1
0The research introduces Conditional Monge Gap (CMonge), a mathematically grounded machine learning framework designed to predict how single cells respond to various treatments. Unlike previous models that require separate training for every individual condition, this global estimator uses optimal transport to learn from hundreds of drug and dose combinations simultaneously. By incorporating conditional information like drug structure and dosage, the system effectively captures the complex diversity within cell populations and identifies patterns across different tasks. This approach allows for generalization to unseen drugs, accurately forecasting cellular changes even for compounds not included in the initial training data. Consequently, CMonge offers a parameter-efficient and scalable tool for accelerating drug discovery and personalized medicine through high-fidelity virtual cell modeling.
References:
Driessen A, Rajwade D A, Harsanyi B, et al. Conditional Monge Gap enables generalizable single-cell perturbation modelling[J]. Nature Machine Intelligence, 2026: 1-13.
