PII-124 - APPLICATION OF MACHINE LEARNING (ML) ON QSP VIRTUAL PATIENT POPULATION (VPOP) FOR IDENTIFICATION OF BIOMARKERS OF RESPONSE AND FOR SELECTION OF PRE-CONDITIONING REGIMEN FOR ENGINEERED T-CELL (ET-CELL) THERAPY.

S. Shah¹, O. Demin Jr², G. Kolesova², L. Cucurull- Sanchez³; ¹GlaxoSmithKline, Waltham, MA, USA, ²InSysBio, Moscow, Russia, ³GlaxoSmithKline, Stevenage, Hertfordshire, United Kingdom.

Background: Clinical studies of eT-cells have unique challenges owing complex pharmacology and limited patient enrollment. QSP model generated by combining the disease mechanistic knowledge and pharmacological information can simulate Vpop (1). We demonstrate the use of ML on Vpop to identify biomarkers of response when lymphodepletion regimens are: 1) high, or 2) high and low.

Methods: ML analysis was performed on 2 data types: 1) one cohort of 298 & 2) two cohorts of 298 Synovial Sarcoma plausible virtual patients (VPs), generated from an existing QSP model of NYESO1 eT-cell therapy. The VP datasets were processed in R. Baseline and longitudinal biomarker-derived parameters were used as ML input descriptors, and RECIST score response was used ML classifier endpoint. The processed dataset was split into training (75%) and test (25%) sets. We applied supervised regression ML algorithms that included decision tree, support vector machines, k-nearest neighbors, and linear regression. ML was carried on Weka (v 3.9.6). 10-fold cross validation was used on training set to avoid over-fitting. Model performance was evaluated by determining the prediction accuracy and precision of the model on test set.

Results: The decision tree model had the best performance for both objectives. Within a regimen, the model predicted, higher levels of week 4 GMCSF was associated with response (Accuracy: 71%, Sen = 0.69, Spe = 0.74, PPV = 0.73, NPV = 0.7). Within two lymphodepletion regimens, lower absolute lymphocyte count in blood on day of eT-cell infusion was associated with response (Accuracy: 72%, Sen = 0.64, Spe = 0.8, PPV = 0.78, NPV = 0.69).

Conclusion: ML applied to QSP Vpop can perform population level exploratory analysis when there is limited clinical data as part of the MID3 framework and design of future clinical study.

1. Cheng, Y., Straube, R., Alnaif, A.E., Huang, L., Leil, T.A., Schmidt, B.J. (2022). Virtual Populations for Quantitative Systems Pharmacology Models. In: Bai, J.P., Hur, J. (eds) Systems Medicine. Methods in Molecular Biology, vol 2486. Humana, New York, NY.