Using the Gamma Distribution to Determine Anti-drug antibody Screening Assay Cut Points with Non-normal Data for Midwest Biopharmaceutical Statistics Workshop May 24-26, 2010 Brian Schlain Nonclinical Statistics ANTI-DRUG ANTIBODY (ADA) TESTING high volume? yes no screening assay yes confirmation assay positive? yes positive? titration assay output titer NEUTRALIZING ANTIBODY ASSAY (NAB) TESTING Anti-drug Antibodies (ADA) detected? yes yes high volume? no titration assay/ output titer screening assay yes positive? Types of Immunogenicity Assays Anti-Drug Antibody (ADA) Testing with Binding Assays Radioimmunoprecipitation Assay (RIPA) Enzyme-Linked Immunosorbent Assay (ELISA) *Standard sandwich ELISA *Bridging ELISA *Electrochemiluminescence (ECL; IGEN) Optical Sensor-based * Surface Plasmon Resonance (SPR; Biacore) *Guided Mode Resonance Filter (BIND; BD Biosci) Neutralizing Antibody (NAB) Testing with Blocking Assays Related to drug mechanism of action Converted PK assays Cell based: Some of the reagents genetically engineered and stored as cell lines Example ADA Assay Positive Control Response Curve Example NAB Assay Positive Control Response Curve 8×12 Example Screening/Confirmation Assay Plate Layout 1 2 3 4 5 6 7 8 9 10 11 12 A 1S 1C 5S 5C NC LPC 9S 9C 13S 13C 17S 17C B 1S 1C 5S 5C NC LPC 9S 9C 13S 13C 17S 17C C 2S 2C 6S 6C NC LPC 10S 10C 14S 14C 18S 18C D 2S 2C 6S 6C NC HPC 10S 10C 14S 14C 18S 18C E 3S 3C 7S 7C NC HPC 11S 11C 15S 15C 19S 19C F 3S 3C 7C 7C NC HPC 11S 11C 15S 15C 19S 19C G 4S 4C 8S 8C NC HPC -C 12S 12C 16S 16C 20S 20C H 4S 4C 8S 8C NC HPC -C 12S 12C 16S 16C 20S 20C Screening Assay Cut Points Fixed: mean and sd do not change across plates. – CP=meanX ± a×sdX – X=Sample assay signal response Floating: mean changes, but sd remains constant across plates. – Multiplicative: CF=meanY ± a×sdY Y=X/NC CP=CF×NC – Additive: CF=meanZ ± a×sdZ Z=X – NC CP=CF + NC Dynamic: mean and sd both change with every plate. Screening Multiplicative Cut Point CF Determination • Screen data for outliers • Normalize:Y= Sample/NC non-normal Model distribution of normalized values normal or lognormal • Gamma • Weibull Estimate CF • Nonparametric percentile Validation Phase CP=CF×NC ADA Screening Assay Cut Point CF Estimation under Normal Theory CUT POINT RULE sample>cut point, judge screening positive. cut point=CF avg(NC). NC: pooled negative sera control NORMAL THEORY CORRECTION FACTOR (CF) ESTIMATION (Guttman I, Statistical Tolerance Regions: Classical and Bayesian, 1970) n negative sera samples. R (or normalized value)=Sample/Avg(NC). mean (MR) and sd (SR) of R. CF=MR + tdf,α SR×(1 + 1/n)0.5 no transformation . CF=exp{MR + tdf,α SR×(1 + 1/n) 0.5 } log transformation. df=n – 1. α=targeted false positive rate=0.05. NAB Screening Assay Cut Point CF Estimation under Normal Theory CUT POINT RULE sample < cut point, judge screening positive. cut point=CF avg(NC). NC: pooled negative sera control NORMAL THEORY CORRECTION FACTOR (CF) ESTIMATION (Guttman I, Statistical Tolerance Regions: Classical and Bayesian, 1970) n negative sera samples. R (or normalized value)=Sample/Avg(NC). mean (MR) and sd (SR) of R. CF=MR - tdf,α SR×(1 + 1/n)0.5 no transformation . CF=exp{MR - tdf,α SR×(1 + 1/n) 0.5 } log transformation. df=n – 1. α=targeted false positive rate=0.05. 2-Parameter Gamma Density Functions Gamma approaches normal as k increases. 2-parameter Gamma 3-parameter Gamma f ( x | k , , s ) ( x s) k 1 exp[( s x ) / ] ( k ) k x s (threshold parameter) Some common gamma fitting methods: maximum likelihood estimation (MLE) -SAS UNIVARIATE, R OPTIM. -becomes unstable when k approaches 1 method of moments (ME) modified moment estimation (MME) (Cohen and Whitten) Maximum Likelihood Estimation (MLE) of 2parameter Gamma Unstable when estimate of k close to 1. ADA distributions tend to be unimodal with k>1. NAB distributions tend to be unimodal with k<1. -Fit gamma to 1/R=Avg(NC)/Sample. Estimation of Gamma Parameters by Method of Moments (ME) Gamma central moment Empirical central moment Mean S + θk =m1 (mean) Variance θ2k =m2 (variance) Third standard moment 3 2 k =m3 /(m2 )3/2 (std. skewness) ME for 3-parameter Gamma Gamma parameter Moment estimator k (shape) =4m23 /m3 θ (scale) =m3 / (2m2 ) 2 s (threshold) =m1 - (2m22 )/m3 Modified Moment Estimators (MME) for 3-parameter Gamma (using SAS PROBGAM function) Cohen AC, Whitten BJ, Modified moment estimation for the 3-parameter Gamma distribution, J. of Qual. Tech., 18, 1, 1986, 53-62 G ( X 1 | sˆ , ˆ , kˆ ) 1 n 1 G (W 1 | s 2 / ˆ 3 ; ˆ 3 / 2 ; k 4 / ˆ 3 ) PROBGAM 2 U 1 (W 1 sˆ ) / ˆ (W 1 2 / ˆ 3 ) /( ˆ 3 / 2 ) W X mean sd W 1 W 2 ... W n ˆ sd ˆ 3 / 2 sˆ mean 2 sd / ˆ 3 2 kˆ 4 / ˆ 3 (U 1 , 4 / ˆ 3 ) 2 1 n 1 Calculating the Cut Point Correction Factor as a Gamma Percentile Cut point correction factor (CF) calculated using SAS GAMINV function: CF ˆ GAMINV ( P , kˆ ) sˆ P=.95 if false positive rate targeted at 5% (ADA assay) -NAB assay: Model reciprocal of normalized values so that skewness > 0. θ=scale parameter s=threshold parameter k=shape parameter CP=CF×NC 3-parameter Gamma Estimation MLE preferred to MME or ME. – SAS UNIVARIATE – R OPTIM MME generally better than ME. MLE unstable for k near 1. (Johnson and Kotz recommend k>2.5). MME can be calculated for any k. MME comparable to MLE with increasing n or α3 (Cohen and Whitten). For 3 2 k < 0.10, consider normal distribution (Cohen and Whitten). Standard Gamma Simulations Comparing MLE and Nonparametric Percentile Cut Point Estimators (targeted false positive rate=5%) Shape n Param. 2 30 60 120 240 4 30 60 120 240 8 30 60 120 240 16 30 60 120 240 MLE se 5.085 4.827 4.794 4.786 8.010 7.856 7.834 7.829 13.458 13.314 13.289 12.278 23.564 23.390 23.363 23.355 1.029 .557 .379 .264 1.206 .695 .487 .335 1.469 . 888 .610 .427 1.862 1.128 .780 .545 FP rate .050 .051 .050 .049 .054 .051 .050 .049 .053 .050 .049 .048 .051 .048 .047 .046 se .037 .023 .016 .011 .036 .023 .016 .011 .035 .023 .015 .011 .033 .022 .015 .010 Nonpar. Est. 4.719 4.744 4.739 4.743 7.701 7.750 7.754 7.752 13.085 13.153 13.141 13.147 23.005 23.089 23.106 23.105 se .924 .648 .471 .337 1.161 .799 .582 .415 1.482 1.043 .735 .532 1.920 1.308 .958 .684 FP rate .065 .057 .054 .052 .065 .057 .053 .052 .065 .056 .053 .052 .065 .056 .053 .051 se .043 .028 .020 .014 .044 .028 .020 .014 .044 .029 .020 .014 .044 .028 .020 .014 Abbreviations: n=sample size; η=gamma shape parameter; MLE=maximum likelihood estimator; se=standard error; FP rate=false positive rate; Nonpar. Est.=nonparametric estimator. Standard Gamma Simulations Comparing MLE with Normal Based Cut Point Estimators (targeted false positive rate=5%) Shape n Param. 2 4 8 16 30 60 120 240 30 60 120 240 30 60 120 240 30 60 120 240 MLE se FP rate se 5.085 4.827 4.794 4.786 8.010 7.856 7.834 7.829 13.458 13.314 13.289 12.278 23.564 23.390 23.363 23.355 1.029 .557 .379 .264 1.206 .695 .487 .335 1.469 .888 .610 .427 1.862 1.128 .780 .545 .050 .051 .050 .049 .054 .051 .050 .049 .053 .050 .049 .048 .051 .048 .047 .046 .037 .023 .016 .011 .036 .023 .016 .011 .035 .023 .015 .011 .033 .022 .015 .010 Normal zmethod 4.272 4.301 4.306 4.322 7.234 7.259 7.273 7.283 12.604 12.628 12.641 12.650 22.525 22.547 22.569 22.584 se FP rate se .640 .466 .332 .237 .808 .572 .412 .286 1.044 .734 .517 .367 1.357 .948 .677 .477 .083 .077 .074 .072 .079 .073 .071 .069 .074 .069 .067 .066 .070 .066 .064 .062 .041 .028 .020 .014 .039 .026 .019 .013 .038 .025 .017 .012 .036 .024 .017 .012 Normal predict. interval 4.385 4.357 4.334 4.322 7.395 7.338 7.312 7.302 12.834 12.740 12.697 12.677 22.851 22.706 22.648 22.623 se FP rate se .662 .473 .335 .237 .834 .581 .415 .287 1.076 .745 .521 .369 1.397 .961 .681 .478 .076 .073 .072 .072 .072 .070 .069 .068 .067 .066 .065 .065 .062 .062 .062 .061 .039 .027 .019 .014 .037 .026 .018 .013 .035 .025 .017 .012 .033 .023 .016 .012 Abbreviations: n=sample size; η=gamma shape parameter; MLE=maximum likelihood estimator; se=standard error; FP rate=false positive rate; normal z-method= mean +1.645×sd; Normal Predict. Interval=upper limit of a 1-sided prediction interval based on normal theory. Standard Gamma Distribution Simulations Comparing MLE with Log-normal Cut Point Estimators (targeted false positive rate=5%) Histogram of ADA ELISA Trial Pre-dose Screening Assay Normalized Values (n=175) outlying values Sample #29 Sample #34 MME Gamma Q-Q Plot (with outlying sample 34 excluded) outlying value Sample # 29 MME Gamma Q-Q Plot (with outlying samples 29 and 34 excluded) Histogram of ADA Elisa Trial Pre-dose Normalized Values (with outlying samples 29 and 34 excluded) skewed to the right SAS UNIVARIATE EDF Goodness of Fit Test p-values (outlying samples 29 and 34 excluded) test gamma normal log-normal Shapiro-Wilk - <.0001 .031 KolmogorovSmirnov .146 .010 .025 Cramer-von Mises .089 .0050 .073 AndersonDarling .151 .0050 .060 Gamma Parameter and CF Estimates (outlying samples 29 and 34 excluded) MME MLE mean (of norm. values) 1.270 sd (of norm. values) 0.170 skewness (of norm. values) 0.902 W1 (smallest stand. value) -2.102 n 173 α3 .474 .727 θ (scale) .040 .061 s (threshold) .553 .809 k (shape) 17.796 7.578 CF (upper 3.9% percentile of gamma) 1.60 1.60 CP=CF×NC target upper gamma percentile= 5% - 100%×(2/175)=3.9% Gamma Distribution CF Determination Target false positive (fp) rate=5% Estimated percentage of outlying samples is 1.1% (=100%×2/175) Target fp rate – percentage of outlying samples =5% - 1.1%=3.9% CF=upper 3.9% percentile of fitted gamma=1.60 – CF=θ×GAMINV(p,k) + s=1.60. p=1- 0.039=.961; k= 17.796; θ=.040; s=.553. CP=CF×NC Empirical fp rate=5.7% (=100%×10/175) References Schlain B et al., A novel gamma-fitting statistical method for anti-drug antibody assays to establish assay cut points for data with non-normal distribution, JIM, V. 352, Issues 1-2, 31Jan. 2010, pp. 161-168. Guttman I, Statistical Tolerance Regions: Classical and Bayesian, 1970). Cohen AC, Whitten BJ, Modified moment estimation for the 3parameter Gamma distribution, J. of Qual. Tech., 18, 1, 1986, 53-62. Cohen AC, Whitten BJ, Modified moment and maximum likelihood estimators for parameters of the three-parameter gamma distribution, commun. Statist.-Simula. Computa., 11(2), 197-216 (1982). Bowman KO and Shenton LR, Properties of Estimators for the Gamma Distribution, Marcel Dekker, 1988. Johnson NL and Kotz S, continuous Univariate Distributions, Vol. 1, Houghton Mifflin company, 1970. Krishnamoorthy K, Mathew T, and Mukherjee S, Normal-based methods for a gamma distribution: prediction and tolerance intervals and stress-strength reliability, Technometrics, Vol. 50, No. 1, pp. 6978. BACKUP SLIDES Standard Gamma Simulations Comparing MLE, Nonparametric, and WH Cut Point Estimators (targeted false positive rate=5%) Shape n Param. MLE se FP rate se Nonpar. Est. se FP rate se 2 5.085 4.827 4.794 4.786 8.010 7.856 7.834 7.829 13.458 13.314 13.289 12.278 23.564 23.390 23.363 23.355 1.029 .557 .379 .264 1.206 .695 .487 .335 1.469 . 888 .610 .427 1.862 1.128 .780 .545 .050 .051 .050 .049 .054 .051 .050 .049 .053 .050 .049 .048 .051 .048 .047 .046 .037 .023 .016 .011 .036 .023 .016 .011 .035 .023 .015 .011 .033 .022 .015 .010 4.719 4.744 4.739 4.743 7.701 7.750 7.754 7.752 13.085 13.153 13.141 13.147 23.005 23.089 23.106 23.105 .924 .648 .471 .337 1.161 .799 .582 .415 1.482 1.043 .735 .532 1.920 1.308 .958 .684 .065 .057 .054 .052 .065 .057 .053 .052 .065 .056 .053 .052 .065 .056 .053 .051 .043 .028 .020 .014 .044 .028 .020 .014 .044 .029 .020 .014 .044 .028 .020 .014 4 8 16 30 60 120 240 30 60 120 240 30 60 120 240 30 60 120 240 WH (cube root) 4.910 4.803 4.746 4.730 7.975 7.851 7.789 7.765 13.447 13.291 13.214 13.180 23.487 23.274 23.190 23.152 se FP rate se .738 .506 .353 .249 .918 .632 .443 .306 1.174 .802 .560 .393 1.503 1.024 .723 .505 .051 .052 .052 .052 .051 .051 .051 .051 .050 .050 .050 .050 .050 .050 .050 .050 .031 .021 .015 .011 .030 .021 .015 .010 .030 .021 .015 .010 .030 .021 .015 .010 Abbreviations: n=sample size; η=gamma shape parameter; MLE=maximum likelihood estimator; se=standard error; FP rate=false positive rate; Nonpar. Est.=nonparametric estimator; WH=Wilson-Hilferty estimator. Need for further research on WH How well does it perform when the real distribution is not quite a gamma, but the gamma is the best approximation that can be found?