Prediction of PLX-4720 Sensitivity in Cancer Cell Lines through Multi-Omics Integration and Attention-Based Fusion Modeling

TL;DR

This study develops an attention-based multi-omics integration framework to predict cancer cell line sensitivity to BRAF inhibitor PLX-4720, demonstrating that combining genomics and transcriptomics yields high predictive accuracy.

Contribution

The paper introduces a novel multi-omics fusion model using attention mechanisms, highlighting the importance of modality selection over data quantity for drug response prediction.

Findings

Genomics and transcriptomics integration achieves R2 > 0.92.

Epigenomics is the strongest single predictor.

Adding more omics layers does not improve accuracy beyond two modalities.

Abstract

Predicting the sensitivity of cancer cell lines to PLX-4720, a preclinical BRAF inhibitor, requires models capable of capturing the multilayered regulation of oncogenic signaling. Single-omics predictors are often insufficient because drug response is shaped by interactions among genomic alterations, epigenetic regulation, transcriptional activity, protein signaling, metabolic state, and network-level context. In this study we develop an attention-based multi-omics integration framework using genomic, epigenomic, transcriptomic, proteomic, metabolomic, and protein interaction data from the GDSC1 panel. Each modality is encoded into a latent representation using feed-forward neural networks or graph convolutional networks, and fused through an attention mechanism that assigns modality-specific importance weights. A regression model is then used to predict PLX-4720 response. Across single- and multi-omics configurations, the best performance is achieved by integrating genomics and transcriptomics, which yields validation R2 values above 0.92. This reflects the complementary roles of mutational status and downstream transcriptional activation in shaping sensitivity to BRAF inhibition. Epigenomics is the strongest single-omics predictor, while metabolomics and PPI data contribute additional context when combined with other modalities. Integration of three to five omics layers improves stability but does not surpass the accuracy of the best two-modality combinations, likely due to information redundancy and sample-size imbalance. These findings highlight the importance of modality selection rather than maximal data depth. The proposed framework provides an efficient and biologically grounded strategy for drug response prediction and supports the development of precision pharmacogenomics.