Few-Shot Drug Synergy Prediction Via Rapid Cross-Tier Adaptation Meta-Optimization

Abstract

Drug combination therapy offers key advantages over monotherapy in personalized oncology by reducing drug resistance and toxicity. However, synergy prediction for rare cell lines remains challenging due to limited experimental data, and current computational methods often struggle with generalization in such data-scarce scenarios. To this end, we present MetaSynergy, a Rapid Cross-tier Adaptation Meta-Optimization(R-CAMO)-based few-shot learning model that enhances drug synergy prediction in data-scarce cell lines. We first designed a multimodal architecture integrating drug pair features with cell line omics profiles, then implemented R-CAMO through a stage-wise training protocol to achieve few-shot drug synergy prediction: (i) The Cross-domain pretraining establishes meta-initialized representations by transferring knowledge from data-rich cell lines to scarce target domains, enhancing feature representation capability in data-scarce scenario. (ii) Cross-tier meta optimization enables rapid adaptation to data-scarce scenario: the inner-tier refines task-specific parameters of the prediction network via gradient descent on target cell line data, and then, the outer-tier meta-learns task-shared, generalizable parameters by minimizing the cross-cell line prediction loss. (iii) The fine-tuning leverages few-shot target cell line samples to further refine task-specific parameters, improving generalizability to novel drug combinations within the same cellular context. Comprehensive experiment results demonstrate that MetaSynergy significantly outperforms conventional machine learning models, transfer learning approaches (TFSynergy), and state-of-the-art meta-learning models (MAML/BOIL/HyperSynergy) across both few-shot (5/10/30 samples) and zero-shot prediction scenarios. Moreover, MetaSynergy exhibited sustained robustness in low-similarity tasks (Jaccard similarity: 0.347-0.538 between meta-training and meta-test cell line genomic profiles) while demonstrating generalizable predictive performance. Furthermore, ablation studies validated the critical role of the R-CAMO strategy in data-scarce cell lines. Significantly, we identified five novel synergistic drug pairs with therapeutic potential in understudied malignancies, including Ewing sarcoma (A-673), breast cancer (MFM-223), and non-small cell lung cancer (NCI-H1299). Beyond establishing a computationally efficient framework for drug synergy prediction, this work advances the paradigm of data-efficient machine learning in precision oncology, enabling rapid derivation of patient-specific combination therapies. Source code and data are at https://github.com/Emmnmusee/MetaSynergy.

Note:
Funding Information: This work was supported by National Natural Science Foundation of China (Grant Nos. 62173271, 62473312).

Conflict of Interests: All authors declare that there are no financial, personal, or professional conflicts of interest that could influence the outcomes or interpretation of this research.

Keywords: Drug Synergy Prediction, Few-shot learning, Cross-tier Optimization, Precision Oncology