Transformer-Based Multi-Omics Integration for Cancer Drug Response Prediction: A PRISMA 2020 Systematic Review

Abstract

Background: Transformer and attention architectures are increasingly used to integrate multi-omics profiles for predicting cancer drug response, but their added value and the robustness of the reported results are unclear.

Objectives: To systematically identify, describe and critically appraise peer-reviewed studies that use transformer or attention mechanisms to integrate at least two omics layers for cancer drug response prediction.

Methods: We searched PubMed, Europe PMC, OpenAlex, arXiv and Semantic Scholar studies published between 1 January 2021 and 31 December 2025 and performed backward and forward citation searching. Eligible studies were peer-reviewed original research in which a transformer or attention mechanism was a core component of the proposed model, at least two omics layers were integrated, and quantitative drug-response predictions were reported. Risk of bias was assessed with PROBAST adapted for machine-learning prediction models. Results were synthesised without meta-analysis (SWiM), stratified by dataset, split strategy and metric, and certainty was rated with an adapted GRADE approach.

Results: From 2,950 database records and 1,627 records from citation searching, 21 studies (24 reports) were included (18 via databases and 3 via citation searching). Most used GDSC cell-line data (18 of 21) and intermediate, attention-based fusion (15 of 21). Random pair splits were used in 13 of 21 studies and were the only evaluation in 4; cell-line-blind, drug-blind, cross-dataset and clinical/external evaluations were used in 14, 8, 5 and 3 studies, respectively. Global Pearson correlations for GDSC-family random pair splits ranged from 0.87 to 0.94 (8 estimates from 8 studies) and from 0.60 to 0.74 under drug-blind evaluation (2 estimates). Overall risk of bias was high in 9, unclear in 9 and low in 3 studies. Certainty of evidence was low for random pair and cell-line-blind performance and very low for drug-blind, cross-dataset, clinical/external and comparative performance.

Discussion: High reported accuracy is concentrated in evaluations that share cell lines and drugs between training and test sets; evidence of generalisation to new drugs, datasets or patients, or of an advantage over well-tuned simpler models, is of very low certainty. Limitations include AI-assisted screening and extraction by a single assessor without a second reviewer and searches not extended to Web of Science and Google Scholar.