Making sense of the frontier where AI meets biology.
Joined November 2026
- Tweets22
- Following7
- Followers116
- Likes18
Nature Communications: Predict the mutation effect, not just absolute kcat — DeltaCata then improved a methyltransferase’s catalytic efficiency for icaritin up to 7.0× in vitro.
Most kinetic models predict absolute kcat and Km and miss what a mutation actually changes. Yuan, Yang, Lu and colleagues at Sun Yat-sen University and the National Supercomputing Center in Shenzhen built DeltaCata, a Siamese geometric graph network that learns Δkcat and ΔKm directly from wild-type/mutant pairs measured under identical conditions.
Hard readout
→ They assembled 12,390 Δkcat and 16,010 ΔKm mutation-effect pairs from matched assays (language-model-assisted parsing plus manual review)
→ On sequence-level tests, DeltaCata reached PCC 0.568 for Δkcat and 0.364 for ΔKm, ahead of existing predictors; the lead holds at low sequence identity
→ Applied to engineer methyltransferase sGmOMT2; wet-lab catalytic efficiency (kcat/Km) for in vitro icaritin synthesis improved by up to 7.0-fold
Worth reading when enzyme models still report a single kcat and leave the mutation effect as an afterthought.
📄 Paper: nature.com/articles/s41467-0…
💻 Code: github.com/biomed-AI/DeltaCa…
#AIforScience #EnzymeEngineering #OpenSource
ALT DeltaCata enzyme mutation-effect prediction improving methyltransferase for icaritin
Nature Communications: Affinity or stability? LAffAb co-optimizes both — and designs didn’t collapse to one winner.
Tennenhouse, Fleishman (Weizmann) and AstraZeneca colleagues start from an antibody–antigen crystal structure and introduce combinatorial CDR mutations that improve the energy of the complex, instead of iterating affinity at the cost of stability.
Hard readout
→ A ~7,000-variant library with up to nine mutations; wet-lab SPR gains of up to ~30× affinity while co-optimizing stability
→ The library did not converge on a single solution. Selection favored diverse variants with a high mutational load
→ A 10-design screen against a potential drug target improved affinity by an order of magnitude with high developability; GitHub release of the RosettaScripts workflow
Worth reading when affinity maturation still feels like a serial trade-off.
📄 Paper: nature.com/articles/s41467-0…
💻 Code: github.com/Fleishman-Lab/LAf…
#AIforScience #AntibodyDesign #OpenSource
ALT LAffAb antibody affinity–stability co-optimization from Nature Communications
Ginkgo sold its first autonomous lab to Novo Nordisk — not a paper, a purchase order.
In its Q3 2026 results (5 Nov), Ginkgo said Novo Nordisk selected it to build an autonomous lab at Novo’s new Waltham, Massachusetts R&D site. This is hardware plus software (RAC + Catalyst) going into a pharma’s own U.S. lab — not another services-only CRO story.
Hard readout
→ First announced autonomous-lab sale to a top-tier pharma company
→ CEO Jason Kelly: the U.S. government is now treating autonomous labs as national infrastructure; the Novo deal “sends a signal to all the other companies in the industry”
→ Q3 revenue $28M (−6% year over year); cash, cash equivalents and marketable securities $285M as of 30 Sept 2026
Worth reading when the question is who is actually buying AI-native wet labs, not who is publishing them.
📄 Earnings PR (SEC): sec.gov/Archives/edgar/data/…
🎧 Earnings call: fool.com/earnings/call-trans…
#AIforScience #LabAutomation #Biopharma
ALT Ginkgo Bioworks autonomous lab sold to Novo Nordisk at Waltham R&D site
AI can accelerate a discovery without resolving its biggest mystery
Anthropic’s biology team — directing Claude at each step — reports the first 3D structure of a vault-like particle assembled from a bacteriophage protein (L05): 96 copies, ~9.5 MDa, shorter/wider than mammalian vaults. Function of vaults remains unknown after ~40 years.
Hard readout
→ Isolation cut from ~3 days to ~90 minutes; first structural model in ~34 hours of processing
→ 7 of 15 expressed candidates formed vault-like particles; cryo-EM confirmed four
→ Humans set goals and wet-lab decisions; Claude assisted search, construct design, purification, and map processing — not a solo AI discovery
Worth reading as an experiment-story on AI-in-the-loop biology — and as a reminder that speed ≠ solved biology.
📄 Anthropic: anthropic.com/research/inves…
📰 Science: science.org/content/article/…
#AIforScience #StructuralBiology #DrugDiscovery
ALT Anthropic research image of a vault-like particle assembled from bacteriophage protein L05
Nature: RNA pretraining that needs base-pairing — not just more sequences
structRFM (Zhu, Xiong, Zhou et al.) is an open structure-guided RNA foundation model trained on 21 million sequence–structure pairs, folding base-pairing into masked language modelling.
Hard readout
→ New secondary-structure benchmarks; tertiary adaptation Zfold reported as comparable to AlphaFold3 on CASP15-natural, CASP16, and RNA-Puzzles
→ Also lifts function tasks and generalizes to new families / lncRNAs
→ Full 21M-pair dataset, weights, and code released (GitHub + Zenodo)
Worth reading when sequence-only RNA LMs feel underpowered for structure.
📄 Paper: nature.com/articles/s41586-0…
💻 Code: github.com/heqin-zhu/structR…
📦 Data: doi.org/10.5281/zenodo.18272…
#AIforScience #RNA #OpenSource
ALT Nature Fig. 1 overview of structRFM, a structure-guided RNA foundation model
Nature: AI agents that design the lab layout — then run it
Most self-driving labs only execute the workflows they were built for. Self-Building Labs (Bai, Zhao, Aspuru-Guzik & co., U Toronto / Acceleration Consortium) start from an empty room or an existing setup and fill the gaps: layout, digital twin, robot skills, and workflows.
Hard readout
→ In a real quantum-dot run: 236 OT-2 liquid-handling commands + 19 robotic transfer/seal/readout stages
→ Twin-first planning, then physical execution under human approval
→ Still a preprint — news coverage, not peer-reviewed validation of a universal unmanned lab
Worth reading when you care about AI-native lab flexibility, not just protocol execution.
📄 Nature: nature.com/articles/d41586-0…
📄 Preprint: arxiv.org/abs/2610.23826
#AIforScience #SelfDrivingLab #LabAutomation
ALT Nature news image of a robot-powered Self-Building Lab. Credit: Jiaru Bai and Yuchi Zhao, University of Toronto
🧬 Can your binder tell mutant KRAS from wild type?
That’s the final task in one of two AI protein design competitions still accepting submissions this November.
🏆 Anthropic × Adaptyv
• Final challenge: design a binder that recognizes KRAS G12D pMHC while avoiding the wild-type complex and peptide-free HLA.
• Across all five challenges, the organizers plan to test 5,000+ designs and publish the experimental results on Proteinbase.
📅 Open-track deadline: Nov 8, 23:59 AoE
proteinbase.com/competitions…
🏆 Changjiang Cup 2026
• 1,800+ teams and 4,000+ participants from 20+ countries and regions
• ~RMB 1.4M in cash prizes and 7,000 free protein measurements
• 7 target-specific tracks for de novo VHH or minibinder design, with rankings based on experimentally measured Kᴅ
📅 Register and submit by Nov 30, 23:59 Beijing time
aiprotein.sjtu.edu.cn/
#ProteinDesign #AIforScience #AIDrugDiscovery #Adaptyv #Anthropic
🧪 Library-scale screening — with a durable lab notebook the agent actually keeps.
El Agente Fármaco (Aspuru-Guzik / Matter Lab et al.; arXiv:2611.03832) isn't another one-shot docking chat. Fármaco Atlas stores molecules, metrics, and provenance in persistent SQL while tools fan out over whole subsets — no per-compound LLM call.
Hard readout:
→ 22-task benchmark: highest mean model-judged score vs plain tools and Codex, with lower median tokens and peak context
→ TBXT: 705 → 34 selected vs expert 39 purchased; common Vina competitive (best −8.5 vs −8.4)
→ Mean QED 0.69 vs 0.52; max Tanimoto 0.30, no shared Murcko scaffolds (Caveat: agent constraints + computational only.)
Worth reading when your "AI scientist" still forgets yesterday's molecules.
📄 arxiv.org/abs/2611.03832
💻 github.com/marcelmbn/farmaco…
#AIforScience #DrugDiscovery #Agents
ALT El Agente Fármaco Atlas overview — persistent memory and library-scale subset screening (arXiv:2611.03832 Fig.1)
🧬 One co-design model for proteins, nucleic acids, and small molecules — with a free web server.
PartiDesign (Jin, Zhang, Kang, Hou; Zhejiang University; bioRxiv 2026.11.03.763860) isn't another modality-specific generator. It jointly builds 3D structure and sequence/chemistry under motif, secondary-structure, and bond controls — including cyclic peptides and antibodies. Trained on 128 Ascend 910B GPUs.
Hard readout:
→ ProtDBench α: ordinary MPNN@1 pass rate 23.3% vs PXDesign 21.2%, BoltzGen 12.1%, RFD3 9.8%
→ Estimated GPU-day Self yield: α 16.5k passing sequences vs BoltzGen MPNN 2.7k and RFD3 1.7k
→ Peptides (Self+1 MPNN): ~17.0k linear / ~13.8k cyclic successful backbones per GPU-day vs BoltzGen ~9.1k / ~9.8k (Caveat: AF2-IG / compute-budget estimates; preprint.)
Worth reading when your binder workflow still needs a different model for each molecular class.
📄 biorxiv.org/content/10.64898…
💻 github.com/MegaGatlingPea/Pa…
🌐 cadd.zju.edu.cn/partidesign/
#AIforScience #DrugDiscovery #OpenSource
ALT PartiDesign framework overview: all-to-all co-design for proteins, nucleic acids, and small molecules (paper Figure 1)
🧬 Alex Rives: the virtual-cell bottleneck isn’t a bigger transformer — it’s measurement.
On The Cognitive Revolution (via AI4S.fyi), Biohub’s Head of Science draws a hard line: models called “virtual cells” today don’t yet simulate a cell in a meaningful way, or serve as an oracle for experimental biology.
His split of the problem:
→ Model — transformers lack multiscale cell structure
→ Data — programmes of the cell at scale
→ Measurement — “even more fundamental”; most biology leaps came from new ways to see
What he thinks is reachable in ~5 years (hedged): from a cell’s state + an intervention, predict a distribution over next states.
Far off: a full “world model of life” at atomic resolution.
On the $1.8B Virtual Biology Initiative: first clear push = spatial maps linking molecular measurements to imaging at hundreds of billions of cells — atlases of health/disease in 1–2 years (first drafts), more comprehensive ~5 with coordinated effort.
📄 Interview digest: ai4s.fyi/interviews/biohub-r…
#AIforScience #VirtualCell #SystemsBiology #Biohub #Measurement #ESM
ALT Alex Rives insight card: virtual-cell bottleneck is measurement, not bigger transformers
🧬 Roche just plugged Chai into multiple discovery programs.
Chai Discovery (CAD suite for protein therapeutics) announced a collaboration: Roche will use Chai’s AI models and platform across a number of drug-discovery programs and therapeutic areas.
Hard facts from the release:
→ Announced 2026-11-05 (Business Wire / Chai site)
→ Stack: Chai AI models + platform; Chai-3 cited as the latest model
→ Claim language: high target success rates and binding affinities, including challenging targets
→ Quotes on both sides (Meier, Chai CEO; Zaïtra, Roche Corp BD)
→ Latest in Chai’s 2026 big-pharma collaboration wave
What is NOT in the PR: deal size, milestones, named assets, or a peer-reviewed Chai-3 scoreboard.
Signal: generative protein design leaving the demo stage — and entering buyer pipelines.
📄 Chai: chaidiscovery.com/news/roche…
📰 Business Wire: businesswire.com/news/home/2…
#AIforScience #ProteinDesign #DrugDiscovery #Biotech #ChaiDiscovery #Roche
ALT Chai Discovery × Roche collaboration card: AI models and platform into multiple discovery programs
📎 Describe the dataset. Agents curate it from the literature — with a source for every cell.
James Zou (Stanford / CZ Biohub) + GXL just shipped Paperclip Datasets: you name the table you want; agents search and read papers, patents, trials, and regulatory docs in parallel, then assemble a refreshable dataset where every value links back to the passage that supports it.
Hard example — human drug half-life (now on HF):
→ 19,952 terminal half-life measurements
→ 2,418 normalized drug names · 3,270 PMC papers
→ ~30× the Obach / TDC benchmark (667 compounds, one value each)
→ keeps cohort, dose, route, and paper — not one number per drug
Bonus check (GXL eval): same agent, same prompt — Paperclip matched direction + significance on 90.3% of outcomes across 20 published meta-analyses vs 72.7% with web search.
🔗 Datasets: paperclip.gxl.ai/datasets
📄 Blog: gxl.ai/blog/introducing-pape…
🤗 HF: huggingface.co/datasets/gxl-…
🧵 James Zou: xitter.gq/james_y_zou/status/211…
#AIforScience #DrugDiscovery #Pharmacokinetics #MetaAnalysis #Datasets #Agents
ALT Paperclip Datasets metrics card: human drug half-life dataset curated by agents from literature
Very excited to introduce Paperclip Datasets!
Curate your own datasets from millions of papers in minutes. Just describe the dataset you want to build!
As example, we release the largest dataset of 20k half-life human measurements curated from 2400+ drugs💊
🖇️ Try it: paperclip.gxl.ai/datasets
📝 Blog: gxl.ai/blog/introducing-pape…
📊 Half-life dataset: huggingface.co/datasets/gxl-…
🧬 Don’t just pick the next mutant — evolve the workflow itself.
EnzymeRSI (Genoria AI; arXiv:2611.06321) is a dual-loop recursive self-improvement system for enzyme design: outer loop evolves harness assets without touching model weights; inner loop develops predictors when the method is the bottleneck.
Hard facts:
→ Outer loop: mean ProteinGym Spearman +0.044 across 5 assays (dev set); CD19 0.314 → 0.335
→ Inner loop: ThermoMPNN PCC 0.849; TadA Spearman 0.307 in 1.13 h
→ RLuc wet-lab: 10 of 32 non-pre-specified candidates ≥2× WT activity
→ A3A: Y130Q → 90.2% 5mC conversion; preference reversed in both contexts
→ KOD-E: path selection raised catalytic efficiency to 31.1× parent on modified dATP
Honest caveat: design rank vs measured rank on RLuc was only ρ = 0.175 — pool hit, ranking miss. Preprint; multi-round closed dry→wet campaign not yet shown.
📄 Paper: arxiv.org/abs/2611.06321
#AIforScience #EnzymeEngineering #ProteinDesign #Agents #LabAutomation #DirectedEvolution
ALT EnzymeRSI dual-loop recursive self-improvement system diagram from paper Figure 1
🧬 First vault particle structure from bacterial life — mapped with Claude.
Anthropic’s biology team just shared cryo-EM of prokaryotic vaults: barrel-shaped nanocapsules long known in our cells, now built from phage/bacterial MVP homologs.
Hard facts:
→ Claude Science screened ~15 billion predicted proteins
→ Homologs across 29 bacterial + 4 archaeal phyla
→ 4 complete vaults confirmed by cryo-EM (2 phage + 2 bacterial)
→ Phage L05: 96 protein chains · 54.5 × 47.5 nm · ~3.2 Å · ~25% more lumen than human vaults
→ End-to-end under 4 months; cells → grid in ~90 min; first model in 34 h
→ Humans still made the key calls (and fixed Claude’s misses — including a mirror-image map)
Bonus for protein design: add 2 wall repeats → a larger closed shell (~50% more volume than the human vault).
Function still unknown. The mystery got deeper — and engineerable.
📄 Blog: anthropic.com/research/inves…
📑 Preprint: www-cdn.anthropic.com/76e55f…
#AIforScience #ProteinDesign #CryoEM #StructuralBiology #SyntheticBiology #Claude
ALT Cryo-EM structure of prokaryotic vault particle from Anthropic Claude Science research
ALT Cryo-EM density of prokaryotic vault particle
ALT Engineered larger vault shell with duplicated wall repeats
📦 RFD4-Proteina just open-sourced its full training-structure store.
Not classic RFdiffusion — this is the Baker Lab IPD × NVIDIA GenAIR release stack. Yesterday’s Hub commit (kierandidi) moved it from a partial PDB+AFDB cut to every source used by the release flow + AE recipes.
Hard facts:
→ huggingface.co/datasets/nvid…
→ 2.13 TB · 19 configs · 33,527 objects
→ AFDB monomers, PDB interfaces, PLINDER ligands, glycans/Re-Glyco (~1.57 TB alone), RNA/DNA, Ab/TCR, NCAA/PTM, TED dimers, OMol, cyclic peptides…
→ Packaging/indices: CC BY 4.0 (structures keep upstream terms)
→ Model weights already BSD-3; you don’t need this dump to run inference
Why it matters: rare for a frontier biomolecular design model to ship weights + code + the actual training shards/provenance. Reproducibility and fine-tune fuel, not just a demo.
Caveat: distill ≠ experimental; Hub mix ≠ every checkpoint’s exact history; preprint, not peer-reviewed.
📄 Dataset: huggingface.co/datasets/nvid…
🧠 Model: huggingface.co/nvidia/RFD4-P…
💻 Code: github.com/RosettaCommons/RF…
📑 Paper: rfd4-proteina.org/assets/pap…
#ProteinDesign #AIforScience #OpenSource #RFD4 #Proteina #NVIDIA
ALT RFD4-Proteina training-structure store open source
ALT RFD4-Proteina training-structure store open source
ALT RFD4-Proteina training-structure store open source
🧬 AI agents aren’t just designing molecules anymore.
Shanghai Jiao Tong University’s AutoDNA (Nature) co-designs AI + robots for multi-instrument DNA wet-lab workflows — a science-as-a-service blueprint, not an AI bolted onto existing automation.
Hard facts from the paper:
→ Orchestrates 20+ instrument types across >11,000 steps
→ Core loop: synthesis → transcription → amplification → sequencing
→ Humans mainly for reagent / labware top-ups
→ Concurrent protocols 3.2× faster
→ DNA storage: exact recovery of “Intelligence is the ability to adapt to change” (11,382 ops in Fig. 5)
→ Clinical-style demo: identified & validated KRAS G13D
Worth reading if your “AI lab” still stops at the notebook.
📄 Paper: nature.com/articles/s41586-0…
💻 Code: github.com/IPADS-X/AutoDNA
#AIforScience #LabAutomation #SyntheticBiology #DNAStorage #DrugDiscovery
ALT AutoDNA Nature paper Figure 1: AI-native multi-instrument DNA wet-lab workflow overview from Shanghai Jiao Tong University
🧬 Stop sampling binders blindly.
**ProteoCraft** (Wu, Choi et al.) reframes protein design as **proposal–verification tree search**: truncated diffusion gradients propose where to mutate; full ESMFold2 decides what’s good — coarse multi-site exploration → single-site refinement.
Hard numbers from the preprint:
→ PD-L1 miniproteins (SPR): **82.9%** binding hit rate (29/35)
→ Best fitted **KD = 0.879 nM** — no affinity maturation
→ IFNAR2 VHH (BLI): binding confirmed; lead ~**153 nM**
→ In silico: strict high-confidence yield **54.3%** vs **19.0%** compute-matched Best-of-N
Worth reading if your binder loop still generates-then-filters — verification belongs *inside* the search.
📄 Paper: arxiv.org/abs/2611.03963
💻 Code: github.com/smiles724/ProteoC…
#ProteinDesign #AIforScience #BinderDesign #OpenSource
ALT ProteoCraft overview schematic — proposal–verification tree search for binder design (paper Figure 2)
🧬 Big Pharma isn’t buying AI demos.
UCB just wired Cradle’s AI protein-design platform into selected antibody discovery projects — hit generation → lead optimization → preclinical readiness — with a continuous lab-in-the-loop.
From the 4 Nov 2026 UCB press release:
→ Strategic collab: UCB (Belgium) × Cradle (Amsterdam)
→ Cradle’s end-to-end AI protein design/engineering platform runs inside UCB programs
→ Wet-lab results feed program-specific models — compounding “lab-in-the-loop”
→ Stated aim: optimize across the full property set, not one assay at a time
No price tag. No named targets. No hit-rate claims. The signal is the wiring — not a slide-deck pilot.
Worth reading: ucb.com/newsroom/press-relea…
#AntibodyEngineering #ProteinDesign #AIforScience #DrugDiscovery
ALT UCB × Cradle — AI protein design wired into antibody discovery with lab-in-the-loop (press-release summary card)
✨ 50+ tools. One open harness. Far fewer wasted tokens.
**Chaperone** (Baker Lab / RosettaCommons) turns RFdiffusion3, ProteinMPNN, AlphaFold3, and 50+ other design tools into composable, reproducible blocks—built for agents and HPC, not brittle shell glue.
From the bioRxiv eval (nine design tasks):
→ **−86%** total tokens vs baseline agents
→ **−64%** GPU-hours
→ Mean pass rate **6.1% → 11.7%**
→ Two-week cortisol campaign → low-nM binders zero-shot (best FP **K_D = 1.32 nM**)
Worth reading for the tool surface, not another “AI scientist” chat wrapper. Preprint.
📄 Paper: biorxiv.org/content/10.64898…
💻 Code: github.com/RosettaCommons/ch…
🔗 Docs: rosettacommons.github.io/cha…
#ProteinDesign #Agents #OpenSource #BakerLab
ALT Chaperone preprint Figure 2 — wet-lab multipanel (cortisol binder, nanoparticle, linker); Butcher et al., bioRxiv 2026.10.19.759896
🔥 Hot take: AlphaFold deserved the Nobel — “press a button → get a medicine” is still mostly hype.
— Mike Nohaile, CEO of Prellis Biologics (ex-CSO, Generate Biomedicines)
What AI is good at today 👇
✅ Optimize around an existing biological seed
✅ Improve sequence / function of a known binder
What it’s not ready for 👇
❌ One-click de novo drugs from scratch
❌ New binding modes on novel epitopes without a seed
Why the sequence math is brutal:
Antibody CDRs (~60 aa) → 20^60 possibilities.
Widen the searchlight all you want — most of sequence space is still black.
And those “AI proteins in the clinic”?
Dig in: most are variants improved from a reference molecule — not true de novo.
His line that sticks:
“If AI can create drugs, where are they?”
His timeline for push-button medicines: a long, data-bottleneck problem — decades, not next year.
🎧 techsurgepodcast.com/podcast…
#AIforScience #DrugDiscovery #ProteinDesign #AIPharma
ALT Quote — Mike Nohaile, CEO, Prellis Biologics: “If AI can create drugs, where are they?” · TechSurge Podcast