Isomorphic goes beyond AlphaFold
-
Nº LIX
- Date
- 20 Jul 2026
- Issue
- 59
- Stories
- Eight
- Editor
- ARC
Monday reset: Isomorphic pushes past structure prediction into design, and a K99 at the 1st percentile lands a multimodal oncology lab.
Isomorphic ships drug-design engine
Isomorphic Labs unveiled its Drug Design Engine, positioning the platform as the step beyond AlphaFold — from predicting protein structures to generating candidate molecules against them. The engine couples structure prediction with generative chemistry and affinity ranking inside one loop, the piece Alphabet's drug-discovery spinout has been building toward since spinning out of DeepMind. Framed as a general-purpose design system rather than a target-specific tool, it lands as the most concrete answer yet to whether structure-prediction leaders can convert accuracy gains into real pipelines.
Multimodal oncology lab funded at 1st percentile
Kevin Boehm's K99/R00 cleared NCI review at the 1st percentile to launch a 2027 lab fusing tumor histopathology with genomics for treatment-selection AI. Multimodal cancer models — image plus omics in one predictor — have been the field's most-promised, least-delivered category; a top-percentile NCI bet signals which architecture the agency thinks is closest to clinical utility.
BioReason-Pro pushes protein function prediction
BioReason-Pro fuses sequence, structure, and text into a single reasoning model for protein function prediction, reporting gains on standard benchmarks over sequence-only baselines. Advances the multimodal-reasoning frontier for functional annotation — the workflow most affected by quarterly UniProt updates and where single-modality models have been the ceiling for two years.
Perturbed flow matching for SBDD
PFM applies perturbed flow matching — a generative technique that learns molecular distributions by mapping noise to structure — to structure-based drug design, improving pocket-conditioned ligand generation. Sharpens the generative-chemistry toolkit that Isomorphic's engine above is productizing at scale.
Clinicians build medical-AI failure benchmark
MedFailBench probes where medical AI models cross safety boundaries — hallucinated dosing, missed red-flag symptoms, unsafe deferrals — using cases written by practicing clinicians. Anchors a reference benchmark for medical-AI safety claims; vendor pitches now have a clinician-authored floor to clear — a gap we've tracked since exam scores stopped predicting bedside performance.
Perturb-and-record assays as next frontier
Live-cell imaging plus time-resolved chemical composition — perturbing cells and recording exactly what changes — is the assay upgrade drug testing still mostly lacks. Names the capability gap between watching cell state and controlling it, the bottleneck most compound-screening pipelines still route around.
AutoSynthesis automates meta-analysis
AutoSynthesis runs meta-analyses end-to-end — search, screening, extraction, synthesis — as an agentic pipeline, targeting the months-long human bottleneck in evidence synthesis. Moves systematic-review automation from demo to workflow-viable.
OpenAI's GPT-Red self-improves on safety
OpenAI released GPT-Red, an automated red-teaming system that uses self-play to harden models against jailbreaks and prompt injection. Raises the floor for what counts as adequate safety testing before a biomedical agent touches patient data.
Reply with your discoveries. A human reads them. Forward freely.
|