Conditional Monge Gap enables generalizable single-cell perturbation modelling

2026-06-01 · Source: Nature Machine Intelligence · Field: Health & Wellbeing — Pharmaceuticals & Biotechnology, Artificial Intelligence & Machine Learning, Life Sciences & Biology · Depth: Expert, extended

Summary

The Conditional Monge Gap (CMonge) is a novel neural optimal transport (OT) framework designed for generalizable single-cell perturbation modeling. This method learns optimal transport maps conditionally on various covariates, such as drug identity, dose, or combinations, addressing the limitations of existing unconditional OT approaches that struggle with unpaired data and generalization to unseen treatments. CMonge predicts single-cell perturbation responses, demonstrating performance comparable to, and often superior to, state-of-the-art single-cell RNA sequencing and multiplexed protein imaging models. Notably, CMonge scales efficiently to hundreds of conditions and drugs, enabling cross-task learning and accurate predictions for unseen drugs using only their compound structure. Evaluations on the SciPlex dataset (187 drugs, 4 doses) and the 4i dataset (35 cancer therapies) showed CMonge outperformed baselines and the chemCPA model in out-of-sample prediction, particularly in capturing cellular heterogeneity.

Key takeaway

For AI Scientists and Machine Learning Engineers developing drug discovery platforms, integrate Conditional Monge Gap (CMonge) for predicting single-cell perturbation responses. This method offers superior generalization to unseen drugs and doses, even with limited training data, by capturing cellular heterogeneity more effectively than previous models. Utilize its parameter efficiency and ability to use compound structure for virtual screening, accelerating drug repurposing and identifying critical cell states.

Key insights

CMonge is a conditional optimal transport framework that accurately predicts single-cell perturbation responses and generalizes to unseen conditions.

Principles

• Conditional modeling improves generalization to unseen contexts.
• Global models reduce computational burden compared to local models.
• Optimal transport inherently captures cellular population heterogeneity.

Method

CMonge learns a single global optimal transport map by minimizing Sinkhorn divergence with a Monge Gap regularizer, conditioned on covariates like drug structure or dose.

In practice

• Accelerate drug discovery and repurposing workflows.
• Screen novel compounds virtually using structure-based embeddings.
• Identify rare, resistant cell states for toxicity assessments.

Topics

• Conditional Monge Gap
• Single-cell Perturbation Modeling
• Optimal Transport
• Drug Discovery
• Cellular Heterogeneity
• scRNA-seq

Code references

• AI4SCR/Conditional-Monge

Best for: AI Scientist, Research Scientist, Machine Learning Engineer

Related on AIssential

• Spectral Audit of In-Context Operator Networks — Prediction accuracy alone is insufficient for neural operators; local operator fidelity requires spectral auditing.
• The Ghost Couple: Correlated LLM Name Priors and Their Haunting of the Web and Academic Publishing — LLMs generate correlated fictional character ensembles, not just individual names, leading to widespread ghost authorship in academic repositories.
• Bridging the Sim-to-Real Gap in Semiconductor Visual Program Synthesis via Input Binarization — Input binarization effectively bridges the sim-to-real gap for VLMs in semiconductor visual program synthesis by abstracting texture.
• Topological texture analysis of microscopy images of dynamic casein gelation and its relation to rheological properties — The toolbox combines TDA, DBC, MFP, and LBP to quantify dynamic microstructural changes in gelation, correlating with rheological properties.
• Learning the coupled dynamics of global climate modes — UniCM unifies global climate mode forecasting by learning coupled ocean-atmosphere dynamics and inter-mode interactions, revealing emergent predictability.
• Science Earth: Towards A Planet-Scale Operating System for AI-Native Scientific Discovery — A planet-scale runtime enables AI capabilities to self-organize for distributed, self-correcting scientific discovery.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Nature Machine Intelligence.