Abstract
In line with recent advances in neural drug design and sensitivity prediction, we propose a novel architecture for interpretable prediction of anticancer compound sensitivity using a multimodal attention-based convolutional encoder. Our model is based on the three key pillars of drug sensitivity: compounds' structure in the form of a SMILES sequence, gene expression profiles of tumors and prior knowledge on intracellular interactions from protein-protein interaction networks. We demonstrate that our multiscale convolutional attention-based (MCA) encoder significantly outperforms a baseline model trained on Morgan fingerprints, a selection of encoders based on SMILES as well as previously reported state of the art for multimodal drug sensitivity prediction (R2 = 0.86 and RMSE = 0.89). Moreover, the explainability of our approach is demonstrated by a thorough analysis of the attention weights. We show that the attended genes significantly enrich apoptotic processes and that the drug attention is strongly correlated with a standard chemical structure similarity index. Finally, we report a case study of two receptor tyrosine kinase (RTK) inhibitors acting on a leukemia cell line, showcasing the ability of the model to focus on informative genes and submolecular regions of the two compounds. The demonstrated generalizability and the interpretability of our model testify its potential for in-silico prediction of anticancer compound efficacy on unseen cancer cells, positioning it as a valid solution for the development of personalized therapies as well as for the evaluation of candidate compounds in de novo drug design.
Introduction
Precision oncology relies on matching a patient's unique tumor profile to the right drug. Machine learning models can predict this "drug sensitivity" (often measured as IC50, the concentration needed to inhibit 50% of cell growth), but they traditionally suffer from two major flaws. First, they often rely on engineered chemical features, such as Morgan fingerprints, which constrain the model's ability to learn novel structural patterns. Second, they frequently act as black boxes; clinicians and chemists require interpretability to understand the biological and chemical drivers behind a prediction.
Furthermore, many existing models are evaluated using "lenient" data splits, which may lead to memorization of specific drugs or cell lines rather than learning generalizable rules. This work addresses the need for a multimodal, highly interpretable model capable of generalizing to entirely new drug-cell pairs through a novel architecture and strict evaluation protocols.
Methods and Data
The proposed architecture integrates three core pillars to capture the complexity of drug-cell interactions:
- Transcriptomics (The Cell): Gene expression profiles represent the cancer cell state. To manage high dimensionality, a network propagation technique over a Protein-Protein Interaction (PPI) network (sourced from STRING) filters ~17,000 genes down to a highly informative subset of 2,128 genes. A gene attention encoder then assigns weights to these genes based on their relevance.
- Drug Structure (The Drug): Rather than using fixed fingerprints, the model processes raw SMILES strings. This allows the network to learn optimal chemical representations directly from the molecular structure.
- Contextual Attention (The Bridge): The novel Multiscale Convolutional Attentive (MCA) encoder uses parallel 1D CNNs with varying kernel sizes to scan SMILES strings for diverse chemical substructures. A contextual attention mechanism uses the cell line's gene expression as context, allowing the model to focus on specific molecular regions depending on the genetic makeup of the target cell.
To ensure generalizability, the model was evaluated using a "strict split" protocol, where no cell line or drug in the test set appeared in the training data. Data was sourced from the Genomics of Drug Sensitivity in Cancer (GDSC) database, PubChem, and LINCS.
Key Results and Insights
The MCA model demonstrates superior performance across multiple metrics and evaluation scenarios:
- Predictive Performance: On the challenging strict split, the MCA model outperformed all baselines, including RNNs and fingerprint-based DNNs, achieving a standardized RMSE of 0.104. In a standard lenient split, the model achieved an $R^2$ of 0.86 and an RMSE of 0.887 on the log-IC50 scale.
- Biological and Chemical Validity: Attention weights align with established biological knowledge; highly attended genes are significantly enriched in apoptotic processes. Chemically, drug attention correlates strongly with Tanimoto structural similarity, indicating the model learns genuine chemical features.
- Explainability Case Study: In a comparison of two similar Tyrosine Kinase (TK) inhibitors, Imatinib and Masitinib, on a leukemia cell line, the model correctly highlighted the specific functional group that distinguishes the two compounds. Simultaneously, gene attention focused on leukemia-relevant genes such as DDR1 and BMP2K.
Conclusion
The Multiscale Convolutional Attentive encoder provides a robust and interpretable framework for anticancer drug sensitivity prediction. By moving beyond engineered features and black-box models, this approach offers a path toward personalized therapy and informed de novo drug design. Future work will explore integrating this predictor as an oracle within generative AI frameworks to design patient-specific compounds.
Summary updated