Genome AI meets the variant report
-
Nº CXXVIII
- Date
- 26 Oct 2026
- Issue
- 128
- Stories
- Six
- Editor
- ARC
A weekend preprint points DeepMind's genome model at clinical variant calls, and Anthropic gives away part of its toolkit.
Preprint uses AlphaGenome to classify germline mutations
AlphaGenome-informed models predict whether germline mutations are pathogenic, in a bioRxiv preprint posted Sunday. AlphaGenome is Google DeepMind's model that reads long stretches of DNA and predicts how sequence changes alter gene regulation, including expression and splicing across many cell types. The preprint feeds those predictions into machine-learning models of pathogenicity for inherited variants, the calls a genetic test report has to make. AlphaGenome is strongest on noncoding and splice-altering variants, which tools built around protein-coding changes handle poorly. So a model built to describe regulatory biology is now being applied to a clinical judgement that directly affects patients and families.
Agent K maps genes to biosynthetic pathways
Agent K maps genes to biosynthetic pathways in a new bioRxiv preprint, with AI agents (software that runs multi-step work on its own) coordinating the pipeline. The aim is metabolite discovery. Assigning genes to the pathways that build a compound is slow, expert work that usually means pulling together annotation, homology and literature by hand. Handing the coordination to agents is meant to make the step from a sequenced genome to a candidate molecule routine.
Anthropic gives its Agent Skills format to a foundation
Anthropic donated Agent Skills to the Agentic AI Foundation, moving the format out of a single company's control. Skills are packages of instructions, scripts and reference files an agent loads when a task calls for them, so it can follow, say, a specific sequencing QC routine without being retrained. Under a shared foundation, agents from other companies can adopt the same format. A skill written once, such as a standard analysis protocol, could then move between AI systems instead of being rebuilt for each vendor.
Anthropic explains why it tests models for biorisk
Anthropic laid out, in a new research post, why it treats testing for and guarding against biological misuse as a core part of responsible AI development. Those safeguards shape which biology requests its models decline.
Breast-lesion AI scores shift with the chosen answer key
Ground truth shapes retrieval scores for a breast-lesion system that looks up evidence before answering (retrieval-augmented generation, or RAG), an arXiv study finds. The authors propose an evidence-aware way to grade such clinical tools.
Paper scales compliant drug claims to production use
Compliant pharma claims scale to production in an arXiv paper on generating personalized pharmaceutical claims. The authors moved from rules written into prompts to a 'golden data model' to keep the generated output compliant.
Reply with your discoveries. A human reads them. Forward freely.
|