Abstract 5011: Integrating multimodal data with survival-based variational autoencoders to predict recurrence-free survival in breast cancer

Abstract Breast cancer is the most commonly diagnosed malignancy globally and responsible for over 670, 000 deaths each year. Accurate prediction of recurrence-free survival (RFS) is crucial for clinical management and personalized care in breast cancer, enabling improved surveillance, risk stratification, and optimized therapy selection. Although extensive multi-omic data from next generation sequencing tests have been collected, there is a notable gap in clinically relevant models that leverage advanced artificial intelligence-based approaches that integrate multiple data modalities. Variational autoencoders (VAEs) have been used to compress omics datasets into lower dimensionality, however the latent dimensions extracted from VAEs typically lack biological significance, hindering interpretation and the performance of downstream predictive modeling. To address this pitfall, we have developed a VAE with a built-in survival model. Our approach uses the performance of the survival model to iteratively optimize the VAE, generating latent dimensions which are relevant for predicting RFS. We developed our model by extracting latent features from approximately 800 breast cancer cases from METABRIC split into training and test datasets. Gene expression, MSigDB pathway enrichment scores, CNAs, and SNV/Indels were used to train four separate VAEs. The model’s encoder compressed high-dimensional omic datasets into a set of latent dimensions, while the decoder then used the latent dimensions to reconstruct the original data; simultaneously, linear representations of the latent dimensions were used as inputs to a cox regression model. The reconstruction loss, Kullback-Leibler divergence, and negative partial log likelihood based on the survival model were used to reparametrize the VAE during each training iteration. Final models for each modality were selected based on survival prediction on the test dataset with c-index ranging from 0.61-0.65. A cox regression model was fit based on a combined dataset of the latent dimensions from each model, and clinical data from each patient. To assess the ability of our model to stratify high- and low-risk patients, an optimal cut-off was established based on the linear predictors of the cox regression model in the training dataset. There was a significant difference in survival between high- and low-risk groups in both the training (p < 0.0001) and the test (p < 0.0001) cohorts. The 5-year survival rates of high- and low-risk groups were 30.7% vs 88.9% in the training dataset and 60.5% vs 87.8% in the test dataset. Overall, our survival-based VAEs successfully extracted relevant latent dimensions for prediction of recurrence-free survival in breast cancer patients. Further implementation of our model may be used to predict treatment-specific response, helping clinicians make informed treatment decisions for cancer patients. Citation Format: Sydney Grant, Aritro Nath. Integrating multimodal data with survival-based variational autoencoders to predict recurrence-free survival in breast cancer [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 5011.