AI x Bio Daily Digest — 2026-11-14
**Saturday, 14 November 2026 — last ~36 hours. AMP 2026 (Seattle) dominates the clinical-genomics side; Nvidia's leadership collapse dominates everything else.**
Saturday, 14 November 2026 — last ~36 hours. AMP 2026 (Seattle) dominates the clinical-genomics side; Nvidia's leadership collapse dominates everything else.
Life sciences × AI
AlphaGenome 2 conditions variant-effect prediction on cell type and state. DeepMind's successor (announced 11 Nov, coverage and analysis landing 12–13 Nov) trains on ~107M single-cell profiles and conditions each prediction on a tissue, cell type or state — including states derived from your own scRNA-seq. On DeepMind's benchmark, recall of causal variants observed in single cells goes from 16% to 30% at matched 90% precision; in an IBD credible set it picked rs2836882 out of seven variants, linked it to ETS2 ~290 kb away, and localized the effect to gut monocytes/macrophages. Free for non-commercial research, private preview for commercial. ( https://www.google.com/url?q=http://implicator.ai&source=gmail&ust=1794714306027000&sa=E, https://www.google.com/url?q=http://unite.ai&source=gmail&ust=1794714306027000&sa=E )
The AlphaGenome numbers are all in-house, and v1 just took a hit. No independent evaluation of v2 has been reported; the gains come from an unreviewed company preprint. Separately, a Charité-led preprint posted 9 Nov tested AlphaGenome 1 on 2,111 in silico-disrupted promoter–gene pairs and found it frequently failed to propagate the promoter-disruption effect across the rest of the transcript. Worth weighting before building either model into a reporting pipeline. ( https://www.google.com/url?q=http://implicator.ai&source=gmail&ust=1794714306027000&sa=E )
ALMA 3: methylation foundation model for hematolymphoid classification. Presented at AMP 2026 by ALMA Genomics with Penn and MIT — trained on DNA methylation from >110,000 patient samples, covering 81 molecular subtypes. Independent testing: 99.5% sensitivity for detecting blood/lymphatic tumors, 90.9% PPV for calling the specific subtype. Methylation-classifier-as-foundation-model is now a credible competitor to morphology-plus-panel workflows in heme. (LabMedica)
Open-source evidence-grounded somatic reporting, with a hard number on hallucination. University of Michigan's Turnkey Precision Oncology (TPO-Agent) ingests DNA/RNA-seq and generates molecular pathology reports, verifying interpretations against cancer genetics databases and literature. Across 50 alterations in 10 patients, ~95% of claims were supported by cited evidence versus 65% for a conventional LLM approach; ~95% agreement with clinical lab findings across 28 reports. This is the closest public analogue to a VariantGPT-style system with a published grounding metric. (LabMedica)
AI triage of molecular testing: real tissue savings, uneven by alteration class. MSK + Mount Sinai had six molecular pathology fellows review 80 consecutive lung adenocarcinoma cases with AI-assisted slide assessment; 26 cases (33%) could have deferred rapid EGFR testing with no confirmed EGFR-positive case missed — preserving tissue for comprehensive profiling. The same approach was notably less reliable for deferring fusion testing. Stanford, on 367 gastric tumors, predicted mutations from H&E with a pretrained pathology foundation model: PIK3CA AUC 0.80, CDH1 0.77 — interesting for sequencing-limited settings, not yet a gate. (LabMedica)
Inocras presents first public results for agentic cancer WGS curation. Multimodal agentic framework for variant curation and clinical interpretation of cancer whole genomes, with expert-in-the-loop oversight — oral presentation I024 at AMP on 12 Nov, poster through 14 Nov. Same company that unveiled a WGS-trained cancer foundation model with learnable tokenization and is targeting a 100,000-patient genome program. (LabMedica)
argenx pays up to $1.15B for AI-guided oral macrocycles; Vertex publishes the counterweight. argenx acquired Unnatural Products and its AI macrocycle platform, though CDMO chemists flag that oral peptide bioavailability and CMC burden remain unsolved. Against that, Vertex told the WSJ it evaluated ~50 AI drug-discovery companies against a 10,000-compound database and none had independently generated a working drug — published the same week Amodei told an IISc audience AI could compress drug development from ~10 years to months. (PharmTech, American Bazaar)
Frontier & industry
Jensen Huang resigned as Nvidia CEO on 12 November. Departure followed reporting on a relationship with a woman suspected of spying for China; CFO Colette Kress is interim CEO, Harvey Jones becomes first board chair. Nvidia closed below $4T market cap on Friday for the first time since July 2025, Nasdaq 100 fell 1.2%, DOJ is probing possible chip-smuggling links, and Huang has agreed to testify before House committees. The $12.9B Hugging Face acquisition is still expected to close in 2027. Nvidia underwrites an estimated $400B in guarantees across the AI buildout — the financing question is the real second-order risk. (CNBC, CNBC)
Anthropic's IPO is six times oversubscribed into next week's listing. Amended 10 Nov prospectus proposes 825M shares at $70–78, listing on Nasdaq as ANT. The filing discloses ~$900.9B in committed spending, substantially all compute and infrastructure, with ~$70.7B due within 12 months and the remainder stretching to 2045; Nvidia has agreed to buy another $10B of Class N stock immediately post-offering. (Bloomberg, Sam on AI)
OpenAI closed $15B at a $1.25T valuation, down from an originally planned $30B at $1.4T. ChatGPT Enterprise also went live on AWS Marketplace, letting spend count against existing AWS commitments — a distribution and procurement move, not a capability one. (CNBC, Sam on AI)
Frontier board as of 13 Nov: Claude Fable 5.2 (9 Nov) leads; Gemini 4 Carbon went GA to paid API developers 10 Nov; Gemini 4 Hydrogen and Qwen4-Max-Preview landed 11–12 Nov. Nine confirmed releases in the 7–13 Nov window. ( https://www.google.com/url?q=http://benchlm.ai&source=gmail&ust=1794714306027000&sa=E )
Research, tools & repos
Two bioRxiv preprints posted 10 Nov say single-cell foundation models still don't beat simple baselines. Cupaioli and Merelli (CNR-ITB) benchmarked scGPT, Geneformer V1/V2, Nicheformer and scFoundation against CellOracle, STATE and task-matched baselines across K562 Perturb-seq, primary cells, isogenic leukodystrophy brain organoids, an organoid CRISPR screen and lung cancer biopsies: within species, FM embeddings annotated cell types no better than principal components. Souza and Mehta reach the same conclusion independently. If you're scoping scFM work for Panbiome or MCED, these are the papers to read before committing GPU budget. (Cupaioli & Merelli, Souza & Mehta)
DECONVersation — scFM embeddings as a deconvolution front end. New York Genome Center (posted 10 Nov) pairs pretrained or fine-tuned single-cell foundation model embeddings with non-negative least squares for bulk RNA-seq deconvolution, eliminating explicit marker selection and reducing batch effects. A narrow, defensible use of scFMs that survives the benchmark critique above — embeddings as features, not as a general-purpose reasoning layer. (bioRxiv)
Worth a second look
TPO-Agent's 95%-vs-65% evidence-grounding delta is the metric to steal. It's the first public clinical-genomics agent result that separates "the report was right" from "the report's claims were supported by what it cited" — exactly the failure mode that blocks regulatory and CAP-style acceptance of LLM reporting. Instrumenting VariantGPT with an equivalent citation-verification score would give you a defensible validation artifact rather than a concordance number.
Cell-state-conditioned variant effects are a near-term VUS lever, not a distant one. AlphaGenome 2 takes conditions generated from a researcher's own single-cell experiment. For hereditary cancer screening and rare-disease exomes where the bottleneck is non-coding and regulatory VUS, that's a plausible path to tissue-specific functional evidence — tempered by the fact that every published number so far is DeepMind's own, and v1 just failed an external promoter-propagation test.
Comments ()