A study published in Nature Machine Intelligence, and featured on the journal’s cover, offers a way to predict how cell populations will react to a drug before any laboratory testing. Rather than scaling up model size, the CMonge method (Conditional Monge Gap) draws on the mathematical framework of optimal transport to learn how entire populations of cells shift in response to compounds, doses and drug combinations.
Using a fraction of the parameters required by comparable approaches, it generalises to compounds it has never encountered while preserving the heterogeneity of cellular data, one of the hardest aspects of this type of data. The work was led by Marianna Rapsomaniki, UNIL professor and group leader at the CHUV’s Biomedical Data Science Center, together with Jannis Born (IBM Research), and carried out by Alice Driessen, Dhruva Rajwade and Benedek Harsanyi.