MetaComb: a meta-learning framework for drug combination response prediction from cell lines to patients

Combination therapy has emerged as a pivotal strategy in oncology to enhance efficacy and overcome drug resistance. Computational prediction models of drug combinations trained on abundant cell line data provide a starting point, but their applicability to patients remains constrained by inherent biological disparities between cultured cell lines and patient-derived tumors. However, due to ethical and cost issues, patient-derived datasets remain scarce, thus, developing patient-level predictive algorithms must explicitly confront the few-shot problem of relevant data.

To break through the small sample bottleneck, we used the Model-Agnostic Meta-Learning (MAML) to develop a Meta-Learning Drug Combination Response Prediction (MetaComb) method for patient ex vivo drug combination response prediction, using drug structures and gene expression profiles of cell lines/patients as the model input. In MetaComb, the meta-model was trained on data-rich cell line-specific drug combination response prediction tasks and subsequently fine-tuned to adapt to scenarios with limited samples.

MetaComb outperformed conventional transfer learning in predicting drug combination response, improving AUROC by 8.5% for data-poor cell lines and by 7.4% for patient ex vivo samples. And for the patients with ex vivo data, MetaComb also achieved superior accuracy over existing methods. Given the limited patient cohort, these results demonstrate the feasibility of MetaComb, but further validation with larger patient datasets is needed.

This study, as a proof-of-concept, provided an initial evidence that the MetaComb meta-learning framework is feasible for patient-derived ex vivo drug combination response prediction under few-shot conditions, by transferring drug-combination response knowledge from preclinical cell lines. Current patient-level assessment is insufficient to support generalization to other cancer types or patient populations, and in the future, with the accumulation of relevant patient-derived data, further validation with larger patient cohorts is required.