Biomni and TusoAI built an open predictor of cellular responses to gene knockouts in one week

Biomni and TusoAI built an open predictor of cellular responses to gene knockouts in one week

On July 30, Phylo introduced Biomni × TusoAI, a system for automating the development of biological models. According to the company, the system evaluated about 500 variants and produced TusoPerturb, an open method for predicting changes in gene activity after CRISPR is used to knock out another gene.

CRISPR can be used to knock out a selected gene. The cell then changes the activity of many other genes, increasing the expression of some and reducing the expression of others. Researchers face a long list of possible knockouts and must decide which experiment to run next. A genetic perturbation predictor learns from previous experiments and produces a table of expected changes in gene activity before the experiment is performed.

Biomni × TusoAI assembled the predictor from several components. The system tested which information about a gene to use as input, how to prepare the data, and which algorithm to use for prediction. TusoPerturb draws on biological reference data, including the gene's role in cellular processes, its relationships with other genes, and information about cell lines. For some tasks, the method uses regression. For others, it also identifies similar genes and transfers their measured responses.

Phylo reports that the method outperformed previous approaches on three independent test sets, including large deep learning models. The TusoPerturb repository provides the code, configuration, and a description of the path from input data to prediction. Other groups can use the same configuration, run it on supported datasets, and compare the results with those of their own methods.

In the Virtual Cell Challenge Arc, most models performing the same task were outperformed by a simple method for predicting the complete gene expression profile. Against this background, TusoPerturb gives researchers a specific method that they can evaluate using the same comparison.

In aging research, such a predictor allows researchers to compare in advance how a cell might respond to knocking out different genes. This makes it easier to select genetic hypotheses for the next experiment, while the open code preserves the path from the biological data to that decision.