat, 17 May 2025 11:55:57
transcripts thanks to the correlated cell states: Signals from the more abundant B and T lymphocytes in blood, including but not limited to their interferon response gene signatures, are sufficient
to capture the state of the correlated cell network. This observation also suggests that setpoints for
autoimmunity (encompassing disease status, activity, endotype) and other immune response outcomes, including those to infection and vaccination, could be shared (100, 101). Understanding
what perturbations could change and modulate these setpoints to influence disease risk and activity could help unravel the association between infection and autoimmunity (e.g., Epstein–Barr
virus and multiple sclerosis; 102) and identify drug targets to mitigate autoimmune disease activity.
Uncovering natural adjuvanted setpoints may also guide the development of vaccine adjuvants.
3.2. Setting the Setpoint
Analyses of blood from healthy individuals over weeks to years (68, 95, 103) suggest that immune
cell frequencies and serum protein levels, including baseline setpoints predictive of influenza vaccination (95), are relatively stable, although these variables can exhibit temporal trends during
pediatric development or aging later in life (42, 61, 104, 105). In general, aside from antigenspecific or bystander memory lymphocytes, most if not all of the acute challenge–induced immune
changes are expected to revert to their prechallenge baseline state (18), a process governed by
tightly regulated cellular circuitry tailored to each person’s unique physiological context. The
degree of temporal stability varies across immune parameters, with some components uniquely
shaped by prior immunological challenges or exhibiting more pronounced fluctuations than others. Furthermore, immune setpoints can be modulated. BCG vaccination and fungal infections can
change chromatin states in monocytes and myeloid progenitor cells to confer antigen-nonspecific
effects on future responses, a phenomenon known as trained innate immunity (48). The induced
nonspecific, protective effects are thought to occur through modification of epigenetic states
in progenitor cells, including those reflected by histone marks such as H3K4me1, H3K4me3,
and H3K27ac (48, 106). Severe COVID-19 can induce similar long-term changes to circulating
monocytes and hematopoietic stem and progenitor cells (HSPCs), including increased activities
in transcription factors like interferon regulatory factors, hyperresponsiveness to inflammatory
stimulations, metabolic reprogramming, and increased inflammatory cytokine production (107,
108). Since circulating monocytes have average half-lives of only a few days (109), longer-term
memory can reside in the HSPCs that give rise to monocytes (110). The wider impacts of such
long-term alterations of HSPCs—such as whether and how they could limit the production
of normal, nontrained monocytes and other HSPC progeny beyond monocytes—remain to be
determined.
Longitudinal multiomics analyses of healthy individuals who recovered earlier from mild
COVID-19 and were not hospitalized revealed that male recoverees exhibited a more poised
baseline immune state associated with stronger innate and adaptive responses to subsequent influenza vaccination (84). This male-specific setpoint state involved multiple interacting immune
cell populations, including virtual memory-like CD8+ T cells and classical monocytes, that form
a mutually reinforcing positive feedback circuit via IFN-γ and IL-15. This finding suggests that
prior viral infections could establish new, sex-dependent setpoints with the potential to shape future responses in an antigen-agnostic manner. Similarly, analyses of human responses to H5N1
vaccination with or without the AS03 adjuvant illustrated that even an antigen-sparing, unadjuvanted vaccine alone can shift the immune system to a new baseline setpoint associated with
enhanced responses (111). Upon AS03-adjuvanted H5N1 influenza vaccination, classical monocytes can also attain a refractory state associated with a global reduction in histone acetylation and
chromatin accessibility, which persisted for at least a month after vaccination. These monocytes
www.annualreviews.org • Systems Human Immunology and AI
701
Downloaded from www.annualreviews.org. Guest (guest) IP: 74.109.18.71 On: Sat, 17 May 2025 11:55:57
had a heightened antiviral program and were more resistant to viral infections (112). These studies
suggest that infection, vaccination, or any immunological challenge can alter immune setpoints,
thereby potentially modulating future responses. Along with better knowledge of the causal mechanisms underlying baseline immune setpoints, these studies could reveal actionable intervention
targets, thus raising the prospect of not only predicting outcomes but also modulating the baseline setpoint to enhance subsequent immune responses (7). As revealed by a recent drug screen
in mouse bone marrow–derived macrophages, even noninflammatory (classically thought of as
anti-inflammatory) agents such as glucocorticoids can induce trained immunity phenotypes (113).
Other promising immune setpoint modulators include drugs targeting epigenetic regulators (e.g.,
histone deacetylase inhibitors), which have shown potential in treating hematologic and solid tumors (114) and altering innate immune training (115), as well as microbiota-targeted strategies
that could promote the efficacy of immune checkpoint inhibitors (116, 117).
3.3. Setpoint and Immune Health: Probing Hidden Health Deviations
The concept of an immune setpoint is closely related to the idea of personal immune health.
For example, predictors of responsiveness to vaccines and therapeutics, as well as resistance to
infections and severe disease, naturally give rise to intuitive measures of immune health and competence (Figure 2c). More broadly, defining and measuring immunological health are still in their
infancy. One approach is the concept of immunological resilience based on age-independent adaptation to antigenic stimulation; in other words, immune systems that are best able to return to
homeostasis after antigenic stimulation such as infection are the most immunologically healthy
(118). When challenged with inflammatory stressors, individuals who resist the degradation of
immunological resilience tend to have lower risks of disease acquisition. Another approach is to
define immunological health by learning common features of diverse pathologies. We pursued
this idea by integrating multiomics and ML to simultaneously assess 22 monogenic immunemediated diseases (15). Despite the penetrance and relatively large effect of monogenic mutations,
interindividual variation in diverse immunological parameters remains the dominant type of variation, significantly above those attributable to age, disease condition, and medication use. Another
unexpected observation was that unsupervised ML—not knowing who has what disease or any
other phenotypic attributes, but solely utilizing individual-level multiomics data—generated a
one-dimensional metric capable of distinguishing between sick patients and age- and sex-matched
healthy controls. This same metric emerged with the use of data from sick patients alone (or from
healthy controls only, albeit less robustly), suggesting that it captures a robust and potentially
meaningful measure of immune states. Indeed, this immune health metric (IHM) was concordant with the probability of being healthy derived using supervised ML distinguishing disease
from health. In independent data sets, the IHM marks healthy aging; measures biological age declines; forecasts age-dependent decreases in vaccination responses; and predicts deviations from
health, including autoimmunity, heart failure, and poor body mass index. Given that immune cells
surveil the body and can detect signals of stress and deviation from physiologic homeostasis in
tissues, the IHM could perhaps capture such signals to report on “hidden” pathology. Thus, as
we observed earlier (15), even clinically healthy individuals can be distributed over a wide range
along the immune health axis, from very healthy to healthy, to inching toward pathology but still
without overt signs of disease, and finally to having overt disease (Figure 2c). Many diseases, including Alzheimer’s and autoimmunity, as well as aging-associated declines, take years to develop
into full-blown symptomatic conditions (119–124). Further development of the IHM concept
and the identification of additional dimensions of immune health could enable early detection
and intervention to revert early health deviation and functional dysregulation.
702
Lei • Tsang
Downloaded from www.annualreviews.org. Guest (guest) IP: 74.109.18.71 On: Sat, 17 May 2025 11:55:57
3.4. Tissue Setpoint
Thus far, our discussion of human immune variability and setpoints has focused on blood-based
assessments reflective of peripheral (or systemic) immune states. The nearly two trillion immune
cells in the average human body reside and function largely in tissues, particularly the bone marrow, the lymphatic system, and barrier sites (93, 125, 126). Thus, understanding tissue setpoints
and deciphering the connection between blood and tissue immune states are major research goals.
Certain immune cell populations are rarely found in blood under normal physiological conditions;
these include tissue-resident macrophages that develop directly from embryonic progenitor cells
and certain memory T cells that establish long-term residency in tissues. The extent to which
tissue immune states can be captured in blood by cells, proteins, metabolites, and so forth is an
active area of investigation. For instance, some circulating protein components of the IHM might
have originated from tissues. However, many tissue-specific responses remain difficult to detect,
as circulating immune parameters often poorly reflect the immune activity present in tissues. For
example, longitudinal immune profiling of paired blood and airway samples from patients with severe COVID-19 showed that T cell frequencies in the airway, but not the blood, predict outcomes
in patients with severe COVID-19 (127).
Tissue setpoints are determined by the intrinsic properties and states of tissue components,
which determine how tissues respond to perturbations. While variable outcomes of identical perturbations are commonly observed, we are only beginning to understand how preexisting tissue
states influence these responses. Studies of mucosal immune responses, such as those of the upper
airway, have provided illuminating examples. When clinically healthy, nonsmoking adults were
experimentally challenged with respiratory syncytial virus (RSV), inflammatory activation of neutrophils in the nasal mucosa before virus exposure predicted symptomatic RSV disease (128). This
baseline setpoint state was associated with the suppression of early mucosal immune responses,
particularly IL-17 and related pathways, during the early presymptomatic phase of the response.
In contrast, individuals who resisted infection exhibited presymptomatic activation of IL-17- and
tumor necrosis factor–related pathways and were able to mount an early IL-17-driven inflammatory response that might have controlled and blocked viral infection. Along similar conceptual
lines, another study showed that children had higher basal levels of innate immune genes, including viral sensors, in upper airway epithelial cells, macrophages, and dendritic cells (129). This
finding may explain why young children are often better protected against SARS-CoV-2 symptomatic infection: Presumably they are better primed for virus sensing, leading to stronger innate
activation upon infection (129).
The broader tissue environment can be shaped by previous experiences or exposures, creating a
kind of tissue memory that influences future responses. For example, after acute skin inflammation,
epithelial stem cells retain inflammatory memory that accelerates barrier restoration during subsequent tissue damage (130). This memory is mechanistically maintained through sustained changes
in chromatin accessibility at key stress response genes activated by the initial stimulus. Genes
governed by these persisting accessible regions are transcribed rapidly in response to secondary
stressors. While these alterations mobilize stem cells more rapidly and are often associated with
accelerated wound repair, they are also associated with increased susceptibility to autoimmune
disorders and cancer. The concept of tissue memory has also been extended to nonimmune cell
progenitors in other tissues, including intestine, respiratory tract, and pancreas, with implications
for infection, tumorigenesis, and chronic inflammatory diseases (131–134).
Similar to immune setpoints in blood, tissue setpoints can be shaped by intrinsic, extrinsic, and
composite factors and underlie interindividual variation in response to immune perturbations.
For example, individuals with chronic respiratory and allergic diseases are at increased risk of
www.annualreviews.org • Systems Human Immunology and AI
703
severe outcomes from respiratory viral infections. One study demonstrated that nasal challenge
with a Toll-like receptor 7/8 agonist led to heightened interferon and chemokine responses in
nasal mucosa in individuals with allergic rhinitis (with or without asthma) compared with healthy
participants (135). This finding suggests that baseline inflammation and immune dysregulation
shape tissue-specific immune responses and thereby influence the susceptibility to and outcome
of respiratory viral infections.
4. PROBING HUMAN IMMUNE SYSTEM STATES
Downloaded from www.annualreviews.org. Guest (guest) IP: 74.109.18.71 On: Sat, 17 May 2025 11:55:57
4.1. Perturb, Then Measure
Natural immune variation, together with perturbations such as vaccination followed by highdimensional immune profiling and statistical and ML analyses, has been key to empowering the
discovery of predictive immune setpoints in humans (1, 5, 7). In addition to being an effective
medical intervention, vaccines can serve as a perturbation to probe the human immune system,
allowing assessment of whether baseline immune variation across individuals can lead to differences in functional outputs (Figure 2a). This approach has uncovered setpoints that predict and
potentially determine outcomes (7). Vaccines naturally elicit in vivo responses linked to function,
including early innate and adaptive responses (e.g., days 1, 3, and 7 following influenza vaccination), as well as longer-term impacts on the immune system, such as the magnitude and durability
of antibody, memory T cell, and trained innate immunity responses (5, 48). Vaccines have also been
used to probe patients to assess how disease-associated immune changes may influence responses;
such in vivo perturbations, together with in vitro cellular stimulation assays (see Section 4.3), can
reveal molecular and cellular defects that become apparent only when evaluating responses to perturbations (84, 85, 136–138). Therapeutics such as checkpoint blockades, cytokines, and cellular
therapies are additional perturbations that can be applied in appropriate populations to identify
immune setpoints.
4.2. Time: A Critical Dimension
As mentioned above, the immune system is inherently dynamic, and immune parameters can
change over time at various scales. These changes range from highly stable parameters (i.e., those
whose temporal variations are significantly less than differences across individuals) to transient
changes to long-lasting shifts, potentially influencing future immune responses or having no
discernible impact—effects that are often difficult for cross-sectional snapshots alone to reveal.
Longitudinal studies with thoughtful study designs are thus desirable for capturing the temporal dimension of immune states and identifying robust predictors and setpoints of functionally
relevant outcomes. A key design decision is the temporal sampling frequency, which is often constrained by budget and logistics. By taking repeated measurements over time, one can observe how
each person’s immune parameters evolve with or without perturbation, allowing for the identification of temporally stable versus varying traits that may be population or individual specific, as well
as temporal changes that could be the key drivers of immune variation observed in cross-sectional
studies. Denser sampling over hours to days can help unveil response dynamics including specific
molecular, pathway, and cellular parameters, as highlighted by early response dynamics to vaccination (111) and neonatal development during the first week of life (139). Sampling in the range of
weeks to months can provide both innate and adaptive response dynamics after a perturbation (5,
140) and reveal temporally stable parameters when there are no apparent perturbations. For example, setpoint parameters predictive of vaccine responses revealed by our earlier research tended
to be temporally stable within healthy individuals for more than 2 months (28, 95). Unexpectedly, longitudinal sampling of patients with different monogenic diseases (with samples collected
704
Lei • Tsang
Downloaded from www.annualreviews.org. Guest (guest) IP: 74.109.18.71 On: Sat, 17 May 2025 11:55:57
months to 2 years apart) suggested that many blood parameters—including transcriptional profiles, cell frequencies, and circulating protein levels—are stable, varying significantly less over time
within a person compared with the extent of differences observed across individuals. Variations
attributable to individuals are also substantially higher than those explained by monogenic disease
labels, suggesting that individual immune state is the dominant unit of variation.
The study of development, aging, seasonality, and so forth naturally requires sampling at
timescales ranging from months to years. In addition, the temporal sequence captured in longitudinal data provides an opportunity to disentangle causal relationships and infer the directionality
of effects (8, 141, 142). Strategic cohort design can address specific facets of immune system behavior; for example, household and twin studies can help distinguish environmental and genetic
effects on immune states and functions (31, 35, 36, 68, 143, 144). However, large cohorts encompassing various demographics are essential in order to capture and utilize a broad spectrum of
setpoints and influencing factors. These cohorts not only ensure proper representation of underrepresented populations but also leverage natural population variation to increase the power for
novel discoveries (5, 7, 18).
4.3. Sample Type, Multimodality, and Single-Cell Analysis
Modern multiomics and immune-profiling technologies allow comprehensive measurements of
human immune system states and behavior (Figure 2b). Together, these approaches can perform
measurements at different biological scales, from the genome and epigenome to cells, tissues, and
physiologic phenotypes. Blood is the most common sample type from which circulating molecules
such as cytokines and immune cells can be interrogated, and similar data can be obtained from
other body fluids, including saliva and urine. While integrative analyses of blood and tissue samples remain sparse at the cohort level, they could provide novel insights into cell trafficking and the
extent of mirroring of immune statuses between blood and tissues (145–147). Biopsies from accessible tissues such as skin and gut are becoming more common (125, 148–150), but longitudinal
sampling of children and clinically healthy individuals remains more challenging. More specialized samples such as draining lymph node aspirates or biopsies (151–155) enable the evaluation of
germinal center reactions and other processes in secondary lymphoid organs; swabs from mucosal
sites such as the nasopharynx have proven informative for dissecting viral infection responses and
protective neutrophil and memory T cell setpoints (128, 156, 157). Recent data suggest that cellfree DNA/RNA from blood could report on tissue status (158, 159); in the future, advances in
cellular recording technologies could enable remote sensing of tissue status in blood and other
body fluids (160–162).
While sample type and amount can limit downstream measurements, modern multiplexed
technologies require relatively small inputs to generate large data sets, as exemplified by targeted
proteomic technologies like SomaScan (163–165) and Olink (166, 167). Advances in mass spectrometry and enrichment approaches have also enabled both targeted and nontargeted profiling
of proteins and metabolites (1, 168–171). The measurement of cells is a pillar of immune profiling. It can involve assessing the frequency and composition of immune cells as well as their
molecular profiles, activation and proliferation status, effector functions, antigen specificities, and
cytokine production. The surge of single-cell omics has enhanced our ability to resolve cell type–
specific behaviors and states with remarkable granularity and resolution. These methods generate
extensive quantities of multimodal data to interrogate multiple aspects of cellular immunity.
Single-cell multiomics technologies are continually evolving to increase both throughput and
modalities that can be profiled in a single experiment, enabling the simultaneous profiling of the
transcriptome, proteome, epigenome, metabolome, and so on (4). A notable example is cellular
www.annualreviews.org • Systems Human Immunology and AI
705
Downloaded from www.annualreviews.org. Guest (guest) IP: 74.109.18.71 On: Sat, 17 May 2025 11:55:57
indexing of transcriptomes and epitopes by sequencing (CITE-seq), which combines multiplexed
surface protein detection via nucleic acid–based barcodes with transcriptome profiling (172)
and, optionally, V(D)J sequencing to simultaneously capture the variable receptor and antibody
repertoires in T and B cells. Surface protein expression profiles allow the use of decades of immunological knowledge and data to annotate immune cells and states pioneered by flow cytometry
(4, 173, 174). CITE-seq has also been extended to the detection of specific TCRs and BCRs via
barcoded antigens (175, 176) and to the labeling of intracellular proteins and their posttranslational modifications to assess signaling status and transcription factor activities (177). In principle,
there is no limit to the number of targets that can be detected, but both cost and the availability
of off-the-shelf barcoded antibody reagents are practical constraints. Additional modalities, including chromatin accessibility, have been added to CITE-seq or single-cell RNA sequencing to
enhance the multimodal characterization of cells (178, 179).
While costs have come down dramatically, they continue to be a major limitation in the application of single-cell multiomics to longitudinal human cohort studies. Thus, the number of cells that
can be profiled per sample is still relatively low compared with that of technologies such as modern
high-dimensional mass or flow cytometry, which remain staples of the human immune-profiling
toolkit as they can measure up to ∼50 parameters in tens of thousands of cells per sample with high
throughput (180, 181). Moreover, mapping of the interactions between single cells within tissues
offers insights into cell–cell interactions and the role of the microenvironment in shaping cellular
behavior. Single-cell omics methods typically require the dissociation of cells from their native tissue, resulting in the loss of spatial information. Spatial technologies bridge this gap by enabling the
simultaneous profiling of multiple modalities down to the single-cell level while maintaining the
spatial coordinates of cells within intact tissues (4, 182). While multimodal experimental methods
such as CITE-seq enable direct linking of multiple modalities to the same individual cells, numerous single-cell data sets are unpaired, with different modalities profiled from cells of the same
sample or tissue but not from the exact same single cells (183). Emerging tools are beginning to
address data set heterogeneity and reconcile the differences between omics-specific feature spaces
(173, 184).
In addition to analyses of cells and molecules, in vitro stimulation of cryopreserved immune
cells (185–187) or whole-blood cultures (188, 189) followed by signaling and transcriptional response measurements provides another layer of information. Stimulations include cytokines such
as interferons, TCR signaling activation triggers (e.g., anti-CD3/CD28), and ligands for pattern
recognition receptors such as LPS. When single-cell omics or flow/mass cytometry is used, responses can be measured at different timescales and down to single-cell resolution. Signaling
responses such as phosphorylation states are typically measured at early time points to measure
cell-intrinsic responsiveness, which can reveal baseline immune setpoints associated with aging
(103), pediatric development (190), and vaccine responses (85). Transcriptional responses measured at later time points can be more challenging to interpret, given the influence of other cells
in the culture and upstream responses in the same cell. Secreted molecules such as cytokines can
also be interrogated by analyzing cell culture supernatant. As discussed above with regard to naturally adjuvanted baseline setpoints (85), comparison of in vitro stimulation responses of cells from
the baseline with in vivo innate responses elicited by vaccination can lead to biological insights.
5. AI AND SYSTEMS HUMAN IMMUNOLOGY
The sheer volume, high dimensionality, and multimodality of immune-profiling data (e.g., capturing variations across individuals over time with high molecular and cellular resolution) require
advanced AI/ML approaches to extract biological insights (10, 12).
706
Lei • Tsang
Downloaded from www.annualreviews.org. Guest (guest) IP: 74.109.18.71 On: Sat, 17 May 2025 11:55:57
5.1. A Brief History of AI
While the field of AI has no definitive beginning, one early milestone was in 1950, when computing pioneer Alan Turing proposed a test to evaluate whether a computer exhibits behavior
indistinguishable from that of humans (191). John McCarthy, a mathematician and another AI pioneer, coined the term AI in 1955, when he was preparing for the first conference in AI (192). For
many years, AI focused on knowledge representation and ways for computers to mimic how humans think, resulting in probabilistic and logic-based representations of the world and associated
algorithms (193). Another branch of AI, often called pattern recognition, focused on approaches
such as neural networks (NNs) and learning algorithms, including classic statistical methods like
linear regression and decision trees (see Section 5.2). This branch evolved into what is now called
ML (Figure 1), which aims to learn mathematical functions that map a vector of numbers (inputs)
into another vector of numbers (output).
5.2. A Brief Introduction to Machine Learning
Uncovering immune setpoints and potential determinants of functional outcomes can be framed
as an ML/predictive modeling problem, in which the inputs are the data on molecular and cellular
immune states of a person and the goal is to develop ML models that use the input data to predict
individual outcomes (Figure 1). For example, given the transcriptional state of T cell subsets in
blood before anti-PD-1 treatment, a model could predict whether the patient will respond to the
treatment. ML models can be divided into interpretable versus less interpretable or black-box
models; the main difference is that the key parameters contributing to the prediction can readily
be extracted in the former, while it is more challenging to do so in the latter (Figure 3a).
Examples of interpretable models include linear regression and random forest models; in general, NN and deep neural network (DNN) models are less interpretable. An NN is a class of
functions that mimics how the brain computes. The simplest unit, called a perceptron, multiplies
each input value by a weight (which is learned from the training data), sums up these weighted values, and computes the output by basically evaluating whether the weighted sum is above or below
a certain threshold (196). While there are variations in how individual units compute and many
different architectures are possible, an NN essentially comprises many perceptron units where
the output of one unit feeds into the input of other units. Units are often organized in layers,
and a DNN contains many layers; learning in such deep networks is computationally demanding.
However, such networks are remarkably expressive (or capable), allowing them to learn to perform
extremely complex data transformations, such as translating text between two natural languages.
NN-based approaches have revolutionized biology in recent years, yielding breakthroughs in
protein structure prediction (197), artificial protein design (198), and prediction of pathologic genetic mutations (199, 200) (Figure 1). Thanks to the abundance of single-cell data, DNN models
of cells including immune cells are emerging (201, 202), even though their predictive capacity
and generalizability are still being evaluated (203). In general, given the complexity of the human
immune system and the variability in the human population, a hurdle AI still faces is the relative
dearth of data covering the vast space of personal immune and molecular recognition states in
individuals and diverse populations.
In addition to interpretability, ML algorithms and associated models can be classified as either
supervised or unsupervised (204). In supervised learning, the algorithm is trained on labeled data,
where inputs are paired with their ground-truth outputs. It learns to map the input features to the
correct output labels, allowing the model to make predictions on new data based on the learned
patterns. A key goal is to achieve generalizability, which refers to the ability of the resulting model
to predict the output with good accuracy and precision when given unseen input (12). By contrast,
www.annualreviews.org • Systems Human Immunology and AI
707
a
Cluster Cell annotation
Results
Data
Downloaded from www.annualreviews.org. Guest (guest) IP: 74.109.18.71 On: Sat, 17 May 2025 11:55:57
AI
1
...
...
...
...
n
b
CD4 central memory
...
...
...
...
Classical monocytes
c
Explanation
Low-dimensional
representation
Why? How?
Reconstructed features
to assess quality of
representation
Input
features
ML model for
specific tasks
Figure 3
Interpretable machine learning (ML) approaches for high-dimensional immune data analysis. (a) ML models can be classified into
explainable versus black-box models. ML models are trained to map input data to output; the reasoning and path to reaching decisions
or computing the output remain largely hidden in black-box models but can be fully or partly retrieved in explainable or interpretable
models, such as random forest classification or regression models. In the context of cell type prediction using single-cell transcriptomic
data, a black-box model might accurately predict that a cell belongs to a specific immune cell type or subset on the basis of its gene or
protein expression patterns, but without revealing which specific expression patterns inform this classification. Such a lack of
interpretability makes it challenging both to verify whether the model’s decision-making is consistent with biological knowledge and to
extract potential new knowledge and insight from the model. In contrast, explainable AI models provide interpretable insights, so when
applied to predicting cell types and states, such models allow users to examine the relative importance and contribution of different
features or combinations of features in determining each cell type assignment. (b) The “curse of dimensionality” refers to the challenges
that arise in analyses of high-dimensional data, such as immunological measurements where each sample is characterized by tens of
thousands of features. As the dimensionality increases, data points become extremely sparsely distributed, making it extremely
challenging to identify meaningful patterns or relationships (in ML lingo: Learning mapping functions that can interpolate the data
values in the vast space without any observed data is a challenging task). However, many features measured at higher dimensions may be
redundant (e.g., correlated with other features) or irrelevant for a given predictive task, thus adding noise that can obscure the true
biological signals. Dimensionality reduction techniques, or representation learning, address these challenges by mapping highdimensional data into a lower-dimensional space while revealing the underlying manifold structure using fewer features. These learned
representations transform the original feature space into new informative features, enabling more informative visualizations.
Importantly, these new features serve as inputs for downstream learning tasks that can benefit from the compressed and more
informative representations (or embeddings). In this hypothetical example, a representation learning procedure can compress and
“unfold” the data in three-dimensional space into a new representation in two dimensions. The new representation captures the
topological structure of the data, allowing easy classification of the data points belonging to different color categories (essentially, a
linear, cutoff-based function will successfully delineate the points). Such a classification function would be significantly more
challenging to learn in the original three-dimensional space. (c) While traditional dimensionality reduction techniques, such as
principal component analysis, are useful for transforming data into lower dimensions, they are often limited to capturing linear
relationships (i.e., summing up contributions from features) and may not fully capture complex nonlinear relationships or topological
features. Neural network models, such as autoencoders, can overcome these limitations by learning nonlinear mapping functions that
compress high-dimensional input features into low-dimensional representations that preserve biologically meaningful information. In
this case, the preservation of information is formally trained and assessed via a decoder, which aims to reconstruct the original data from
the low-dimensional representation. The extent to which the original data can be reconstructed from the compressed representation
also provides a way to assess the quality of the learned representations, ensuring their biological relevance. The low-dimensional
representations can be used for downstream ML tasks, such as clustering, classification, or prediction of outcomes. Panel b adapted
from Reference 194 (CC BY 4.0). Panel c adapted from Reference 195.
708
Lei • Tsang
Downloaded from www.annualreviews.org. Guest (guest) IP: 74.109.18.71 On: Sat, 17 May 2025 11:55:57
in unsupervised learning, the model learns patterns and structures from unlabeled data and is
typically used for clustering, dimensionality reduction, and anomaly detection.
Unsupervised learning can serve as a first step toward the discovery of structure or patterns in
the data, which can then be combined with the original labeled data set to enhance the performance of the subsequent supervised ML step. This process is closely related to a key ML approach
called representation learning (205), which aims to overcome the “curse of dimensionality,” a problem that arises when working with high-dimensional immune state measurements (Figure 3b).
ML models are inherently data interpolation schemes in that they attempt to learn mathematical
functions to fill in the unseen data points (or data volumes) in the input data space. A single-variable
input is easy to interpolate even after seeing a relatively small quantity of training data—when a linear model is used, this corresponds to the classic single-variable regression problem. However, the
data volume grows exponentially as a function of the number of input dimensions. Thus, given the
large number of immune-profiling parameters that are now routinely being measured in systems
human immunology studies (minimally in the 105 –106 range when transcriptomics data are used),
interpolation becomes extremely challenging. Representation learning aims to discover compressions of high-dimensional data into lower-dimensional spaces to capture the correlated structure
and relationship between the input features. The resulting, more succinct representations, or embeddings, are often more useful encodings of the input data, such that downstream supervised
learning becomes easier and more robust (Figure 3c). This paradigm has been successfully applied in classical ML and statistical modeling, such as principal component analysis followed by
linear regression. In modern ML, it is exemplified by methods such as deep autoencoders (206),
which first derive latent representations of the input data and then feed these representations into
additional NN layers to perform supervised learning and downstream predictive tasks (207).
5.3. Methods Linking Personal Immune States to Functions and Outcomes
Supplemental Table 1 lists the key ML techniques used in systems human immunology studies to
uncover predictive insights on human immune setpoints and immune response predictors. Classic regression-based approaches are often used to model relationships between input and output
variables and to quantify the input variables’ effects on functional outputs of the immune system.
These approaches include linear regression for continuous outcomes (72, 208), logistic regression
for binary outcomes (28, 209), and generalized linear models to incorporate diverse distributions
beyond the normal (84, 210). Regression-based models can also be extended to handle more complex data structures and relationships. For example, mixed effects models, which incorporate both
fixed and random effects—the former can model contributions from variables such as sex, while
the latter can account for effects such as stable baseline differences across individuals—are useful for analyzing hierarchical data structures and repeated measures common in immunological
studies, such as longitudinal vaccine responses (85, 211) and clinical outcomes (212). We recently
used this approach to model the hierarchical variations arising from multimodal single-cell data
obtained from different individuals sampled longitudinally over time before and after vaccination
(85). Moreover, regularized regression methods such as lasso and ridge regression are well-suited
to handling multicollinearity, or correlations between predictors. They reduce model complexity
and prevent overfitting by applying penalties to regression coefficients, making them particularly
valuable for analyzing the many correlated variables generated by immune profiling, including
single-cell multiomics (204).
Although ML tools excel at extracting potentially meaningful patterns that might be obscured
by the messiness of biological systems, they cannot inherently prove causal relationships. While
one could drop out specific input variables (or combinations thereof ) to computationally assess
www.annualreviews.org • Systems Human Immunology and AI
709
their contributions to prediction performance, rigorous experimental validations are required to
elucidate the mechanistic basis, establish causality, and provide feedback to refine and improve
computational models. The synergy between computational and experimental methods fosters
an iterative cycle positioned to catalyze deeper insights and drive the field toward more accurate
predictions and clinically transformative applications.
6. PERSPECTIVE AND CONCLUDING REMARKS
Downloaded from www.annualreviews.org. Guest (guest) IP: 74.109.18.71 On: Sat, 17 May 2025 11:55:57
The integration of systems human immunology and AI is poised for significant advances, which
will build on current methodological, technological, and conceptual developments to improve our
understanding of human immune variation. As we look ahead, several emerging trends will shape
the advancement of the field. Addressing the historical bias in systems immunology studies toward
relatively small cohorts in economically more advanced regions, large-scale studies are slowly expanding the scope to include traditionally underrepresented populations and a broader range of
variables shaping human immunity (83), including factors related to urban and rural environments (75). To enhance data comparability and reproducibility across diverse research sites and
populations, initiatives are underway to develop globally applicable and replicable standardized
immune-profiling kits and protocols (83, 213–215; see https://www.humanimmunomeproject.
org/scientific-approach). These efforts could lead to a unified framework for immune profiling,
spanning sample collection, processing, and data reporting. Such standardization will not only
facilitate seamless data integration and reduce cost but also enhance the robustness and generalizability of immunological findings. Adding to highly standardized immune profiling, in vivo
immune cell–based sensors represent an exciting frontier (216; see https://www.czbiohub.org).
Immune cells such as T cells surveil diverse tissues and are thus natural sensors of tissue and
physiologic states. Harnessing such microscale biosensors both naturally and synthetically via
engineered immune cells could lead to real-time, high-resolution monitoring and reporting of
physiologic changes within the human body (217). These technologies could lead to the generation of massive temporal and spatial data sets within and across individuals, which would fuel
advanced computational and AI modeling, potentially revealing new immune health dimensions
to enable early disease detection and precise immune modulation strategies.
6.1. Toward Human Immune State Embeddings
While current statistical and ML methods have been instrumental in identifying predictors of
immune features and outcomes, as the multiplicity and complexity of data modalities increase
together with more extensive longitudinal profiling of ever larger populations, the complex,
nonlinear relationships inherent in the resulting high-dimensional data sets will demand more
expressive ML models, such as DNNs. DNN models are adept at scaling effectively and often
show more pronounced performance gains with increasing data size, leveraging their deep architectures to learn richer, more nuanced representations. Building on these strengths, foundation
models—DNNs with a very large number of parameters trained in an unsupervised fashion on
massive, diverse data sets—that have been developed for genomes (218), proteins (219, 220), and
cells (201–203, 221, 222) could become a core tool to transform heterogeneous data into predictive immunological insights. They also facilitate the exchange of knowledge and AI models across
immunological contexts through transfer learning (223). The few-shot learning capabilities (i.e.,
requiring very few additional training examples for new applications) and support for fine-tuning
of foundational models enable quick adaptation to new immunological subdomains and predictive
tasks, even when data specific to that subdomain are limited (224). We envision that DNNs and
other emerging AI advances will provide novel representations or embeddings of human immune
710
Lei • Tsang
Downloaded from www.annualreviews.org. Guest (guest) IP: 74.109.18.71 On: Sat, 17 May 2025 11:55:57
states and their correlation structure to empower a predictive and, ultimately, causal mechanistic
understanding of the human immune system and its interactions with physiology. Achieving such
informative, compressed representations also has practical implications: The number and type of
immune state measurements could be dramatically reduced by focusing on these key embedding
states. Doing so would lead to highly standardized immune profiling with reduced cost and enhanced robustness and could thus allow for a significant expansion of cohort sizes and sampling
frequency.
6.2. Toward Effective Integration of Data-Driven and Mechanistic Models
Successful integration of AI in systems immunology is not without challenges. As discussed above,
ML models such as DNNs often function as black boxes, making their predictive and decisionmaking processes inscrutable (225). The lack of interpretability hinders our ability to extract
causal, mechanistic insights that can guide further hypothesis generation and experimentation.
Without a clear understanding of how the model arrives at its predictions, identifying and mitigating inherent biases become exceedingly difficult. When models generate incorrect predictions
or exhibit unexpected behavior, interpretability is essential for pinpointing the root cause and driving model refinement. These issues collectively contribute to limited trust and adoption among
researchers and clinicians, who justifiably expect concrete proof that ML predictions are grounded
in biologically relevant features and relationships (226). Moreover, most ML models are designed
to recognize patterns and correlations in the data but lack the inherent ability to discern whether
these relationships are causal (12, 227). For example, correlations between cytokine levels and
observed immune responses do not necessarily imply causation. Such findings, though valuable
for prediction, require further investigation to establish causal relationships in immune system
behavior. These limitations underscore the need to integrate complementary approaches such as
quantitative mechanistic models, which extend beyond statistical associations to simulate and interrogate immunological processes bottom up (228, 229). These models are typically constructed
on the basis of existing knowledge of the central players in a particular process or, more rarely, by
incorporating key molecular and cellular components identified through statistical and ML-based
approaches. These models can be formulated, parameterized, and simulated computationally using agent-based approaches, systems of ordinary (or partial) differential equations (230), or the
Gillespie algorithm (231), for example. Thus, mechanistic models can in principle predict how
molecular-level changes can propagate across scales over time under a variety of conditions, which
would be impractical to test experimentally (229). However, a major challenge is the large number of parameters with unknown values; how they vary across individuals in the population is also
largely unknown. New concepts are needed to bridge this scale gap: Often, top-down, data-driven
approaches powered by AI result in features that are not easy to model directly via mechanistic
modeling approaches. Perhaps AI will again come to the rescue; for example, by relatively sparse
sampling of the parameter space of stochastic mechanistic models, as discussed above, ML can
learn the mapping between parameter configuration and phenotypic outputs of the system under
simulation (232, 233). Thus, we envision that, to bridge this top-down (AI- and data-driven) versus bottom-up (mechanistic modeling) scale gap, ML can map human immune state embeddings
to molecular and cellular state embeddings to reveal key dynamic variables for incorporation into
mechanistic models.
Systems human immunology has made remarkable strides in delineating human immunological diversity and health. The convergence of high-throughput technologies generating massive
data and AI is setting the stage for mapping and predicting the dynamic landscape and trajectory of individual human immunity and physiologic health. Envisioning the path ahead,
www.annualreviews.org • Systems Human Immunology and AI
711
interpretable AI models capable of forecasting immune responses based on an array of individual
factors—including immune profiles, ancestry, age, sex, socioeconomic factors, and other pertinent
information represented as personal immune state embeddings—hold immense promise. Emerging research strategies and tools could propel us toward a future where immune health is precisely
monitored, predicted, and modulated.
DISCLOSURE STATEMENT
Downloaded from www.annualreviews.org. Guest (guest) IP: 74.109.18.71 On: Sat, 17 May 2025 11:55:57
J.S.T. serves on the scientific advisory boards of CytoReason Inc. and Immunoscape Inc. and as
the co–chief scientific officer (as an unpaid volunteer) of the Human Immunome Project.
AUTHOR CONTRIBUTIONS
J.S.T. conceptualized and guided the review. Y.L. and J.S.T. wrote and finalized the manuscript.
ACKNOWLEDGMENTS
We thank Shubham Tripathi for comments on the manuscript. This work is supported by a Chan
Zuckerberg Biohub Investigator award to J.S.T. and by the following grants from the National
Institute of Allergy and Infectious Diseases of the National Institutes of Health (NIH) to J.S.T.:
R01AI170116, U01AI153700, U01AI165745, and U19AI145825. Y.L. is supported in part by the
NIH Medical Scientist Training Program at the Yale University School of Medicine.
LITERATURE CITED
1. Pulendran B, Davis MM. 2020. The science and medicine of human immunology. Science
369(6511):eaay4014
2. Meizlish ML, Franklin RA, Zhou X, Medzhitov R. 2021. Tissue homeostasis and inflammation. Annu.
Rev. Immunol. 39:557–81
3. Rankin LC, Artis D. 2018. Beyond host defense: emerging functions of the immune system in regulating
complex tissue physiology. Cell 173(3):554–67
4. Baysoy A, Bai Z, Satija R, Fan R. 2023. The technological landscape and applications of single-cell
multi-omics. Nat. Rev. Mol. Cell Biol. 24(10):695–713
5. Tsang JS. 2015. Utilizing population variation, vaccination, and systems biology to study human
immunology. Trends Immunol. 36(8):479–93
6. Liston A, Humblet-Baron S, Duffy D, Goris A. 2021. Human immune diversity: from evolution to
modernity. Nat. Immunol. 22(12):1479–89
7. Tsang JS, Dobaño C, VanDamme P, Moncunill G, Marchant A, et al. 2020. Improving vaccine-induced
immunity: Can baseline predict outcome? Trends Immunol. 41(6):457–65
8. Forlin R, James A, Brodin P. 2023. Making human immune systems more interpretable through systems
immunology. Trends Immunol. 44(8):577–84
9. Perez-Lopez R, Ghaffari Laleh N, Mahmood F, Kather JN. 2024. A guide to artificial intelligence for
cancer researchers. Nat. Rev. Cancer 24:427–41
10. Wang H, Fu T, Du Y, Gao W, Huang K, et al. 2023. Scientific discovery in the age of artificial
intelligence. Nature 620(7972):47–60
11. Germain RN, Van Allen EM, Trynka G, Tsang JS, Grün D, et al. 2024. AI and immunology. Immunity
57(6):1177–81
12. Camacho DM, Collins KM, Powers RK, Costello JC, Collins JJ. 2018. Next-generation machine
learning for biological networks. Cell 173(7):1581–92
13. Xhonneux L-P, Knight O, Lernmark Å, Bonifacio E, Hagopian WA, et al. 2021. Transcriptional
networks in at-risk individuals identify signatures of type 1 diabetes progression. Sci. Transl. Med.
13(587):eabd5666
712
Lei • Tsang
Downloaded from www.annualreviews.org. Guest (guest) IP: 74.109.18.71 On: Sat, 17 May 2025 11:55:57
14. Zamecnik CR, Sowa GM, Abdelhak A, Dandekar R, Bair RD, et al. 2024. An autoantibody signature
predictive for multiple sclerosis. Nat. Med. 30(5):1300–8
15. Sparks R, Rachmaninoff N, Lau WW, Hirsch DC, Bansal N, et al. 2024. A unified metric of human
immune health. Nat. Med. 30:2461–72
16. Szeto GL, Finley SD. 2019. Integrative approaches to cancer immunotherapy. Trends Cancer 5(7):400–10
17. Wimmers F, Pulendran B. 2020. Emerging technologies for systems vaccinology—multi-omics
integration and single-cell (epi)genomic profiling. Curr. Opin. Immunol. 65:57–64
18. Brodin P, Davis MM. 2017. Human immune system variation. Nat. Rev. Immunol. 17(1):21–29
19. Jongo SA, Shekalaghe SA, Church LWP, Ruben AJ, Schindler T, et al. 2018. Safety, immunogenicity,
and protective efficacy against controlled human malaria infection of Plasmodium falciparum sporozoite
vaccine in Tanzanian adults. Am. J. Trop. Med. Hyg. 99(2):338–49
20. Fine PEM. 1995. Variation in protection by BCG: implications of and for heterologous immunity. Lancet
346(8986):1339–45
21. Dockrell HM, Smith SG. 2017. What have we learnt about BCG vaccination in the last 20 years? Front.
Immunol. 8:1134
22. Muyanja E, Ssemaganda A, Ngauv P, Cubas R, Perrin H, et al. 2014. Immune activation alters cellular
and humoral responses to yellow fever 17D vaccine. J. Clin. Investig. 124(7):3147–58
23. Mbow M, Lell B, Jochems SP, Cisse B, Mboup S, et al. 2020. COVID-19 in Africa: dampening the
storm? Science 369(6504):624–26
24. Krieg C, Nowicka M, Guglietta S, Schindler S, Hartmann FJ, et al. 2018. High-dimensional single-cell
analysis predicts response to anti-PD-1 immunotherapy. Nat. Med. 24(2):144–53
25. Lozano AX, Chaudhuri AA, Nene A, Bacchiocchi A, Earland N, et al. 2022. T cell characteristics associated with toxicity to immune checkpoint blockade in patients with melanoma. Nat. Med.
28(2):353–62
26. Yoosuf N, Maciejewski M, Ziemek D, Jelinsky SA, Folkersen L, et al. 2021. Early prediction of clinical response to anti-TNF treatment using multi-omics and machine learning in rheumatoid arthritis.
Rheumatology 61(4):1680–89
27. Lee JWJ, Plichta D, Hogstrom L, Borren NZ, Lau H, et al. 2021. Multi-omics reveal microbial determinants impacting responses to biologic therapies in inflammatory bowel disease. Cell Host Microbe
29(8):1294–304.e4
28. Kotliarov Y, Sparks R, Martins AJ, Mulè MP, Lu Y, et al. 2020. Broad immune activation underlies shared
set point signatures for vaccine responsiveness in healthy individuals and disease activity in patients with
lupus. Nat. Med. 26(4):618–29
29. Robinson GA, Peng J, Dönnes P, Coelewij L, Naja M, et al. 2020. Disease-associated and patient-specific
immune cell signatures in juvenile-onset systemic lupus erythematosus: patient stratification using a
machine-learning approach. Lancet Rheumatol. 2(8):e485–96
30. Liston A, Carr EJ, Linterman MA. 2016. Shaping variation in the human immune system. Trends
Immunol. 37(10):637–46
31. Orrù V, Steri M, Sole G, Sidore C, Virdis F, et al. 2013. Genetic variants regulating immune cell levels
in health and disease. Cell 155(1):242–56
32. Orrù V, Steri M, Sidore C, Marongiu M, Serra V, et al. 2020. Complex genetic signatures in immune
cells underlie autoimmunity and inform therapy. Nat. Genet. 52(10):1036–45
33. Yan K, Cai W, Cao F, Sun H, Chen S, et al. 2013. Genetic effects have a dominant role on poor responses
to infant vaccination to hepatitis B virus. J. Hum. Genet. 58(5):293–97
34. Schaid DJ, Haralambieva IH, Larrabee BR, Ovsyannikova IG, Kennedy RB, Poland GA. 2017.
Heritability of vaccine-induced measles neutralizing antibody titers. Vaccine 35(10):1390–94
35. Brodin P, Jojic V, Gao T, Bhattacharya S, Angel CJL, et al. 2015. Variation in the human immune system
is largely driven by non-heritable influences. Cell 160(1):37–47
36. Roederer M, Quaye L, Mangino M, Beddall MH, Mahnke Y, et al. 2015. The genetic architecture of the
human immune system: a bioresource for autoimmunity and disease pathogenesis. Cell 161(2):387–403
37. Cheung P, Vallania F, Warsinske HC, Donato M, Schaffert S, et al. 2018. Single-cell chromatin
modification profiling reveals increased epigenetic variations with aging. Cell 173(6):1385–97.e14
www.annualreviews.org • Systems Human Immunology and AI
713
Downloaded from www.annualreviews.org. Guest (guest) IP: 74.109.18.71 On: Sat, 17 May 2025 11:55:57
38. Dunn SE, Perry WA, Klein SL. 2024. Mechanisms and consequences of sex differences in immune
responses. Nat. Rev. Nephrol. 20(1):37–55
39. Klein SL, Flanagan KL. 2016. Sex differences in immune responses. Nat. Rev. Immunol. 16(10):626–38
40. Abdullah M, Chai P-S, Chong M-Y, Tohit ERM, Ramasamy R, et al. 2012. Gender effect on in vitro
lymphocyte subset levels of healthy individuals. Cell. Immunol. 272(2):214–19
41. Furman D, Hejblum BP, Simon N, Jojic V, Dekker CL, et al. 2014. Systems analysis of sex differences reveals an immunosuppressive role for testosterone in the response to influenza vaccination. PNAS
111(2):869–74
42. Simon AK, Hollander GA, McMichael A. 2015. Evolution of the immune system in humans from infancy
to old age. Proc. R. Soc. B 282(1821):20143085
43. Torow N, Hand TW, Hornef MW. 2023. Programmed and environmental determinants driving
neonatal mucosal immune development. Immunity 56(3):485–99
44. Olin A, Henckel E, Chen Y, Lakshmikanth T, Pou C, et al. 2018. Stereotypic immune system
development in newborn children. Cell 174(5):1277–92.e14
45. Jain N. The early life education of the immune system: moms, microbes and (missed) opportunities. Gut
Microbes 12(1):1824564
46. Edler P, Schwab LSU, Aban M, Wille M, Spirason N, et al. 2024. Immune imprinting in early life shapes
cross-reactivity to influenza B virus haemagglutinin. Nat. Microbiol. 9:2073–83
47. Vashishtha VM, Kumar P. 2024. The durability of vaccine-induced protection: an overview. Expert Rev.
Vaccines 23(1):389–408
48. Netea MG, Domínguez-Andrés J, Barreiro LB, Chavakis T, Divangahi M, et al. 2020. Defining trained
immunity and its role in health and disease. Nat. Rev. Immunol. 20(6):375–88
49. Benn CS, Fisker AB, Rieckmann A, Sørup S, Aaby P. 2020. Vaccinology: time to change the paradigm?
Lancet Infect. Dis. 20(10):e274–83
50. Franceschi C, Garagnani P, Vitale G, Capri M, Salvioli S. 2017. Inflammaging and ‘garb-aging.’ Trends
Endocrinol. Metab. 28(3):199–212
51. Nikolich-Žugich J. 2018. The twilight of immunity: emerging concepts in aging of the immune system.
Nat. Immunol. 19(1):10–19
52. Ferrucci L, Fabbri E. 2018. Inflammageing: chronic inflammation in ageing, cardiovascular disease, and
frailty. Nat. Rev. Cardiol. 15(9):505–22
53. Furman D. 2019. Chronic inflammation in the etiology of disease across the life span. Nat. Med. 25:1822–
32
54. Han S, Georgiev P, Ringel AE, Sharpe AH, Haigis M. 2023. Age-associated remodeling of T cell
immunity and metabolism. Cell Metab. 35(1):36–55
55. Frasca D, Blomberg BB. 2020. Aging induces B cell defects and decreased antibody responses to influenza
infection and vaccination. Immun. Ageing 17(1):37
56. Collier DA, Ferreira IATM, Kotagiri P, Datir RP, Lim EY, et al. 2021. Age-related immune response
heterogeneity to SARS-CoV-2 vaccine BNT162b2. Nature 596(7872):417–22
57. Hou Y, Chen M, Bian Y, Hu Y, Chuan J, et al. 2024. Insights into vaccines for elderly individuals: from
the impacts of immunosenescence to delivery strategies. npj Vaccines 9(1):77
58. Franceschi C, Campisi J. 2014. Chronic inflammation (inflammaging) and its potential contribution to
age-associated diseases. J. Gerontol. A 69(Suppl. 1):S4–9
59. Fulop T, Witkowski JM, Olivieri F, Larbi A. 2018. The integration of inflammaging in age-related
diseases. Semin. Immunol. 40:17–35
60. Biragyn A, Ferrucci L. 2018. Gut dysbiosis: a potential link between increased cancer risk in ageing and
inflammaging. Lancet Oncol. 19(6):e295–304
61. Alpert A, Pickman Y, Leipold M, Rosenberg-Hasson Y, Ji X, et al. 2019. A clinically meaningful metric
of immune age derived from high-dimensional longitudinal monitoring. Nat. Med. 25(3):487–95
62. Sayed N, Huang Y, Nguyen K, Krejciova-Rajaniemi Z, Grawe AP, et al. 2021. An inflammatory aging
clock (iAge) based on deep learning tracks multimorbidity, immunosenescence, frailty and cardiovascular
aging. Nat. Aging 1:598–615
63. Belsky DW, Caspi A, Houts R, Cohen HJ, Corcoran DL, et al. 2015. Quantification of biological aging
in young adults. PNAS 112(30):E4104–10
714
Lei • Tsang
Downloaded from www.annualreviews.org. Guest (guest) IP: 74.109.18.71 On: Sat, 17 May 2025 11:55:57
64. Chen BH, Marioni RE, Colicino E, Peters MJ, Ward-Caviness CK, et al. 2016. DNA methylation–based
measures of biological age: meta-analysis predicting time to death. Aging 8(9):1844–59
65. Jylhävä J, Pedersen NL, Hägg S. 2017. Biological age predictors. eBioMedicine 21:29–36
66. Ferrucci L, Gonzalez-Freire M, Fabbri E, Simonsick E, Tanaka T, et al. 2020. Measuring biological
aging in humans: a quest. Aging Cell 19(2):e13080
67. Tian YE, Cropley V, Maier AB, Lautenschlager NT, Breakspear M, Zalesky A. 2023. Heterogeneous aging across multiple organ systems and prediction of chronic disease and mortality. Nat. Med.
29(5):1221–31
68. Carr EJ, Dooley J, Garcia-Perez JE, Lagou V, Lee JC, et al. 2016. The cellular composition of the human
immune system is shaped by age and cohabitation. Nat. Immunol. 17(4):461–68
69. Whitney AR, Diehn M, Popper SJ, Alizadeh AA, Boldrick JC, et al. 2003. Individuality and variation in
gene expression patterns in human blood. PNAS 100(4):1896–901
70. Furman D, Jojic V, Sharma S, Shen-Orr SS, Angel CJL, et al. 2015. Cytomegalovirus infection enhances
the immune response to influenza. Sci. Transl. Med. 7:281ra43
71. Schluter J, Peled JU, Taylor BP, Markey KA, Smith M, et al. 2020. The gut microbiota is associated with
immune cell dynamics in humans. Nature 588(7837):303–7
72. Investig. IBDMDB, Lloyd-Price J, Arze C, Ananthakrishnan AN, Schirmer M, et al. 2019. Multi-omics
of the gut microbial ecosystem in inflammatory bowel diseases. Nature 569(7758):655–62
73. Hagan T, Cortese M, Rouphael N, Boudreau C, Linde C, et al. 2019. Antibiotics-driven gut microbiome
perturbation alters immunity to vaccines in humans. Cell 178(6):1313–28.e13
74. Nakaya HI, Hagan T, Duraisingham SS, Lee EK, Kwissa M, et al. 2015. Systems analysis of immunity to influenza vaccination across multiple years and in diverse populations reveals shared molecular
signatures. Immunity 43(6):1186–98
75. van Dorst MMAR, Pyuza JJ, Nkurunungi G, Kullaya VI, Smits HH, et al. 2023. Immunological factors
linked to geographical variation in vaccine responses. Nat. Rev. Immunol. 24(4):250–63
76. Chao F, You D, Pedersen J, Hug L, Alkema L. 2018. National and regional under-5 mortality rate by
economic status for low-income and middle-income countries: a systematic assessment. Lancet Glob.
Health 6(5):e535–47
77. Hill DL, Carr EJ, Rutishauser T, Moncunill G, Campo JJ, et al. 2020. Immune system development
varies according to age, location, and anemia in African children. Sci. Transl. Med. 12(529):eaaw9522
78. RTS, S Clin. Trials Partnersh. 2015. Efficacy and safety of RTS,S/AS01 malaria vaccine with or without a booster dose in infants and children in Africa: final results of a phase 3, individually randomised,
controlled trial. Lancet 386(9988):31–45
79. White MT, Verity R, Griffin JT, Asante KP, Owusu-Agyei S, et al. 2015. Immunogenicity of the
RTS,S/AS01 malaria vaccine and implications for duration of vaccine efficacy: secondary analysis of
data from a phase 3 randomised controlled trial. Lancet Infect. Dis. 15(12):1450–58
80. Kazmin D, Nakaya HI, Lee EK, Johnson MJ, van der Most R, et al. 2017. Systems analysis of protective
immune responses to RTS,S malaria vaccination in humans. PNAS 114(9):2425–30
81. Mordmüller B, Surat G, Lagler H, Chakravarty S, Ishizuka AS, et al. 2017. Sterile protection against
human malaria by chemoattenuated PfSPZ vaccine. Nature 542(7642):445–49
82. Ishizuka AS, Lyke KE, DeZure A, Berry AA, Richie TL, et al. 2016. Protection against malaria at 1 year
and immune correlates following PfSPZ vaccination. Nat. Med. 22(6):614–23
83. Leslie M. 2024. Megaproject will chart human immune diversity. Science 383(6678):13–14
84. Sparks R, Lau WW, Liu C, Han KL, Vrindten KL, et al. 2023. Influenza vaccination reveals sex
dimorphic imprints of prior mild COVID-19. Nature 614(7949):752–61
85. Mulè MP, Martins AJ, Cheung F, Farmer R, Sellers BA, et al. 2024. Integrating population and singlecell variations in vaccine responses identifies a naturally adjuvanted human immune setpoint. Immunity
57(5):1160–76.e7
86. Ravichandran S, Erra-Diaz F, Karakaslar OE, Marches R, Kenyon-Pesce L, et al. 2024. Distinct baseline immune characteristics associated with responses to conjugated and unconjugated pneumococcal
polysaccharide vaccines in older adults. Nat. Immunol. 25(2):316–29
www.annualreviews.org • Systems Human Immunology and AI
715
Downloaded from www.annualreviews.org. Guest (guest) IP: 74.109.18.71 On: Sat, 17 May 2025 11:55:57
87. Rubelt F, Bolen CR, McGuire HM, Vander Heiden JA, Gadala-Maria D, et al. 2016. Individual heritable
differences result in unique cell lymphocyte receptor repertoires of naïve and antigen-experienced cells.
Nat. Commun. 7:11112
88. Jiang N, Malone M, Chizari S. 2023. Antigen-specific and cross-reactive T cells in protection and disease.
Immunol. Rev. 316(1):120–35
89. Wang Q, Feng D, Jia S, Lu Q, Zhao M. 2024. B-cell receptor repertoire: recent advances in autoimmune
diseases. Clin. Rev. Allergy Immunol. 66(1):76–98
90. Yuuki H, Itamiya T, Nagafuchi Y, Ota M, Fujio K. 2024. B cell receptor repertoire abnormalities in
autoimmune disease. Front. Immunol. 15:1326823
91. Sun X, Nguyen T, Achour A, Ko A, Cifello J, et al. 2022. Longitudinal analysis reveals age-related
changes in the T cell receptor repertoire of human T cell subsets. J. Clin. Investig. 132(17):e158122
92. Ghraichy M, Galson JD, Kovaltsuk A, von Niederhäusern V, Pachlopnik Schmid J, et al. 2020.
Maturation of the human immunoglobulin heavy chain repertoire with age. Front. Immunol. 11:1734
93. Sender R, Weiss Y, Navon Y, Milo I, Azulay N, et al. 2023. The total mass, number, and distribution of
immune cells in the human body. PNAS 120(44):e2308511120
94. HIPC-CHI Signat. Proj. Team, HIPC-I Consort., Avey S, Cheung F, Fermin D, et al. 2017. Multicohort analysis reveals baseline transcriptional predictors of influenza vaccination responses. Sci. Immunol.
2(14):eaal4656
95. Tsang JS, Schwartzberg PL, Kotliarov Y, Biancotto A, Xie Z, et al. 2014. Global analyses of human
immune variation reveal baseline predictors of postvaccination responses. Cell 157(2):499–513
96. Fourati S, Cristescu R, Loboda A, Talla A, Filali A, et al. 2016. Pre-vaccination inflammation and B-cell
signalling predict age-related hyporesponse to hepatitis B vaccination. Nat. Commun. 7:10369
97. Fourati S, Tomalin LE, Mulè MP, Chawla DG, Gerritsen B, et al. 2022. Pan-vaccine analysis reveals
innate immune endotypes predictive of antibody responses to vaccination. Nat. Immunol. 23(12):1777–87
98. Nehar-Belaid D, Sokolowski M, Ravichandran S, Banchereau J, Chaussabel D, Ucar D. 2023. Baseline immune states (BIS) associated with vaccine responsiveness and factors that shape the BIS. Semin.
Immunol. 70:101842
99. Tao W, Concepcion AN, Vianen M, Marijnissen ACA, Lafeber FPGJ, et al. 2021. Multiomics and machine learning accurately predict clinical response to adalimumab and etanercept therapy in patients
with rheumatoid arthritis. Arthritis Rheumatol. 73(2):212–22
100. Hromić-Jahjefendić A, Lundstrom K, Adilović M, Aljabali AAA, Tambuwala MM, et al. 2024.
Autoimmune response after SARS-CoV-2 infection and SARS-CoV-2 vaccines. Autoimmun. Rev.
23(3):103508
101. Pisetsky DS. 2023. Pathogenesis of autoimmune disease. Nat. Rev. Nephrol. 19(8):509–24
102. Bjornevik K, Münz C, Cohen JI, Ascherio A. 2023. Epstein–Barr virus as a leading cause of multiple
sclerosis: mechanisms and implications. Nat. Rev. Neurol. 19(3):160–71
103. Shen-Orr SS, Furman D, Kidd BA, Hadad F, Lovelace P, et al. 2016. Defective signaling in the JAKSTAT pathway tracks with chronic inflammation and cardiovascular risk in aging humans. Cell Syst.
3(4):374–84.e4
104. Kollmann TR, Levy O, Montgomery RR, Goriely S. 2012. Innate immune function by Toll-like
receptors: distinct responses in newborns and the elderly. Immunity 37(5):771–83
105. Van De Sandt CE, Nguyen THO, Gherardin NA, Crawford JC, Samir J, et al. 2023. Newborn and childlike molecular signatures in older adults stem from TCR shifts across human lifespan. Nat. Immunol.
24(11):1890–907
106. Fanucchi S, Domínguez-Andrés J, Joosten LAB, Netea MG, Mhlanga MM. 2021. The intersection of
epigenetics and metabolism in trained immunity. Immunity 54(1):32–43
107. Gu J, Liu Q, Zhang J, Xu S. 2023. COVID-19 and trained immunity: the inflammatory burden of long
covid. Front. Immunol. 14:1294959
108. Cheong J-G, Ravishankar A, Sharma S, Parkhurst CN, Grassmann SA, et al. 2023. Epigenetic memory
of coronavirus infection in innate immune cells and their progenitors. Cell 186(18):3882–902.e24
109. Patel AA, Ginhoux F, Yona S. 2021. Monocytes, macrophages, dendritic cells and neutrophils: an update
on lifespan kinetics in health and disease. Immunology 163(3):250–61
716
Lei • Tsang
Downloaded from www.annualreviews.org. Guest (guest) IP: 74.109.18.71 On: Sat, 17 May 2025 11:55:57
110. Sun SJ, Aguirre-Gamboa R, de Bree LCJ, Sanz J, Dumaine A, et al. 2024. BCG vaccination alters the
epigenetic landscape of progenitor cells in human bone marrow to influence innate immune responses.
Immunity 57(9):2095–107.e8
111. Apps R, Biancotto A, Candia J, Kotliarov Y, Perl S, et al. 2024. Acute and persistent responses of H5N1
vaccination in humans. Cell Rep. 43(9):114706
112. Wimmers F, Donato M, Kuo A, Ashuach T, Gupta S, et al. 2021. The single-cell epigenomic and
transcriptional landscape of immunity to influenza vaccination. Cell 184(15):3915–35.e21
113. Knight HR, Ketter E, Ung T, Weiss A, Ajit J, et al. 2024. High-throughput screen identifies non
inflammatory small molecule inducers of trained immunity. PNAS 121(29):e2400413121
114. Cheng Y, He C, Wang M, Ma X, Mo F, et al. 2019. Targeting epigenetic regulators for cancer therapy:
mechanisms and advances in clinical trials. Signal Transduct. Target Ther. 4(1):62
115. Mulder WJM, Ochando J, Joosten LAB, Fayad ZA, Netea MG. 2019. Therapeutic targeting of trained
immunity. Nat. Rev. Drug Discov. 18(7):553–66
116. Baruch EN, Youngster I, Ben-Betzalel G, Ortenberg R, Lahat A, et al. 2021. Fecal microbiota transplant
promotes response in immunotherapy-refractory melanoma patients. Science 371(6529):602–9
117. Davar D, Dzutsev AK, McCulloch JA, Rodrigues RR, Chauvin J-M, et al. 2021. Fecal microbiota
transplant overcomes resistance to anti-PD-1 therapy in melanoma patients. Science 371(6529):595–602
118. Ahuja SK, Manoharan MS, Lee GC, McKinnon LR, Meunier JA, et al. 2023. Immune resilience despite
inflammatory stress promotes longevity and favorable health outcomes including resistance to infection.
Nat. Commun. 14:3286
119. Deane KD, O’Donnell CI, Hueber W, Majka DS, Lazar AA, et al. 2010. The number of elevated cytokines and chemokines in preclinical seropositive rheumatoid arthritis predicts time to diagnosis in an
age-dependent manner. Arthritis Rheumatol. 62(11):3161–72
120. Kokkonen H, Söderström I, Rocklöv J, Hallmans G, Lejon K, Rantapää Dahlqvist S. 2010. Upregulation of cytokines and chemokines predates the onset of rheumatoid arthritis. Arthritis Rheumatol.
62(2):383–91
121. Powers AC. 2021. Type 1 diabetes mellitus: much progress, many opportunities. J. Clin. Investig.
131(8):e142242
122. Guo J, Huang X, Dou L, Yan M, Shen T, et al. 2022. Aging and aging-related diseases: from molecular
mechanisms to interventions and treatments. Signal Transduct. Target Ther. 7(1):391
123. Tahami Monfared AA, Byrnes MJ, White LA, Zhang Q. 2022. Alzheimer’s disease: epidemiology and
clinical progression. Neurol. Ther. 11(2):553–69
124. Kuhlmann T, Moccia M, Coetzee T, Cohen JA, Correale J, et al. 2023. Multiple sclerosis progression:
time for a new mechanism-driven framework. Lancet Neurol. 22(1):78–88
125. Farber DL. 2021. Tissues, not blood, are where immune cells act. Nature 593(7860):506–9
126. Szabo PA, Levitin HM, Miron M, Snyder ME, Senda T, et al. 2019. Single-cell transcriptomics of human
T cells reveals tissue and activation signatures in health and disease. Nat. Commun. 10:4706
127. Szabo PA, Dogra P, Gray JI, Wells SB, Connors TJ, et al. 2021. Longitudinal profiling of respiratory
and systemic immune responses reveals myeloid cell–driven lung inflammation in severe COVID-19.
Immunity 54(4):797–814.e6
128. Habibi MS, Thwaites RS, Chang M, Jozwik A, Paras A, et al. 2020. Neutrophilic inflammation in the
respiratory mucosa predisposes to RSV infection. Science 370(6513):eaba9301
129. Loske J, Röhmel J, Lukassen S, Stricker S, Magalhães VG, et al. 2022. Pre-activated antiviral innate immunity in the upper airways controls early SARS-CoV-2 infection in children. Nat. Biotechnol.
40(3):319–24
130. Naik S, Larsen SB, Gomez NC, Alaverdyan K, Sendoel A, et al. 2017. Inflammatory memory sensitizes
skin epithelial stem cells to tissue damage. Nature 550(7677):475–80
131. Naik S, Fuchs E. 2022. Inflammatory memory and tissue adaptation in sickness and in health. Nature
607(7918):249–55
132. Lim AI, McFadden T, Link VM, Han S-J, Karlsson R-M, et al. 2021. Prenatal maternal infection
promotes tissue-specific immunity and inflammation in offspring. Science 373(6558):eabf3002
133. Ordovas-Montanes J, Dwyer DF, Nyquist SK, Buchheit KM, Vukovic M, et al. 2018. Allergic
inflammatory memory in human respiratory epithelial progenitor cells. Nature 560(7720):649–54
www.annualreviews.org • Systems Human Immunology and AI
717
Downloaded from www.annualreviews.org. Guest (guest) IP: 74.109.18.71 On: Sat, 17 May 2025 11:55:57
134. Del Poggetto E, Ho I-L, Balestrieri C, Yen E-Y, Zhang S, et al. 2021. Epithelial memory of inflammation
limits tissue damage while promoting pancreatic tumorigenesis. Science 373(6561):eabj0486
135. Jha A, Thwaites RS, Tunstall T, Kon OM, Shattock RJ, et al. 2021. Increased nasal mucosal interferon
and CCL13 response to a TLR7/8 agonist in asthma and allergic rhinitis. J. Allergy Clin. Immunol.
147(2):694–703.e12
136. Izmirly PM, Kim MY, Samanovic M, Fernandez-Ruiz R, Ohana S, et al. 2022. Evaluation of immune
response and disease status in systemic lupus erythematosus patients following SARS-CoV-2 vaccination.
Arthritis Rheumatol. 74(2):284–94
137. Sheridan PA, Paich HA, Handy J, Karlsson EA, Hudgens MG, et al. 2012. Obesity is associated with
impaired immune response to influenza vaccination in humans. Int. J. Obes. 36(8):1072–77
138. Frasca D, Ferracci F, Diaz A, Romero M, Lechner S, Blomberg BB. 2016. Obesity decreases B cell
responses in young and elderly individuals. Obesity 24(3):615–25
139. Lee AH, Shannon CP, Amenyogbe N, Bennike TB, Diray-Arce J, et al. 2019. Dynamic molecular
changes during the first week of human life follow a robust developmental trajectory. Nat. Commun.
10:1092
140. Pulendran B. 2014. Systems vaccinology: probing humanity’s diverse immune systems with vaccines.
PNAS 111(34):12300–6
141. Kidd BA, Peters LA, Schadt EE, Dudley JT. 2014. Unifying immunology with informatics and multiscale
biology. Nat. Immunol. 15(2):118–27
142. Davis MM, Tato CM, Furman D. 2017. Systems immunology: just getting started. Nat. Immunol.
18(7):725–32
143. Mangino M, Roederer M, Beddall MH, Nestle FO, Spector TD. 2017. Innate and adaptive immune
traits are differentially affected by genetic and environmental factors. Nat. Commun. 8:13850
144. Monto AS, Malosh RE, Evans R, Lauring AS, Gordon A, et al. 2019. Data resource profile: Household
Influenza Vaccine Evaluation (HIVE) study. Int. J. Epidemiol. 48(4):1040
145. Wells SB, Rainbow DB, Mark M, Szabo PA, Ergen C, et al. 2024. Multimodal profiling reveals tissuedirected signatures of human immune cells altered with age. bioRxiv 2024.01.03.573877. https://doi.
org/10.1101/2024.01.03.573877
146. Connors TJ, Matsumoto R, Verma S, Szabo PA, Guyer R, et al. 2023. Site-specific development and
progressive maturation of human tissue-resident memory T cells over infancy and childhood. Immunity
56(8):1894–909.e5
147. Wong MT, Ong DEH, Lim FSH, Teng KWW, McGovern N, et al. 2016. A high-dimensional atlas of
human T cell diversity reveals tissue-specific trafficking and cytokine signatures. Immunity 45(2):442–56
148. Lacouture ME, Goleva E, Shah N, Rotemberg V, Kraehenbuehl L, et al. 2024. Immunologic profiling of immune-related cutaneous adverse events with checkpoint inhibitors reveals polarized actionable
pathways. Clin. Cancer Res. 30(13):2822–34
149. Canales-Herrerias P, Uzzan M, Seki A, Czepielewski RS, Verstockt B, et al. 2024. Gut-associated lymphoid tissue attrition associates with response to anti-α4 β7 therapy in ulcerative colitis. Sci. Immunol.
9(94):eadg7549
150. Burke Schinkel SC, Barros PO, Berthoud T, Byrareddy SN, McGuinty M, et al. 2024. Comparative
analysis of human gut- and blood-derived mononuclear cells: contrasts in function and phenotype. Front.
Immunol. 15:1336480
151. D’Souza MP, Allen MA, Baumblatt JAG, Boggiano C, Crotty S, et al. 2018. Innovative approaches to
track lymph node germinal center responses to evaluate development of broadly neutralizing antibodies
in human HIV vaccine trials. Vaccine 36(38):5671–77
152. Austin JW, Buckner CM, Kardava L, Wang W, Zhang X, et al. 2019. Overexpression of T-bet in HIV
infection is associated with accumulation of B cells outside germinal centers and poor affinity maturation.
Sci. Transl. Med. 11(520):eaax0904
153. Turner JS, O’Halloran JA, Kalaidina E, Kim W, Schmitz AJ, et al. 2021. SARS-CoV-2 mRNA vaccines
induce persistent human germinal centre responses. Nature 596(7870):109–13
154. Kim W, Zhou JQ, Horvath SC, Schmitz AJ, Sturtz AJ, et al. 2022. Germinal centre–driven maturation
of B cell response to mRNA vaccination. Nature 604(7904):141–45
718
Lei • Tsang
Downloaded from www.annualreviews.org. Guest (guest) IP: 74.109.18.71 On: Sat, 17 May 2025 11:55:57
155. Alsoussi WB, Malladi SK, Zhou JQ, Liu Z, Ying B, et al. 2023. SARS-CoV-2 omicron boosting induces
de novo B cell response in humans. Nature 617(7961):592–98
156. Siggins MK, Thwaites RS, Openshaw PJM. 2021. Durability of immunity to SARS-CoV-2 and other
respiratory viruses. Trends Microbiol. 29(7):648–62
157. Swadling L, Diniz MO, Schmidt NM, Amin OE, Chandran A, et al. 2022. Pre-existing polymerasespecific T cells expand in abortive seronegative SARS-CoV-2. Nature 601(7891):110–17
158. Tao Y, Xing S, Zuo S, Bao P, Jin Y, et al. 2023. Cell-free multi-omics analysis reveals potential biomarkers
in gastrointestinal cancer patients’ blood. Cell Rep. Med. 4(11):101281
159. Loy CJ, Sotomayor-Gonzalez A, Servellita V, Nguyen J, Lenz J, et al. 2023. Nucleic acid biomarkers
of immune response and cell and tissue damage in children with COVID-19 and MIS-C. Cell Rep. Med.
4(6):101034
160. Sankaran VG, Weissman JS, Zon LI. 2022. Cellular barcoding to decipher clonal dynamics in disease.
Science 378(6616):eabm5874
161. Choi J, Chen W, Minkina A, Chardon FM, Suiter CC, et al. 2022. A time-resolved, multi-symbol
molecular recorder via sequential genome editing. Nature 608(7921):98–107
162. Chen W, Choi J, Li X, Nathans JF, Martin B, et al. 2024. Symbolic recording of signalling and cisregulatory element activity to DNA. Nature 632(8027):1073–81
163. Gold L, Ayers D, Bertino J, Bock C, Bock A, et al. 2010. Aptamer-based multiplexed proteomic
technology for biomarker discovery. PLOS ONE 5(12):e15004
164. Candia J, Daya GN, Tanaka T, Ferrucci L, Walker KA. 2022. Assessment of variability in the plasma 7k
SomaScan proteomics assay. Sci. Rep. 12:17147
165. Candia J, Cheung F, Kotliarov Y, Fantoni G, Sellers B, et al. 2017. Assessment of variability in the
SOMAscan assay. Sci. Rep. 7:14248
166. Eldjarn GH, Ferkingstad E, Lund SH, Helgason H, Magnusson OT, et al. 2023. Large-scale plasma
proteomics comparisons through genetics and disease associations. Nature 622(7982):348–58
167. Assarsson E, Lundberg M, Holmquist G, Björkesten J, Thorsen SB, et al. 2014. Homogenous 96Plex PEA immunoassay exhibiting high sensitivity, specificity, and excellent scalability. PLOS ONE
9(4):e95192
168. Li S, Sullivan NL, Rouphael N, Yu T, Banton S, et al. 2017. Metabolic phenotypes of response to
vaccination in humans. Cell 169(5):862–77.e17
169. Ferdosi S, Tangeysh B, Brown TR, Everley PA, Figa M, et al. 2022. Engineered nanoparticles enable deep
proteomics studies at scale by leveraging tunable nano–bio interactions. PNAS 119(11):e2106053119
170. Viode A, van Zalm P, Smolen KK, Fatou B, Stevenson D, et al. 2023. A simple, time- and cost-effective,
high-throughput depletion strategy for deep plasma proteomics. Sci. Adv. 9(13):eadf9717
171. Phan HV, Tsitsiklis A, Maguire CP, Haddad EK, Becker PM, et al. 2024. Host–microbe multiomic
profiling reveals age-dependent immune dysregulation associated with COVID-19 immunopathology.
Sci. Transl. Med. 16(743):eadj5154
172. Stoeckius M, Hafemeister C, Stephenson W, Houck-Loomis B, Chattopadhyay PK, et al. 2017.
Simultaneous epitope and transcriptome measurement in single cells. Nat. Methods 14(9):865–68
173. Heumos L, Schaar AC, Lance C, Litinetskaya A, Drost F, et al. 2023. Best practices for single-cell analysis
across modalities. Nat. Rev. Genet. 24(8):550–72
174. Mulè MP, Martins AJ, Tsang JS. 2022. Normalizing and denoising protein expression data from dropletbased single cell profiling. Nat. Commun. 13:2099
175. Setliff I, Shiakolas AR, Pilewski KA, Murji AA, Mapengo RE, et al. 2019. High-throughput mapping of
B cell receptor sequences to antigen specificity. Cell 179(7):1636–46.e15
176. Zhang B, Upadhyay R, Hao Y, Samanovic MI, Herati RS, et al. 2023. Multimodal single-cell datasets
characterize antigen-specific CD8+ T cells across SARS-CoV-2 vaccination and infection. Nat. Immunol.
24(10):1725–34
177. Blair JD, Hartman A, Zenk F, Dalgarno C, Treutlein B, Satija R. 2023. Phospho-seq: integrated, multimodal profiling of intracellular protein dynamics in single cells. bioRxiv 2023.03.27.534442. https://
doi.org/10.1101/2023.03.27.534442
178. Swanson E, Lord C, Reading J, Heubeck AT, Genge PC, et al. 2021. Simultaneous trimodal single-cell
measurement of transcripts, epitopes, and chromatin accessibility using TEA-seq. eLife 10:e63632
www.annualreviews.org • Systems Human Immunology and AI
719
Downloaded from www.annualreviews.org. Guest (guest) IP: 74.109.18.71 On: Sat, 17 May 2025 11:55:57
179. Mimitou EP, Lareau CA, Chen KY, Zorzetto-Fernandes AL, Hao Y, et al. 2021. Scalable, multimodal
profiling of chromatin accessibility, gene expression and protein levels in single cells. Nat. Biotechnol.
39(10):1246–58
180. Spitzer MH, Nolan GP. 2016. Mass cytometry: single cells, many features. Cell 165(4):780–91
181. Konecny AJ, Mage P, Tyznik AJ, Prlic M, Mair F. 2023. 50-color phenotyping of the human immune
system with in-depth assessment of T cells and dendritic cells. bioRxiv 2023.12.14.571745. https://doi.
org/10.1101/2023.12.14.571745
182. Vandereyken K, Sifrim A, Thienpont B, Voet T. 2023. Methods and applications for single-cell and
spatial multi-omics. Nat. Rev. Genet. 24(8):494–515
183. Efremova M, Teichmann SA. 2020. Computational methods for single-cell omics across modalities. Nat.
Methods 17(1):14–17
184. Cao Z-J, Gao G. 2022. Multi-omics single-cell data integration and regulatory inference with graphlinked embedding. Nat. Biotechnol. 40(10):1458–66
185. Fernandez R, Maecker H. 2015. Cytokine-stimulated phosphoflow of PBMC using CyTOF mass
cytometry. Bio-Protocol 5(11):e1496
186. Rubin SJS, Bai L, Haileselassie Y, Garay G, Yun C, et al. 2019. Mass cytometry reveals systemic and
local immune signatures that distinguish inflammatory bowel diseases. Nat. Commun. 10:2686
187. Arunachalam PS, Wimmers F, Mok CKP, Perera RAPM, Scott M, et al. 2020. Systems biological assessment of immunity to mild versus severe COVID-19 infection in humans. Science
369(6508):1210–20
188. Piasecka B, Duffy D, Urrutia A, Quach H, Patin E, et al. 2018. Distinctive roles of age, sex, and genetics in shaping transcriptional variation of human immune responses to microbial challenges. PNAS
115(3):E488–97
189. Urrutia A, Duffy D, Rouilly V, Posseme C, Djebali R, et al. 2016. Standardized whole-blood transcriptional profiling enables the deconvolution of complex induced immune responses. Cell Rep.
16(10):2777–91
190. Farmer R, Apps R, Quiel J, Sellers BA, Cheung F, et al. 2022. Multiparameter stimulation mapping
of signaling states in single pediatric immune cells reveals heightened tonic activation during puberty.
bioRxiv 2022.11.14.516371. https://doi.org/10.1101/2022.11.14.516371
191. Turing AM. 1950. I. Computing machinery and intelligence. Mind 59(236):433–60
192. McCarthy J. 1971. The present state of research on artificial intelligence. Turing Award Lecture presented at
26th Annual Conference of the ACM, Chicago, Aug. 3–5
193. Russell S, Norvig P. 2020. Artificial Intelligence: A Modern Approach. London: Pearson. 4th ed.
194. Jia W, Sun M, Lian J, Hou S. 2022. Feature dimensionality reduction: a review. Complex Intell. Syst.
8:2663–93
195. Aqel N, Reusser L, Margreth S, Carminati A, Lehmann P. 2024. Prediction of hysteretic matric potential
dynamics using artificial intelligence: application of autoencoder neural networks. Geosci. Model Dev.
17(18):6949–66
196. Bishop CM. 2006. Pattern Recognition and Machine Learning. New York: Springer
197. Jumper J, Evans R, Pritzel A, Green T, Figurnov M, et al. 2021. Highly accurate protein structure
prediction with AlphaFold. Nature 596(7873):583–89
198. Dauparas J, Anishchenko I, Bennett N, Bai H, Ragotte RJ, et al. 2022. Robust deep learning–based
protein sequence design using ProteinMPNN. Science 378(6615):49–56
199. Frazer J, Notin P, Dias M, Gomez A, Min JK, et al. 2021. Disease variant prediction with deep generative
models of evolutionary data. Nature 599(7883):91–95
200. Cheng J, Novati G, Pan J, Bycroft C, Žemgulytė A, et al. 2023. Accurate proteome-wide missense variant
effect prediction with AlphaMissense. Science 381(6664):eadg7492
201. Cui H, Wang C, Maan H, Pang K, Luo F, et al. 2024. scGPT: toward building a foundation model for
single-cell multi-omics using generative AI. Nat. Methods 21:1470–80
202. Heimberg G, Kuo T, DePianto D, Heigl T, Diamant N, et al. 2023. Scalable querying of human cell
atlases via a foundational model reveals commonalities across fibrosis-associated macrophages. bioRxiv
2023.07.18.549537. https://doi.org/10.1101/2023.07.18.549537
720
Lei • Tsang
Downloaded from www.annualreviews.org. Guest (guest) IP: 74.109.18.71 On: Sat, 17 May 2025 11:55:57
203. Liu T, Li K, Wang Y, Li H, Zhao H. 2024. Evaluating the utilities of foundation models in single-cell
data analysis. bioRxiv 2023.09.08.555192. https://doi.org/10.1101/2023.09.08.555192
204. Greener JG, Kandathil SM, Moffat L, Jones DT. 2022. A guide to machine learning for biologists. Nat.
Rev. Mol. Cell Biol. 23(1):40–55
205. Goodfellow I, Bengio Y, Courville A. 2016. Deep Learning. Cambridge, MA: MIT Press
206. Hinton GE, Salakhutdinov RR. 2006. Reducing the dimensionality of data with neural networks. Science
313(5786):504–7
207. LeCun Y, Bengio Y, Hinton G. 2015. Deep learning. Nature 521(7553):436–44
208. Matsuura T, Fukushima W, Nakagama Y, Kido Y, Kase T, et al. 2024. Factors impacting antibody
kinetics, including fever and vaccination intervals, in SARS-CoV-2-naïve adults receiving the first four
mRNA COVID-19 vaccine doses. Sci. Rep. 14:7217
209. Wang XQ, Danenberg E, Huang C-S, Egle D, Callari M, et al. 2023. Spatial predictors of
immunotherapy response in triple-negative breast cancer. Nature 621(7980):868–76
210. Pérez-Alós L, Hansen CB, Almagro Armenteros JJ, Madsen JR, Heftdal LD, et al. 2023. Previous
immunity shapes immune responses to SARS-CoV-2 booster vaccination and Omicron breakthrough
infection risk. Nat. Commun. 14:5624
211. Khoury DS, Cromer D, Reynaldi A, Schlub TE, Wheatley AK, et al. 2021. Neutralizing antibody levels are highly predictive of immune protection from symptomatic SARS-CoV-2 infection. Nat. Med.
27(7):1205–11
212. Gygi JP, Maguire C, Patel RK, Shinde P, Konstorum A, et al. 2024. Integrated longitudinal multiomics study identifies immune programs associated with acute COVID-19 severity and mortality.
J. Clin. Investig. 134(9):e176640
213. Aicher TP, Carroll S, Raddi G, Gierahn T, Wadsworth MH, et al. 2019. Seq-Well: a sample-efficient,
portable picowell platform for massively parallel single-cell RNA sequencing. In Single Cell Methods, ed.
V Proserpio, pp. 111–32. Methods Mol. Biol. 1979. New York: Humana
214. Phelan JE, O’Sullivan DM, Machado D, Ramos J, Oppong YEA, et al. 2019. Integrating informatics
tools and portable sequencing technology for rapid detection of resistance to anti-tuberculous drugs.
Genome Med. 11:41
215. Viode A, van Zalm P, Smolen KK, Fatou B, Stevenson D, et al. 2023. A simple, time- and cost-effective,
high-throughput depletion strategy for deep plasma proteomics. Sci. Adv. 9(13):eadf9717
216. Yale Sch. Med. 2023. Yale joins Chan Zuckerberg Initiative Biohub NY. Press Release, Yale Sch. Med.,
New Haven, CT, Oct. 18. https://medicine.yale.edu/news-article/yale-joins-chan-zuckerberginitiative-biohub-ny-with-goal-of-bioengineering-immune-cells-for-early-disease-detectionand-prevention
217. Dixon TA, Williams TC, Pretorius IS. 2021. Sensing the future of bio-informational engineering. Nat.
Commun. 12:388
218. Nguyen E, Poli M, Durrant MG, Thomas AW, Kang B, et al. 2024. Sequence modeling and design from
molecular to genome scale with Evo. bioRxiv 2024.02.27.582234. https://doi.org/10.1101/2024.02.
27.582234
219. Lin Z, Akin H, Rao R, Hie B, Zhu Z, et al. 2023. Evolutionary-scale prediction of atomic-level protein
structure with a language model. Science 379(6637):1123–30
220. Verkuil R, Kabeli O, Du Y, Wicky BIM, Milles LF, et al. 2022. Language models generalize beyond
natural proteins. bioRxiv 2022.12.21.521521. https://doi.org/10.1101/2022.12.21.521521
221. Theodoris CV, Xiao L, Chopra A, Chaffin MD, Al Sayed ZR, et al. 2023. Transfer learning enables
predictions in network biology. Nature 618(7965):616–24
222. Hao M, Gong J, Zeng X, Liu C, Guo Y, et al. 2024. Large-scale foundation model on single-cell
transcriptomics. Nat. Methods 21:1481–91
223. Theodoris CV, Xiao L, Chopra A, Chaffin MD, Al Sayed ZR, et al. 2023. Transfer learning enables
predictions in network biology. Nature 618(7965):616–24
224. Chen RJ, Ding T, Lu MY, Williamson DFK, Jaume G, et al. 2024. Towards a general-purpose foundation
model for computational pathology. Nat. Med. 30(3):850–62
225. Rudin C. 2019. Stop explaining black box machine learning models for high stakes decisions and use
interpretable models instead. Nat. Mach. Intell. 1(5):206–15
www.annualreviews.org • Systems Human Immunology and AI
721
Downloaded from www.annualreviews.org. Guest (guest) IP: 74.109.18.71 On: Sat, 17 May 2025 11:55:57
226. Petch J, Di S, Nelson W. 2022. Opening the black box: the promise and limitations of explainable
machine learning in cardiology. Can. J. Cardiol. 38(2):204–13
227. Jordan MI, Mitchell TM. 2015. Machine learning: trends, perspectives, and prospects. Science
349(6245):255–60
228. Vodovotz Y. 2023. Towards systems immunology of critical illness at scale: from single cell ’omics to
digital twins. Trends Immunol. 44(5):345–55
229. Vodovotz Y, Xia A, Read EL, Bassaganya-Riera J, Hafler DA, et al. 2017. Solving immunology? Trends
Immunol. 38(2):116–27
230. Handel A, La Gruta NL, Thomas PG. 2020. Simulation modelling for immunologists. Nat. Rev.
Immunol. 20(3):186–95
231. Gillespie DT. 2007. Stochastic simulation of chemical kinetics. Annu. Rev. Phys. Chem. 58:35–55
232. Park K, Prüstel T, Lu Y, Tsang JS. 2019. Machine learning of stochastic gene network phenotypes.
bioRxiv 825943. https://doi.org/10.1101/825943
233. Wong HS, Park K, Gola A, Baptista AP, Miller CH, et al. 2021. A local regulatory T cell feedback circuit
maintains immune homeostasis by pruning self-activated T cells. Cell 184(15):3981–97.e22
722
Lei • Tsang