AlphaFold learns protein partners
Good Morning!
AlphaFold grows up, curated agents outpunch frontier giants, and chemistry gets an evolutionary committee.
1. AlphaFold adds protein pairing
AlphaFold's database now predicts which proteins bind which, expanding from solo-structure prediction to pairwise interaction calls across the proteome. The DeepMind-EMBL release adds confidence scores for millions of candidate complexes, the layer everyone has been hand-rolling on top of monomer predictions for two years. Reviewers in Nature describe it as the database's "next level" — moving structural biology's reference resource from "what does this protein look like" to "who does it talk to." Read More →
Why it matters: Interaction prediction at database scale resets the cost ceiling for hypothesis generation in cell biology — anyone designing a screen, mapping a pathway, or prioritizing a drug target now starts from a precomputed interactome instead of a blank docking job.
2. Curated AI beats frontier LLMs
A purpose-built curated model outperforms GPT-class frontier LLMs on pharmaceutical asset discovery, according to a new arXiv paper from Kidziński and colleagues. The system pairs domain-specific training data with retrieval over licensing and pipeline databases, suggesting that for narrow, high-stakes pharma workflows, curation still beats raw scale — a counterargument to the "just use the biggest model" default that has dominated bio-AI vendor pitches. Read More →
3. Evolutionary multi-agent for synthesis
An evolutionary multi-agent framework lets LLM reasoners propose, critique, and mutate synthetic routes for target molecules, with a chemistry-knowledge module scoring each generation. The bioRxiv preprint reports gains over single-agent baselines on retrosynthesis benchmarks — pushing autonomous chemistry agents from one-shot suggestion toward iterative search, the regime medicinal chemists actually work in. Read More →
4. LLM pipeline maps PTMs
Post-translational modifications get an LLM-driven detection-and-modeling pipeline that pulls candidate PTMs from proteomics output and threads them into structural predictions. Closes a longstanding gap where PTM calls and structural context lived in separate tools. Read More →
5. Causal flows handle missing EHR data
Treatment-effect estimation on incomplete healthcare records gets a temporal causal normalizing flow with LLM-driven imputation for missing-not-at-random data. Targets one of the dirtiest problems in real-world-evidence work, where missingness itself encodes clinical decisions. Read More →
6. OpenAI cuts agent-loop latency
OpenAI rewired Codex's agent loop with WebSockets and connection-scoped caching in the Responses API, cutting per-step overhead on long agent runs. Lowers the latency tax on multi-tool agent workflows — the regime where most production biology agents now live. Read More →
7. ChatGPT workspace agents launch
OpenAI shipped workspace agents in ChatGPT, packaging tool-connected automations for team-shared workflows. Brings agent deployment closer to the no-code tier where most lab ops actually get built. Read More →
Reply with your discoveries. A human reads them. Forward freely.