Back

DINOcell: Learning Generalizable Perturbation Effects through Self-Distillation

Li, J.; Tasdemir, N.; Kyriazopoulou Panagiotopoulou, S.; Xu, S.; Merrell, A.; Susanto, J.; Cass, A.; DeTomaso, D.; Steach, H.; Chow, J.; Wang, L.; Ravindra, N.; Kang, A.; Zheng, G.; Chua, M.; Knudsgaard, P.; He, E.; Ingrao, F.; Boroughs, A.; Copperman, B.; Charrington, N.; Abboud, K.; Berthoin, L.; Baca, D.; Leon, L.; Kotov, J.; Wozniak, G.; Cardozo, A.; Vucci, K.; Barner, A.; Yi, C.; Oh, C.; Simon, M.; Paliwal, S.; Drever, M.; Galvin, B.; Tan, M.; Guzman, G.; Loh, K.; Ong, A.; Sandoval, M.; Lim, M.; Ng, E.; Lincoln-Cabatu, B.; Wu, J.; Nguyen, A.; Patel, M.; Fua, A.; Wong, K.; Chen, J.; Yashin

2025-12-19 bioinformatics
10.64898/2025.12.16.694747 bioRxiv
Show abstract

AO_SCPLOWBSTRACTC_SCPLOWPredicting cellular responses to therapeutics is a promising approach for novel target discovery. However, state-of-the-art computational models designed to predict perturbation effects struggle to generalize and outperform simple baselines. We present DINOcell, a weakly supervised framework that adapts self-distillation to single-cell transcriptomics for predicting perturbation effects. We demonstrate that DINOcell outperforms baselines in predicting the effects of single gene perturbations. Furthermore, DINOcell accurately predicts non-additive effects of combination perturbations, indicating its capacity to model complex genetic interactions. Finally, we show that DINOcell learns representations that capture biological signals and is a promising, generalizable approach for in silico perturbation modeling, providing a valuable tool for accelerating therapeutic target discovery.

Matching journals

The top 4 journals account for 50% of the predicted probability mass.

50% of probability mass above

"Similar papers" are the closest papers from that journal in the model's embedding space. They show what the match is built on, but the ranking comes mostly from a classifier over the whole training set, not from these examples alone.