MetaComb: Drug Combination Response Predictions from Cell Lines to Patients

A meta-learning framework for predicting drug combination responses, developed by Congcong Guo and last updated in June 2026. It uses drug structures and gene expression profiles from cell lines and patient ex vivo samples to train a model that adapts to data-scarce patient scenarios. The method reportedly improved prediction AUROC by 8.5% for data-poor cell lines and 7.4% for patient ex vivo samples compared to conventional transfer learning.

Use Cases

• Predicting patient ex vivo drug combination responses based on drug structures and gene expression profiles.
• Fine-tuning pre-trained models on small patient-derived datasets for personalized therapy prediction.
• Benchmarking meta-learning approaches against conventional transfer learning for drug response prediction.
• Developing few-shot learning models for clinical scenarios where patient data is scarce.

Strengths

• Framework reportedly improved AUROC by 8.5% for data-poor cell lines and 7.4% for patient ex vivo samples.
• Uses a meta-learning approach designed to adapt from data-rich cell line tasks to data-scarce patient tasks.
• Released under a permissive CC-BY-4.0 license for open use.

Limitations

• Dataset size is only 16.4 KB, suggesting limited raw data or a documentation-only file.
• Row count and column-level documentation are absent, requiring manual inspection after download.
• The description notes that validation with larger patient cohorts is needed for generalization.

Provenance

Source

Congcong Guo via figshare.

Collection Method

Developed using Model-Agnostic Meta-Learning (MAML) on cell line data and fine-tuned on patient ex vivo samples.

Freshness

Last updated 2026-06-01 22:02:55.