Real world prediction and biological characterization of sotorasib sensitivity using multimodal AI and liquid biopsy genomic inputs

Real world prediction and biological characterization of sotorasib sensitivity using multimodal Al and liquid biopsy genomic inputs

Maayan Baron $ ^{1} $ , Momeneh Foroutan $ ^{1} $ , Sunil Kumar $ ^{1} $ , Zong Miao $ ^{1} $ , Ripple Khera $ ^{1} $ , Felicia Kuperwaser $ ^{1} $ , Dillon Tracy $ ^{1} $ , Kevin Freisen $ ^{1} $ , Brandon Funkhouser $ ^{1} $ , Nathaniel Tann $ ^{1} $ , Fahad Khan $ ^{1} $ Nick Lee $ ^{1} $ , Sean Klei $ ^{1} $ , Jordan Wolinsky $ ^{1} $ , Kevin Shah $ ^{2} $ , Aaron Hardin $ ^{2} $ , Jean-François Martini $ ^{2} $ , Jean-Michel Rouly $ ^{1} $ , Allen Chao $ ^{1} $ , Anshu Jain $ ^{1} $ , Jeff Sherman $ ^{1} $ Emily Vucic $ ^{1} $

1. Zephyr AI, McLean, VA
2. Guardant Health, Palo Alto, CA

Heterogeneous Response to KRAS $ ^{G12C} $ Inhibitors Highlights the Need for Functional Biomarkers Beyond Mutation Status

• KRAS $ ^{G12C} $ inhibitors (e.g., sotorasib) have expanded treatment options in NSCLC, yet clinical benefit remains heterogeneous and often short-lived.

• KRAS $ ^{G12C} $ mutation status alone is insufficient to explain response heterogeneity, suggesting that functional KRAS dependency is not fully captured by current companion diagnostics.

• Reduced benefit from KRAS inhibition can arise through: MAPK pathway reactivation; alternative RTK-driven signaling; tumor plasticity; and EMT-associated states.

• These functional tumor states are not effectively captured by single-gene biomarkers

• Liquid biopsy NGS is increasingly used at treatment initiation and longitudinally in clinical trials and routine care, creating an opportunity to infer functional drug sensitivity from routine clinical data across the patient journey.

• Multimodal AI approaches that capture complex tumor biology relevant to drug response from clinically available inputs, may enable improved and streamlined patient stratification and more precise deployment of KRAS and other therapies.

Study Objectives:

• Evaluate whether a biologically interpretable multimodal AI model, AIM-Bx (Figure 1), can predict real-world benefit from sotorasib using minimal clinical features, commercially available liquid biopsy DNA inputs and drug structure properties in NSCLC (Table 1).

Clinically Available Inputs

• ctDNA-derived genomic inputs from the Guardant360 CDx liquid biopsy assay

Drug Response Supports prospective & retrospective real-world clinical validation

Explainable Drug Response Predictions

For each patient & drug, AIM-Bx predicts...

AI-enabled Stratification Using ctDNA Inputs Identifies Clinically Meaningful Benefit From Sotorasib in Real-World NSCLC patients

Predicted Network- and Program-Level Tumor States Define Sensitivity to KRAS Inhibition Beyond KRAS $ ^{G12C} $ Dependency

Figure 2. AIM-Bx, stratifies real-world NSCLC patients by clinical response to sotorasib. (A-B) Kaplan-Meier analyses of real-world progression-free survival (rwPFS) and overall survival (rwOS) in sotorasib-treated NSCLC patients stratified by model-predicted sensitivity. Predicted-sensitive patients demonstrated improved rwPFS (14 vs. 4 months) and rwOS (14 vs. 7 months; log-rank p < 0.05). (C-D) Univariable Cox analyses for rwPFS and rwOS adjusting for clinical and molecular variables available at treatment initiation. Model-predicted sensitivity was the only variable significantly associated with improved outcomes. (E) Aggregate feature importance highlighting genomic alterations contributing to model

A

0.000 0.005 0.010 0.015 0.020 0.025 Feature Importance Score

Figure 3. Functional KRAS dependency does not fully explain sensitivity to KRAS inhibitors in preclinical models. (A) Observed drug sensitivity (AUC) versus CRISPR-derived KRAS dependency in cancer cell lines treated with KRAS G12C inhibitors, colored by KRAS mutation status. (B) Predicted sensitivity versus KRAS dependency across test-set and additional cell lines, demonstrating limited association between predicted response and single-gene KRAS dependency. For both plots: green = KRAS $ ^{\mathrm{WT}} $ ; red = KRAS $ ^{G12C} $ ; blue = other KRAS mutation.

Preclinical analyses sensitivity to KRAS in independent of CRISPR-d

demonstrated that hibitors was largely fined KRAS

Figure 4. Distinct predicted biological states associated with differential sotorasib sensitivity. (A) Vulnerability Network (VN) analysis of predicted-sensitive tumors showing KRAS-centered subnetworks enriched for MAPK signaling, cell-cycle control, metabolism, and apoptosis. (B) VN analysis of predicted less-sensitive tumors highlighting alternative signaling programs, including adhesion, cell polarity, vesicle trafficking, RTK signaling, and stress-response pathways. (C) GSEA of reconstructed expression profiles showing enrichment of proliferative and metabolic programs in predicted-sensitive tumors, and EMT, extracellular matrix remodeling, angiogenesis, and RTK-driven signaling in predicted less-sensitive tumors (nomininal $ p < 0.05 $).

Conclusion: Multimodal AI Identifies Functional Target Dependence Beyond Genomic Biomarkers Using Routine Clinical Inputs

Multimodal AI-enabled software operating on routine clinical data enables biologically grounded, scalable stratification for sotorasib and other KRAS inhibitors, with broader applicability to targeted therapies.