INTERNATIONAL CONGRESS OF
CancerGenetics
Can Multimodal Deep Learning Reveal Resistance Mechanisms and Improve Osimertinib Response Prediction in NSCLC?
-
Elaheh AskariZadeh,1,*
1. Apadana Institute, Shiraz, Department of Engineering Technology, Medical Engineering
- Introduction: Non-small cell lung cancer (NSCLC) with Epidermal Growth Factor Receptor (EGFR) mutations is one of the most challenging diseases, as patients show highly variable and unpredictable responses to targeted therapies. The third-generation EGFR inhibitor osimertinib has shown great promise and has had significant success; however, drug resistance and clinical variability remain major obstacles. To date, no study has simultaneously integrated molecular, imaging, and clinical data to predict osimertinib response in NSCLC. Existing univariate machine learning models predict patient response relatively well, but they fall short in addressing resistance pathways and multimodal integration. This review takes a comprehensive and narrative look at multimodal AI models to both address the gap in accurately predicting response to osimertinib and to shed light on potential avenues for future research. Therefore, this review seeks to answer the question of whether multimodal AI frameworks can both increase the accuracy of predicting response to osimertinib and reveal novel resistance pathways.
- Methods: To write this review, the workflow was designed according to the PRISMA 2020 guideline and the PICOS framework, and PROSPERO standards were also considered in the design of the protocols to ensure transparency and reproducibility. strategy was designed to cover PubMed, Scopus, and IEEE Xplore databases for articles published between 2024 and 2025. Keywords such as NSCLC, Osimertinib, AI, Deep Learning, Drug Response Prediction, Resistance Mechanisms were used. References were planned to be managed with EndNote, with a workflow outlined for screening using Rayyan. to maintain clinical relevance and transferability, studies that focused solely on single-omics data or traditional statistical approaches were excluded. Study selection was planned in two stages (abstract and full text). Only experimental studies, multiomics data modeling, and machine learning analyses were included in the review. QUADAS-2 and ROBINS-I tools were planned for quality assessment according to study type. The review focuses on reported deep and multimodal learning models, including PASA, DeepDRA, GNN and GPT-based approaches, with the ability to analyze genomic, transcriptomic, proteomic, metabolomic and advanced imaging (radiomic and pathomic) data. Studies reported using MOFA+, DIABLO, and SNF for data integration, with Nested Cross-Validation and Bootstrapping to assess model robustness. According to the included studies, Explainable AI approaches including SHAP, LIME and Integrated Gradients were used to elucidate molecular pathways and resistance patterns. The main criteria included the accuracy of predicting drug response, the generalizability of the model and the identification of emerging trends in resistance. To measure the performance, indicators such as AUC, C-index, Accuracy and F1-score were used.
- Results: 1. Multimodal ML models in NSCLC/osimertinib Recent studies that integrated multi-source data (imaging, genetics, clinical, and demographics) showed that this approach has higher accuracy than univariate models in predicting response to osimertinib (C-index≈0.82). This directly covers the research gap, as it is the first time that molecular and imaging data were used simultaneously. 2. DeepDRA and similar models Although DeepDRA was originally tested on general drug data, it has shown great ability to reduce multidimensionality and predict drug response (AUPRC≈0.99). Its main limitation was the instability in cross-sectional data; therefore, for osimertinib in NSCLC, it can be used as an accurate tool at the molecular level, but needs to be strengthened in generalizability. 3. Pathformer Transformers These models increase accuracy by preserving the interaction of biological pathways such as MAPK and PI3K/AKT, which are specifically implicated in osimertinib resistance. Although they have not yet been directly applied in NSCLC/osimertinib, their potential to fill the gap in understanding resistance pathways is very high. This could answer the main question of the paper in terms of discovering new resistance pathways. 4. Graph Neural Networks (GNN + GSR-PPI) Graphic models were able to capture more accuracy in predicting response to osimertinib in relevant laboratory cells by analyzing protein interactions. Extending this approach to NSCLC clinical data could lead to improvement in prediction. These models excel at handling patient heterogeneity, which is exactly what raised as a clinical obstacle. 5. Large Language Models (LLMs like ChatGPT) Although not yet practically applied in NSCLC/osimertinib, hybrid models (LLM+GNN/Transformer) Most were able to improve by up to 25% in similar scenarios. This suggests that in the future, LLMs could be the same multimodal data integration tool that was identified as the main gap. Overall, DeepDRA has the highest raw accuracy but the lowest inter-data stability, while GNN and Pathformer strike a better balance between accuracy and generalizability by focusing on heterogeneity and signaling pathways (MAPK, PI3K/AKT). Multimodal clinical-imaging models are the most direct response to the data integration gap, and LLMs, despite their lack of practical application, have the most prospective potential for multi-source integration and discovery of novel resistance pathways.
- Conclusion: With the innovation of simultaneously integrating molecular, imaging, and clinical data with multimodal AI models, this review demonstrated that prediction of response to osimertinib in NSCLC is significantly improved due to simultaneous analysis of biological and clinical pathways, resistance pathways are revealed, research gaps are filled, and future research should focus on clinical validation and personalized platforms.
- Keywords: Non-Small Cell Lung Cancer (NSCLC), Osimertinib resistance, Multimodal artificial intelligence, Drug
