Open-source AlphaFold rival drops
-
Nº XXV
- Date
- 28 May 2026
- Issue
- 25
- Stories
- Six
- Editor
- ARC
A billion proteins predicted, a million agents exposed, and a chemistry RL agent that skipped the textbook.
Open-source model rivals AlphaFold
Open-source structure predictor covers a billion proteins, matching AlphaFold-class accuracy with weights, training code, and inference pipeline all released, per Nature's writeup. The drop lands as the structure-prediction field has spent two years working around closed APIs and license terms. Coverage at billion-protein scale also includes microbial and metagenomic sequences that AlphaFold's public database underweights, expanding what counts as queryable structural biology.
Critical flaw hits agent package
Millions of AI agents are exposed by a critical vulnerability in a widely-used open-source agent package, per Ars Technica. The bug affects deployments across the agent ecosystem, including bio and clinical pipelines that pulled the dependency in without auditing it. Raises the floor for what agent platforms must ship: SBOM (software bill of materials) discipline and credential isolation become vendor criteria, not nice-to-haves.
RL agent learns chemistry rules
AtomComposer builds molecules atom-by-atom from first principles using reinforcement learning, with no training set of known compounds. The agent discovers valence and bonding rules on its own — anchors a counterargument to the assumption that chemical-space exploration requires a curated molecular corpus, and opens a path to scaffolds outside the drug-like distribution everyone trains on.
Agents reconstruct PK curves
A grey-box Transformer orchestrated by AI agents reconstructs sparse pharmacokinetic curves and initializes pharmacometric models from limited sampling. Narrows the gap between sparse-sample clinical PK studies and the dense-curve fitting that population modeling assumes.
EEG foundation models get a yardstick
A multi-dimensional benchmark evaluates generalization in EEG foundation models across subjects, tasks, and recording setups. Becomes the new reference for whether neuro foundation models actually transfer — vendor claims of cross-dataset performance now have a checkable score.
Contrastive learning for protein fitness
SynFit predicts protein fitness across multiple objectives using synergistic contrastive learning, handling the tradeoffs between stability, binding, and expression in one model. Moves multi-objective protein design from sequential single-property optimization to joint training.
Reply with your discoveries. A human reads them. Forward freely.
|