Does pretraining pay off for single-cell biology?
-
Nº CXXXIX
- Date
- 10 Nov 2026
- Issue
- 139
- Stories
- Six
- Editor
- ARC
Today the big single-cell models lose to simpler ones, and phase I trial design gets its own AI exam.
Simple models beat big pretrained single-cell AI models
Simple models outperform single-cell foundation models on downstream benchmarks, according to a new bioRxiv preprint. Foundation models here are large AI models pretrained broadly on single-cell data and then reused for many tasks, rather than built for any one of them. Downstream benchmarks are shared tests of whether that pretraining helps once a model is put to a specific job. The result bears on the main promise of these models, which is that broad pretraining yields a representation of cells that simpler methods can't match. Before judging any one model, it's worth checking how wide the margin is and which tasks it holds on.
TrialDesignBench grades AI on phase I trial design
TrialDesignBench grades language models on phase I trial design, scoring their proposals against established clinical knowledge, according to a new arXiv preprint. A benchmark in this sense is a shared test, a fixed set of problems with reference answers, so different models are compared on equal footing. Phase I is where first-in-human decisions carry the most direct safety stakes. Anyone claiming a model can help draft a trial protocol now has a published standard to measure it against.
Stanford and MIT labs run genome-wide studies with Claude
Stanford and MIT labs built systems on Claude, Anthropic's AI model, that run genome-wide studies in minutes and automate the interpretation of gene clusters, according to an Anthropic research post. The post is the company's own account, so the speed figures come from Anthropic, not from an independent evaluation. Both are working systems built inside university research groups, not product demos put together by the vendor.
SciMARS lets AI research systems improve themselves
SciMARS lets AI agents (software that runs multi-step work on its own) improve their own methods round by round as they carry out closed-loop scientific discovery, according to an arXiv preprint. The name stands for Multi-Agent Recursive Self-Improvement.
GeneEmbedLLM predicts gene links with language models
GeneEmbedLLM uses language models to represent gene function as embeddings, lists of numbers in which related genes sit close together, and then uses those representations to predict associations between genes, according to a bioRxiv preprint.
Anthropic explains how Claude sorts what to remember
Anthropic detailed how Claude decides what to remember before it knows what will be useful, and described a process it calls "dreaming" that reorganizes stored memories and corrects mistakes. The post presents memory as what lets Claude learn from experience.
Reply with your discoveries. A human reads them. Forward freely.
|