Modeling Dabrafenib Response Using Multi-Omics Modality Fusion and Protein Network Embeddings Based on Graph Convolutional Networks

Cancer cell response to targeted therapy arises from complex molecular interactions, making single omics insufficient for accurate prediction. This study develops a model to predict Dabrafenib sensitivity by integrating multiple omics layers (genomics, transcriptomics, proteomics, epigenomics, and metabolomics) with protein network embeddings generated using Graph Convolutional Networks (GCN). Each modality is encoded into low dimensional representations through neural network preprocessing. Protein interaction information from STRING is incorporated using GCN to capture biological topology. An attention based fusion mechanism assigns adaptive weights to each modality according to its relevance. Using GDSC cancer cell line data, the model shows that selective integration of two modalities, especially proteomics and transcriptomics, achieves the best test performance (R2 around 0.96), outperforming all single omics and full multimodal settings. Genomic and epigenomic data were less informative, while proteomic and transcriptomic layers provided stronger phenotypic signals related to MAPK inhibitor activity. These results show that attention guided multi omics fusion combined with GCN improves drug response prediction and reveals complementary molecular determinants of Dabrafenib sensitivity. The approach offers a promising computational framework for precision oncology and predictive modeling of targeted therapies.

💡 Research Summary

This study presents a novel computational framework for predicting cancer cell line sensitivity to the targeted therapy Dabrafenib by integrating multi-omics data with biological network information. Recognizing that drug response is a complex phenotype arising from interconnected molecular layers, the authors move beyond single-omics approaches. Their model simultaneously leverages genomics, transcriptomics, proteomics, epigenomics, and metabolomics data, and enriches this information with functional context derived from protein-protein interaction (PPI) networks.

The methodological pipeline is meticulously designed. First, each omics modality undergoes independent preprocessing and is transformed into a low-dimensional embedding via a dedicated neural network encoder. In parallel, a Graph Convolutional Network (GCN) is employed to learn embeddings from a PPI network constructed using STRING database data. The GCN step is crucial as it encodes the topological and functional relationships between proteins, providing a structured biological context that pure numerical omics features lack. The core innovation lies in the fusion stage: instead of simply concatenating all embeddings, an attention mechanism dynamically assigns importance weights to each modality (including the PPI embedding) based on its relevance to the prediction task. This allows the model to focus on the most informative data sources. The final fused representation is then used for two downstream tasks: regression of continuous IC50/AUC values and binary classification of sensitivity.

The evaluation, conducted using data from the GDSC resource, yielded compelling and somewhat counterintuitive results. The model’s performance was rigorously tested across all possible combinations of the five omics modalities and the PPI embedding. Strikingly, the best-performing configuration was not the model using all available data, but the selective integration of just two modalities: proteomics and transcriptomics. This combination achieved an exceptional test performance of R² ≈ 0.96 and a very low RMSE of 0.48, surpassing all single-omics models and every multi-omics model with three or more modalities. Among single modalities, proteomics alone showed the strongest predictive power (R² ≈ 0.75), followed by metabolomics. In stark contrast, genomics and epigenomics data alone provided almost no predictive value (R² near zero or negative). The PPI network embedding offered moderate predictive value on its own but acted as a complementary source of information when fused with proteomics or transcriptomics.

A key insight from the analysis is that adding more data modalities is not always beneficial. Many three-, four-, five-, and six-modality combinations showed significantly worse performance than the top two-modality model, indicating that indiscriminate integration can introduce noise, redundancy, and lead to overfitting. The success of the proteomics-transcriptomics pair is biologically interpretable. Dabrafenib is a BRAF inhibitor targeting the MAPK signaling pathway. Therefore, the actual abundance and activity state (e.g., phosphorylation) of proteins within this pathway (captured by proteomics) and the transcriptional response of related genes (captured by transcriptomics) are more direct determinants of cellular response than the static presence of a genomic mutation. The study computationally validates that phenotypic signals are strongest in these functional layers closest to cellular activity.

In conclusion, this research demonstrates that an attention-guided, selective fusion of high-value omics layers, combined with GCN-derived biological network embeddings, can dramatically improve the accuracy of drug response prediction. It provides a sophisticated framework that balances data integration with model focus, offering a promising path for computational models in precision oncology. Future work would benefit from validation on larger, independent cohorts and deeper biological interpretation of the attention weights and features driving the predictions.