Double perturbation prediction. credit: Natural Method (2025). doi:10.1038/s41592-025-02772-6
Deep learning models show great potential in predicting and engineering functional enzymes and proteins. Does this skill extend to other fields of biology?
Contrary to expectations, recent studies have found that deep learning-based basic models are not superior to simple baseline methods in predicting how genetic perturbations (changes in gene expression or function) affect the transcriptome, a cell's gene expression profile. In the case of double perturbations where two genes were simultaneously altered, we added a combined effect of gene change, instead of using complex machine learning, as the prediction error was higher in the deep learning model compared to the baseline additive model.
The basic model is a deep learning model trained with a huge amount of data. In this context, they refer to single-cell models trained with recently published Transcompritomics data covering millions of cells.
Published in Nature,In this study, we leveraged published single-cell CRISPR perturbation datasets to benchmark five prominent basic models, including SCGPT and SCFoundation, addressing two other deep learning models for four intentionally simple baselines.
Recent research on basic models based on deep learning aims to revolutionize understanding of biology by training vast amounts of data in the hope that the model will generally understand how cells work, rather than remembering specific experimental results. This ability allows for prediction of results without experimentation, greatly accelerating drug discovery and disease research.
Single perturbation prediction. credit: Natural Method (2025). doi:10.1038/s41592-025-02772-6
However, biology is a very complex science, with the behavior of cells, genes, and organisms dependent on many factors, many of which have not been discovered. The models developed to understand these complexities are very computationally expensive as they require time, energy and powerful machines. Before pouring further resources into building such models, it is important to pause and ask: are they really effective and are they better than the models we already have?
Previous studies conducted benchmark experiments, but most of them pitted one deep learning model with another deep learning model, lacking comparisons with simple models. The researchers in this study aimed to change this by comparing a simple, interpretable baseline model with a complex model.
They found that complex models, such as non-modified, mean, or linear model-based predictions, do not consistently outperform simple baselines when predicting the effects of single or double perturbations on gene expression. Most models struggled to accurately predict complex genetic interactions.
These findings made a great deal of clarity that higher costs and complexity do not necessarily translate to better performance compared to simpler, resource intensive methods. It also established the importance of rigorous testing of new models against existing models and benchmarking.
Researchers have learned a generalizable understanding of cellular states and concluded that the ambitious goals of a fundamental model for predicting this knowledge-based outcome are still out of reach.
Written for you by author Sanjukta Mondal, edited by Gaby Clark, fact-checked and reviewed by Robert Egan. This article is the result of the work of a careful human being. We will rely on readers like you to keep independent scientific journalism alive. If this report is important, consider giving (especially every month). You'll get No ads Account as a thank you.
detail:
Constantin Ahlmann-Eltze et al, Deep learning-based gene perturbation effect prediction has not yet exceeded a simple linear baseline; Natural Method (2025). doi:10.1038/s41592-025-02772-6
©2025 Science X Network
Quote: Complex deep learning models are less good at understanding genetic perturbations than simple baseline ones, the study obtained on August 16, 2025 from https://phys.org/news/2025-08.
This document is subject to copyright. Apart from fair transactions for private research or research purposes, there is no part that is reproduced without written permission. Content is provided with information only.
