Abstract 5235: Real world prediction and biological characterization of sotorasib sensitivity using multimodal AI and liquid biopsy genomic inputs

Abstract Background: KRAS G12C inhibitors such as sotorasib have expanded options for NSCLC, but clinical benefit is variable and no validated biomarkers beyond the KRAS G12C mutation exist. Reported resistance mechanisms include MAPK reactivation, co-occurring alterations, lineage plasticity, and TME programs. Liquid biopsy is widely used for treatment selection— including Guardant360 CDx, the FDA-approved test for sotorasib—motivating computational approaches that operate on standard clinical assays. AIM-Bx is a multimodal, biologically interpretable AI model that predicts small-molecule response from routine inputs and outputs drug-response predictions, Vulnerability Networks (predicted CRISPR perturbation sensitivities), and reconstructed expression profiles. We evaluated whether AIM-Bx could predict real-world sotorasib benefit using commercial liquid biopsy data. Methods: NSCLC patients with KRAS G12C mutations treated with sotorasib (n = 39) were identified using Guardant360 CDx liquid biopsy linked to outcomes (rwOS, rwPFS) from Optum® Market Clarity. ctDNA profiling within 24 months prior to treatment was required. AIM-Bx inputs included mutation and copy-number alterations from the 74-gene panel, limited clinical features, and drug-specific structural embeddings. For each patient, AIM-Bx generated a sensitivity prediction, a Vulnerability Network, and a reconstructed expression profile. Performance was assessed using Kaplan-Meier analysis and Cox regression; available clinical covariates were evaluated as confounders. Results: AIM-Bx significantly stratified patients by benefit. Predicted-sensitive patients showed longer rw-PFS (median 14 vs. 4 months; p < 0.05) and rw-OS (14 vs. 7 months; p < 0.05). In Cox analyses, KRAS mutation category, histology, sex, and age > 65 were not significantly associated with outcomes. Biological interpretation revealed distinct features: predicted-sensitive tumors showed enriched KRAS/RAS-effector signaling, GEFs, CCND1-linked cell-cycle programs, and metabolic pathways including OXPHOS, HIF, and YAP/TAZ. Predicted non-responders displayed more heterogeneous dependencies and were enriched for downregulation of RAS-pathway programs and upregulation of EMT- and NOTCH-associated signatures. Conclusions: This retrospective study demonstrates the feasibility of predicting sotorasib benefit directly from liquid biopsy NGS using a multimodal AI model. AIM-Bx identifies biological programs beyond KRAS mutation status that may influence therapeutic sensitivity, suggesting hypotheses for vulnerability-based patient selection and potential combination strategies. Because AIM-Bx operates on routinely available inputs—including the FDA-approved CDx assay used for sotorasib eligibility—it may support future prospective evaluation without new assays. Citation Format: Maayan Baron, Felicia Kuperwaser, Sepideh Foroutan, Sunil Kumar, Dillon Tracy, Kevin Freisen, Brandon Funkhouser, Zong Miao, Nathaniel Tann, Fahad Khan, Sean Klei, Jordan Wolinsky, Taylor Wood, Jean Michel Rouly, Nick Lee, Ripple Khera, Anshu Jain, Jeff Sherman, Emily A. Vucic. Real world prediction and biological characterization of sotorasib sensitivity using multimodal AI and liquid biopsy genomic inputs [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2026; Part 1 (Regular Abstracts); 2026 Apr 17-22; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2026;86(7 Suppl):Abstract nr 5235.