Altos Labs algorithm predicts cell responses to interventions
What if you could test how a cell responds to a genetic intervention before running a single experiment?
In a preprint posted to arXiv, the team described a machine learning algorithm capable of end-to-end prediction of how gene expression changes in response to specific interventions. Gene expression is the process by which instructions in DNA are converted into proteins. It determines which genes are active and at what level. Predicting that reliably has been a longstanding challenge, because cells involve thousands of simultaneous interacting components.
Why this is harder than protein folding
Tools like AlphaFold transformed structural biology by predicting the three-dimensional shape of proteins from their amino acid sequences. But the full biochemistry of a living cell is, according to the researchers, orders of magnitude more complex. Proteins, RNA, metabolites, signalling pathways, and environmental inputs all interact at once.
The new algorithm takes a different approach from earlier models. It operates end-to-end: given a description of an intervention, such as knocking out a gene or introducing a compound, it predicts the full downstream pattern of altered gene activity. Previous approaches required combining multiple separate models, compounding the risk of error at each step.
The relevance for aging research
For longevity science, this matters for a practical reason. The number of possible interventions is far too large to test in the laboratory one by one. Which gene to silence? Which molecule to add? Which combination of changes has the best chance of slowing cellular aging? A reliable predictive model can dramatically narrow that search space and set priorities for follow-up experiments.
The study remains a preprint and has not yet undergone formal peer review. But the approach, using machine learning for end-to-end prediction of cellular behaviour, is widely considered a promising direction for accelerating biological discovery.