New AI Algorithm Predicts How Genes Respond to Treatments With Record Accuracy
Altos Labs' PRiMeFlow algorithm predicts cellular gene expression changes from interventions, outperforming rivals on key benchmarks.
Summary
Researchers at Altos Labs have developed a machine learning algorithm called PRiMeFlow that predicts how cells' genes respond to experimental interventions — without needing to run every experiment in a lab. Working directly in gene expression space rather than compressing data, the algorithm topped performance charts on three major benchmarks and matched real lab results more closely than any competing model when tested on human embryonic stem cells. This kind of tool could dramatically speed up the search for aging interventions by filtering promising candidates computationally before costly and time-consuming biological experiments begin. While still a preprint and far from clinical application, it represents a meaningful step toward building virtual cell models that could one day simulate entire organisms.
Detailed Summary
As the search for longevity interventions accelerates, one major bottleneck is the sheer number of possible genetic and molecular targets researchers must test. A new machine learning algorithm called PRiMeFlow, developed by scientists at Altos Labs and described in an arXiv preprint, aims to cut through that complexity by predicting how cells will respond to perturbations before a single experiment is run.
Unlike previous approaches that compress gene expression data into simplified lower-dimensional representations, PRiMeFlow operates directly within the full gene expression space. This preserves more biological detail and allows the model to make richer, more accurate predictions. Surprisingly, the team used a U-net architecture — typically designed for spatially structured data like images — rather than the conventionally preferred multi-layer perceptron. They found the U-net outperformed alternatives despite its seemingly poor fit for the task, though they acknowledge they cannot yet explain why.
In head-to-head testing on the PerturBench platform, PRiMeFlow achieved state-of-the-art results on three benchmarks, including two measuring how well it generalizes to cell types not seen during training. On a private test set of human embryonic stem cells, the fine-tuned model produced predictions closer to real laboratory results than any other model on the leaderboard.
For longevity research, the implications are significant. Tools like this could help researchers rapidly identify which genetic targets or small molecules are most likely to beneficially alter aging-related cell behavior — narrowing a near-infinite search space before expensive in vitro or animal studies begin. Altos Labs, notably, is a company specifically focused on cellular rejuvenation and longevity biology.
Important caveats apply: this is a preprint, not yet peer-reviewed, and the researchers openly note that substantial computational and algorithmic challenges remain. The vision of fully simulated virtual organisms is still distant. But PRiMeFlow marks a credible, measurable advance in the tools available to aging researchers.
Key Findings
- PRiMeFlow predicts gene expression responses to interventions directly in full expression space, preserving more biological detail.
- Outperformed competing models on 3 PerturBench benchmarks, including generalizing to unseen cell types.
- Fine-tuned model matched real lab results more closely than any rival on a human embryonic stem cell test set.
- Unexpectedly, a U-net architecture outperformed the standard model choice, raising new questions about AI design for biology.
- Could accelerate longevity drug discovery by filtering promising targets computationally before lab experiments begin.
Methodology
This is a research summary based on an arXiv preprint from Altos Labs, meaning it has not yet undergone formal peer review. The source, Lifespan.io, is a credible longevity-focused nonprofit outlet. Evidence is computational and benchmark-based, with no clinical or animal data presented.
Study Limitations
The paper is a preprint and has not been peer-reviewed, so findings should be treated as preliminary. Benchmark performance does not guarantee real-world biological accuracy across diverse cell types or disease contexts. The authors themselves note unresolved questions about why their architectural choices work, suggesting the model is not yet fully understood.
Enjoyed this summary?
Get the latest longevity research delivered to your inbox every week.