[Paper Review] Deep learning methods for drug response prediction in cancer: predominant and emerging trends

Alexander Partin, Thomas Brettin+5 more|arXiv (Cornell University)|Nov 18, 2022

Computational Drug Discovery Methods4 citations

TL;DR

This paper provides a comprehensive review of deep learning methods for predicting cancer drug response, analyzing 60 models to identify predominant and emerging trends. It reveals that multiomics integration offers only marginal gains over single-omics approaches, while graph-based drug representations show limited performance improvement, highlighting key challenges in model generalization and data representation for precision oncology.

ABSTRACT

Cancer claims millions of lives yearly worldwide. While many therapies have been made available in recent years, by in large cancer remains unsolved. Exploiting computational predictive models to study and treat cancer holds great promise in improving drug development and personalized design of treatment plans, ultimately suppressing tumors, alleviating suffering, and prolonging lives of patients. A wave of recent papers demonstrates promising results in predicting cancer response to drug treatments while utilizing deep learning methods. These papers investigate diverse data representations, neural network architectures, learning methodologies, and evaluations schemes. However, deciphering promising predominant and emerging trends is difficult due to the variety of explored methods and lack of standardized framework for comparing drug response prediction models. To obtain a comprehensive landscape of deep learning methods, we conducted an extensive search and analysis of deep learning models that predict the response to single drug treatments. A total of 60 deep learning-based models have been curated and summary plots were generated. Based on the analysis, observable patterns and prevalence of methods have been revealed. This review allows to better understand the current state of the field and identify major challenges and promising solution paths.

Motivation & Objective

• To map the current landscape of deep learning models for drug response prediction in cancer using a curated dataset of 60 models.
• To identify predominant and emerging methodological trends in data representation, neural network architectures, and learning strategies.
• To evaluate the impact of multiomics data integration on predictive performance and assess the value of graph-based molecular representations.
• To highlight persistent challenges such as poor generalization to unseen compounds and limited adoption of ranking-based prediction for clinical decision support.
• To guide future research by identifying underexplored but promising directions, including transformer-based SMILES modeling and 2D-CNNs on molecular descriptors.

Proposed method

• Conducted an extensive literature search and curated 60 deep learning-based drug response prediction (DRP) models from peer-reviewed publications.
• Systematically analyzed model architectures, data representations (e.g., gene expression, copy number variation, molecular fingerprints, molecular graphs), and learning methodologies.
• Generated summary plots to visualize the prevalence and performance trends of different model components across studies.
• Evaluated the predictive performance of models using standardized metrics such as Pearson correlation coefficient (PCC) and area under the curve (AUC).
• Compared the impact of single-omics versus multiomics input, and assessed the added value of graph-based drug representations (e.g., GNNs) versus traditional descriptors.
• Reviewed the use of response metrics such as IC50, AUC, AAC, and discrete or ranked outcomes for clinical applicability.

Experimental results

Research questions

• RQ1 What are the most prevalent data representations and neural network architectures used in deep learning-based drug response prediction models?
• RQ2 How does integrating multiomics data affect the predictive performance of DRP models compared to single-omics approaches?
• RQ3 To what extent do graph-based models (e.g., GNNs) improve prediction accuracy over traditional molecular descriptor or fingerprint-based methods?
• RQ4 Why do many models show poor generalization to previously unseen compounds, and what are the key limitations in current model design?
• RQ5 What are the underexplored but promising methodological directions in DRP, particularly for clinical translation and personalized treatment recommendation?

Key findings

• Multiomics integration is frequently employed but only provides a significant performance boost in a minority of studies, with most reporting only marginal improvements.
• Despite theoretical advantages, graph-based drug representations (e.g., GNNs) show minimal performance gains—less than 0.5% improvement in one side-by-side comparison—suggesting limited added value over traditional descriptors.
• The use of IC50 as a response metric remains dominant, but recent studies are shifting toward more robust global metrics like AUC and AAC to reduce noise from curve fitting.
• Only a small number of models explore ranking-based prediction, despite its direct relevance to personalized treatment recommendation and clinical decision-making.
• Generalization to unseen compounds remains a major challenge, with model performance dropping significantly on out-of-distribution drugs, limiting utility in virtual screening.
• Underexplored areas include transformer-based modeling of SMILES sequences, 2D-CNNs applied to image-converted molecular descriptors, and kinase inhibition profiles for targeted therapies.