Deep learning-based predictions of gene perturbation effects do not yet outperform simple linear methods DOI Creative Commons
Constantin Ahlmann-Eltze, Wolfgang Huber, Simon Anders

et al.

bioRxiv (Cold Spring Harbor Laboratory), Journal Year: 2024, Volume and Issue: unknown

Published: Sept. 19, 2024

Abstract Advanced deep-learning methods, such as foundation models, promise to learn representations of biology that can be employed predict in silico the outcome unseen experiments, effect genetic perturbations on transcriptomes human cells. To see whether current models already reach this goal, we benchmarked five and two other deep learning against deliberately simplistic linear baselines. For combinatorial genes for which only individual single had been seen, find learning-based approaches did not perform better than a simple additive model. yet outper-form baseline predicting mean across training perturbations. We hypothesize poor performance is partially because pre-training data observational; show model reliably outperforms all when pre-trained another perturbation dataset. While neural networks representation biological systems prediction experimental outcomes plausible, our work highlights need clear setting objectives critical benchmarking direct research efforts. Contact [email protected]

Language: Английский

Deep learning-based predictions of gene perturbation effects do not yet outperform simple linear methods DOI Creative Commons
Constantin Ahlmann-Eltze, Wolfgang Huber, Simon Anders

et al.

bioRxiv (Cold Spring Harbor Laboratory), Journal Year: 2024, Volume and Issue: unknown

Published: Sept. 19, 2024

Abstract Advanced deep-learning methods, such as foundation models, promise to learn representations of biology that can be employed predict in silico the outcome unseen experiments, effect genetic perturbations on transcriptomes human cells. To see whether current models already reach this goal, we benchmarked five and two other deep learning against deliberately simplistic linear baselines. For combinatorial genes for which only individual single had been seen, find learning-based approaches did not perform better than a simple additive model. yet outper-form baseline predicting mean across training perturbations. We hypothesize poor performance is partially because pre-training data observational; show model reliably outperforms all when pre-trained another perturbation dataset. While neural networks representation biological systems prediction experimental outcomes plausible, our work highlights need clear setting objectives critical benchmarking direct research efforts. Contact [email protected]

Language: Английский

Citations

8