Cross-attention enables deep learning on limited omics-imaging-clinical data of 130 lung cancer patients

Abstract

Summary: Deep-learning tools that extract prognostic factors derived from multi-omics data have recently contributed to individualized predictions of survival outcomes. However, the limited size of integrated omics-imaging-clinical datasets poses challenges. Here, we propose two biologically interpretable and robust deep-learning architectures for survival prediction of non-small cell lung cancer (NSCLC) patients, learning simultaneously from computed tomography (CT) scan images, gene expression data, and clinical information. The proposed models integrate patient-specific clinical, transcriptomic, and imaging data and incorporate Kyoto Encyclopedia of Genes and Genomes (KEGG) and Reactome pathway information, adding biological knowledge within the learning process to extract prognostic gene biomarkers and molecular pathways. While both models accurately stratify patients in high- and low-risk groups when trained on a dataset of only 130 patients, introducing a cross-attention mechanism in a sparse autoencoder significantly improves the performance, highlighting tumor regions and NSCLC-related genes as potential biomarkers and thus offering a significant methodological advancement when learning from small imaging-omics-clinical samples. Motivation: Multimodal deep-learning models can be used to obtain personalized survival predictions. However, the small size of most matched omics-imaging-clinical studies currently poses significant challenges to the development and application of such tools. Furthermore, the lack of interpretability makes it difficult to understand the biological rationale behind the predictions, leading to a lack of trust and reluctance to adopt them in clinical settings. Specifically, the inability to explain how specific features contribute to the predictions limits the potential for new insights and identification of prognostic biomarkers. We propose two biologically interpretable and robust deep-learning architectures for survival prediction of 130 non-small cell lung cancer (NSCLC) patients, integrating patient-specific clinical, transcriptomic, and imaging data. We incorporate KEGG and Reactome pathway information, adding biological knowledge within the learning process. Introducing a cross-attention mechanism in a sparse autoencoder allows extracting prognostic gene biomarkers and molecular pathways that are biologically interpretable even in the presence of small samples and highlights tumor regions successfully validated by two radiologists.

Keywords

• CP: Systems biology
• CP: Cancer biology