Monotherapy cancer drug-blind response prediction is limited to intraclass generalization

Sep 23, 2026

Uploaded by Angel

Monotherapy cancer drug response prediction (DRP) models aim to forecast how cell lines respond to drugs. Drug-blind evaluation, where test drugs were unseen in training, often shows severely reduced performance or complete failure. This work quantifies how drug-blind generalizability depends on mechanistic overlap between training and testing drug splits, identifying that mixed-set performance largely reflects drug overfitting. Sources of transferable drug features concentrate on shared mechanisms of action and related pathways, and restricting training to a single mechanism can improve performance for certain cases.

LanguageEnglish

CategoryHealthcare

FormatPDF

Pages22

Reading Time55 min

AccessFree to download

Abstract

Drug-blind prediction limitations

Mechanistic overlap and generalizability

Drug overfitting and performance interpretation

Improving performance by single-mechanism training

OPEN ACCESS Citation: Herbert WG, Chia N, Jensen PA, Walther-Antonio MRS (2026) Monotherapy cancer drug-blind response prediction is limited to intraclass generalization. PLoS Comput Biol 22(4): e1013232. [https://doi.org/10.1371/](https://doi.org/10.1371/)[ ](https://doi.org/10.1371/)[journal.pcbi.1013232](journal.pcbi.1013232) Editor: Hailin Chen, East China JiaoTong University, CHINA Received: June 13, 2025 Accepted: March 26, 2026 Published: April 22, 2026 Peer Review History: PLOS recognizes the benefits of transparency in the peer review process; therefore, we enable the publication of all of the content of peer review and author responses alongside final, published articles. The editorial history of this article is available here: [https://doi.org/10.1371/journal](https://doi.org/10.1371/journal). pcbi.1013232 Copyright: This is an open access article, free of all copyright, and may be freely reproduced, distributed, transmitted, modified, built upon, or otherwise used by anyone for any lawful purpose. The work is made available under RESEARCH ARTICLE Monotherapy cancer drug-blind response prediction is limited to intraclass generalization William G. Herbert1,2,3,4*, Nicholas Chia5, Paul A. Jensen6,7, Marina R. S. Walther-Antonio2,3,4 1 Graduate School of Biomedical Sciences, Mayo Clinic, Rochester, Minnesota, United States of America, 2 Department of Obstetrics and Gynecology, Mayo Clinic, Rochester, Minnesota, United States of America, 3 Department of Surgery, Mayo Clinic, Rochester, Minnesota, United States of America, 4 Microbiome Program, Center for Individualized Medicine, Mayo Clinic, Rochester, Minnesota, United States of America, 5 Computing, Environment, and Life Sciences, Argonne National Laboratory, Lemont, Illinois, United States of America, 6 Department of Biomedical Engineering, University of Michigan, Ann Arbor, Michigan, United States of America, 7 Department of Chemical Engineering, University of Michigan, Ann Arbor, Michigan, United States of America * [willherb@umich.edu](willherb@umich.edu) Abstract Monotherapy cancer drug response prediction (DRP) models predict the response of a cell line to a given drug. Analyzing these models’ performance includes assessing their ability to predict the response of cell lines to new drugs, i.e. , drugs that are not in the training set. Drug-blind prediction displays greatly diminished performance or outright failure across a wide range of model architectures and different large pharmacogenomic datasets. Drug-blind failure is hypothesized to be caused by the relatively limited set of drugs present in these datasets. The time and cost associated with further cell line experiments is significant, and it is impossible to predict beforehand how much data would be enough to overcome drug-blind failure. We must first define how current data contributes to drug-blind failure before attempting to remedy drug-blind failure with further data collection. In this work, we quantify the extent to which drug-blind generalizability relies on mechanistic overlap of drugs between training and testing splits. We first identify that the majority of mixed set DRP model performance can be attributed to drug overfitting, likely inhibiting generalization and preventing accurate analysis. Then, by specifically probing the drug-blind ability of models, we reveal the sources of generalizable drug features are confined to shared mechanisms of action and related pathways. Furthermore, we observed that, for certain mechanisms, we can significantly improve performance by limiting the training of models to a single mechanism compared to training on all drugs simultaneously. Across multiple different model architectures examined in this paper, we observe that drug-blind performance is a poor benchmark for DRP as it does not describe model behavior, it describes dataset behavior. Our investigation displays that these deep learning models trained on large, monotherapy

FAQs

What is drug-blind response prediction in cancer DRP models?

Drug-blind prediction evaluates whether a model can accurately predict responses for drugs that were not included in the training set.

Why do drug-blind predictions often fail?

The paper attributes poor drug-blind performance largely to limited mechanistic overlap between drugs across training and testing splits and to drug overfitting that inhibits generalization.

How can performance be improved for certain mechanisms?

For certain mechanisms, performance improves when models are trained using only a single mechanism rather than training on all drugs together.

Sep 23, 2026

Uploaded by Angel