CAV 2014 Finding Instability in Biological Models Byron Cook1,2, Jasmin Fisher1,3, Benjamin A. Hall1, Samin Ishtiaq1, Garvit Juniwal4, Nir Piterman5 1Microsoft Research 3University of Cambridge 5University of Leicester 2University 4UC College London Berkeley 1 Biological Signaling Networks • Network of chemicals (usually proteins) regulating each others’ concentration, evolving the overall state. • Qualitative Networks offer a good level of abstraction for modeling and analysis. [BMC Syst Biol ’07] Schaub et. al. • Have been applied to blood cell differentiation, skin homeostasis and cancer cell development. 2 Homeostasis • Homeostasis (or stability) is a natural requirement for most of these systems. It represents the ability to stay at robust equilibrium. • Model of a naturally occurring phenomenon stability desired. starting from every state, the same stable state reached. • Required sanity check during development stage. • Central problem: Enabling biologists to be able to quickly check stability and examine instability. 3 Bio Model Analyzer (BMA) http://biomodelanalyzer.research.microsoft.com/ [CAV’12] Visual Tool for Modeling and Analyzing Biological Networks. Benque et. al. 4 Contribution • Previous: BMA’s stability algorithm from [VMCAI’11 Cook et. al.] • Problem: takes too long (~ 2 hours) for certain classes of models. • This work: divide/conquer based algorithm to tackle hard cases. • Benefit: Up to 3 orders of magnitude of speed up. Can solve all existing models in matter of seconds making BMA truly interactive. 5 Outline • Definitions – Qualitative Networks and stability • Summary of previous algorithm and its issues • New divide/conquer algorithm • Evaluation and conclusion 6 Qualitative Networks (QNs) • Variables X, Y, Z ∈ 0, 𝑁 𝐹: 0, 𝑁 × 0, 𝑁 → [0, 𝑁] • finite domain – [0, N] • Dependencies • Target functions • Z moves towards F(X,Y) in increments/decrements of 1 • Typically monotonic • Synchronous updates • Unique next state X Y Z Z′:= 𝑍 + 1 𝑖𝑓 𝐹 𝑋, 𝑌 > 𝑍 𝑍 𝑖𝑓 𝐹 𝑋, 𝑌 = 𝑍 𝑍 − 1 𝑖𝑓 𝐹 𝑋, 𝑌 < 𝑍 7 Typical Size of a QN • • • • Variables (|V|) -- 50 Domain size (N+1) – 4 Average in-degree – 3 State space – 450 8 Stabilization Behavior in QNs (1/2) For a QN with variables V • State Space ∑ = {0, …, N}|V| • Transition Function δ: ∑→∑ (defined via target functions) • All states are initial • A state s said to be recurring if it is possible to reach from s to itself with finite applications of δ. • Unique recurring state Stabilizing QN 9 Stabilization Behavior in QNs (2/2) State Stable Self-loop state Cycle Unstable Cycle (length >=2) Cyclic Instability (CI) Multiple self-loops Bifurcation (BF) 10 Proving Stability [VMCAI’11 Proving Stabilization of Biological Systems. Cook et. al.] • Goal: compute the set of recurring states • Idea: over-approximate via rectangular abstraction • If over-approximation has single state - STABLE • Else: 1) Check for multiple self-loop states by encoding to a SAT query 2) Check for cycles of increasing length starting at 2 (BMC) 3) If no cycles found for a length > diameter – STABLE (using a naïve over-approximation of diameter) 12 Over-Approximation Technique • Keep track of an interval per variable • Start with the set of all states 0, 𝑁 × ⋯ × 0, 𝑁 • For a variable, given the current intervals of its inputs and the target function, is it possible to tighten its interval? X Y [0,2] [0,9] X Y [0,2] [1,3] 𝐹(𝑌) = 𝑋 + 1 • Order of updating variables doesn’t matter Pick arbitrarily from a work list. 13 Pitfalls A 0 1 2 B 0 1 2 3 Trivially Stable (TS) Bifurcation (BF) SAT !! Cyclic Instability (CI) Computationally prohibitive to distinguish these two cases Non-trivially Stable (NTS) 14 Main Contribution: Finding Instability • New instability finding algorithm to distinguish between Cyclic Instability (CI) and Non-Trivial Stability (NTS) • Rectangular abstraction. A Cartesian product of intervals is called a region. For example, [1, 4] × [0, 2] • Uses two new generic procedures SHRINK and CUT. 15 SHRINK • SHRINK: region region. Given a region ρ, it returns another region ρ’ contained in ρ s.t. all cycles and self-loops in ρ are within ρ’ A 0 1 2 B 0 1 2 3 ρ ρ' • The previous interval update technique is one way to implement SHRINK • The old algorithm can be thought of as a single application of SHRINK, now we use it within a recursive procedure. 16 CUT • A cut of a region ρ is a pair (ρ1, ρ2) of regions s.t. ρ1∪ρ2 = ρ. • A frontier of a cut (ρ1, ρ2) is a pair of sets of states around the cut. • A frontier can be two/one/zero-way. ρ frontier (two-way) ρ2 ρ1 cut • CUT: region cut × frontier. 17 FINDINSTABILITY // distinguish between CI and NTS // returns either stable or a cycle FINDINSTABILITY(ρ): ρ SHRINK(ρ) if ρ contains single state then return stable else (ρ1, ρ2) CUT(ρ) res1 FINDINSTABILITY(ρ1) if res1 is cycle then return res1 res2 FINDINSTABILITY(ρ2) if res2 is cycle then return res2 return (cycle) FINDCYCLEACROSSCUT(ρ1, ρ2) 18 Concrete Implementation of SHRINK • Reduce region while retaining all contained cycles; update intervals • Issue: Due to cuts, might get outgoing transitions causing intervals to grow • Fix: Change target functions by limiting with a min/max value. • TB’ = max(min(TB, 1), 0) B 0 1 2 3 19 Concrete Implementation of CUT • Split the interval of one of the |V| variables. • Variables change by at most 1 the slice around the split point is a frontier. • Enumerate over N*|V| choices until a zero/one-way frontier is found because no loops can exist across such frontiers. • Zero/one-way property can be checked via a SAT query. cut 20 FINDCYCLEACROSSCUT • If the frontier is zero/one-way Return none • Else enumerate over states in the frontier and run simulations of the transition function. Stop when a cycle of length >= 2 is found. • Works well because • SHRINK is effective in reducing the search space. • One-way cuts are prevalent due to simple and monotonic target functions. CYCLE!! 21 Examples one-way frontier ` Non-trivially Stable (NTS) Cyclic Instability (CI) 22 Benchmarks and Evaluation Model Variables Depend N+1 BMA(old) encies (ms) BMA(new) (ms) Speed up Dicty Population 35 71 2 60066 2514 23.6x Firing Neuron 21 25 21 105 2 4 218 43934 458 9865 0.5x 4.5x SSkin 1D 57 40 92 46 3 5 4497 T.O. 446 132350 10.1x >6.8x SSkin 2D 2 layers 40 64 5 T.O. 2706 >322.6x Ion Channel 10 7 2 499 173 2.9x Lambda Phage 8 13 2 3113 197 15.8x Resting Neuron 21 28 2 T.O. 244949 >3.7x E. Coli Chemotaxis 9 10 5 T.O. 250 >3600x L Model Leukemia T.O. = 15 minutes http://www.cs.le.ac.uk/people/npiterman/publications/2014/instability/ 23 Conclusion • SHRINK and CUT work effectively for systems under consideration. • Running time down from hours to milliseconds • The added capabilities make the tool clinically relevant for industrial biomedicine. • Found stability and instability results for previously untractable and biologically important models: • bacterial chemotaxis • dictyostelium discoideum 24 Thank You! 25