Abstract 3644: Mechanistically explainable AI model for predicting synergistic cancer therapy combinations

Abstract Background: Resistance to single-agent cancer therapies often arises through complex mechanisms such as genetic mutations, compensatory signaling, and tumor microenvironment changes. These adaptive strategies make multi-drug combinations essential to effectively target canonical pathways simultaneously. However, common in vitro and in vivo models often fail to capture these complexities, resulting in limited translational success to human clinical settings. This highlights the need for predictive frameworks that not only identify synergistic drug combinations from human clinical trials, but also provide mechanistic insights to inform clinical application. This study presents a novel Large Language Model (LLM)-based framework that integrates extensive drug combination data with a knowledge graph to predict drug combinations with biological explainability. Methods: A retrieval-augmented generation (RAG) approach was employed to enhance predictive specificity and reduce hallucinations. In vitro data on drugs, targets, cell lines, tumor types, and synergy scores were compiled from over 50, 000 cell line results from the DrugComboDb database. Using NCBI’s APIs, BeautifulSoup, and GPT-4, biomedical information was extracted from 1, 092 phase I-III clinical trials covering 700+ drug combinations across 22 tumor types. This RAG corpus, combined with targeted prompts, fed into the LLM Mistral v0.2 for prediction generation, while a knowledge graph - PrimeKG provided mechanistic insights. Model validation included unseen cases (N=42) across various drug modality combinations (N=8), including 34 immuno-oncology therapies (IO) combined with an IO, chemo, or targeted therapy, 1 targeted therapy combination and 7 antibody drug conjugate (ADC) combinations. Results: The model achieved an F1 score of 0.80 on the entire validation set, with accuracy at 0.74, precision at 0.71, and recall at 0.92. Among drug modalities, combinations of ADC agents with IO/targeted/chemo/other ADC agents performed best (F1 score: 100%), while IO-IO combinations scored the lowest (F1 score: 62%). Each prediction included detailed mechanistic explanations, highlighting whether the drug combination was supported or unsupported with biological rationale. Conclusions: This study presents a novel framework integrating LLMs with biological knowledge graphs to predict the efficacy when combining 2-3 oncology drugs across a range of indications. By linking generative AI with explainability tools, we address two critical challenges in AI-driven drug discovery: predictive accuracy and mechanistic transparency. Such predictive models can accelerate the discovery of synergistic therapies, thereby reducing costly experiments, and improve overall patient outcomes. Specific test cases with explainable results will be presented. Citation Format: Han Si, Sanyam Kumar, Sneh Lata, Arshad Ahmad, Saurav Ghosh, Karen Stephansen, Deepti Nagarkar, Eda Zhou, Brandon W. Higgs. Mechanistically explainable AI model for predicting synergistic cancer therapy combinations [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 3644.