Lab Newsletter — August 12, 2026: The Handshake

AlphaFold answered what shape does this sequence fold into? — one protein, one snapshot. But almost nothing in a cell works alone. A signal passes because one protein grips another; a drug works because a small molecule settles into a pocket at just the right depth; a gene switches because a transcription factor clamps onto a stretch of DNA. Biology is a chain of handshakes, and for years predicting a handshake — the joint structure of two or more molecules locked together — was a separate, harder problem than folding a single chain. In 2024–25 that wall came down twice: first the structure of the complex, then the strength of the grip.

🤝 One model for the whole complex

The first breakthrough was co-folding — predicting the entire assembly at once instead of folding parts and docking them afterward. AlphaFold3 (Abramson, Jumper et al., Nature, 2024) rebuilt AlphaFold around a diffusion architecture that generates atomic coordinates directly and leans far less on the evolutionary alignments its predecessor depended on. The payoff is reach: one unified model that predicts “the joint structure of complexes including proteins, nucleic acids, small molecules, ions and modified residues” — and does it with “far greater accuracy for protein–ligand interactions compared with state-of-the-art docking tools,” much higher accuracy for protein–nucleic acid complexes, and better antibody–antigen prediction than AlphaFold-Multimer. One framework, spanning most of the molecular vocabulary of a cell. The catch was political, not technical: AlphaFold3 arrived not fully open-source and not licensed for commercial use — a server, not a model you could hold — which is precisely what lit a fire under the open-source community. Why it matters for the lab: the complex, not the monomer, is where biology’s decisions get made — and an open version is what lets a lab build on it rather than query it.

⚡ Boltz: open, then the leap to affinity

That open version came from down the road at MIT. Boltz-1 (Wohlwend, Corso, Passaro, Barzilay & Jaakkola, MIT Jameel Clinic; bioRxiv, Nov 2024) was the first fully open-source, commercially usable structure model to reach AlphaFold3-reported accuracy — training code, inference code, weights, and benchmarks all released under the MIT license — predicting protein, RNA, DNA and small-molecule structures in 30 to 60 seconds per complex. Named for the Boltzmann distribution, it matched Chai-1 (the first closed-but-public AF3 replication) “and therefore AlphaFold3.” Then came the step no structure predictor had taken. Boltz-2 (MIT CSAIL + Jameel Clinic with Recursion; bioRxiv, June 2025) is the first co-folding model to jointly predict structure and binding affinity — not just where a drug sits, but how tightly it holds. Its headline: the first deep-learning model to approach the accuracy of physics-based free-energy perturbation (FEP) — the gold-standard, wildly expensive way to compute binding — while running about 1,000× faster. On the held-out FEP+ (OpenFE) benchmark it reaches a Pearson of ≈0.62, comparable to the FEP pipeline itself; in the CASP16 affinity challenge, run out-of-the-box with no fine-tuning, it outbid every submitted method across 140 protein–ligand pairs. The point isn’t a leaderboard — it’s that accurate virtual screening becomes practical: score millions of candidates in silico instead of synthesizing them. Why it matters for the lab: open weights and training code is the BioImage Model Zoo / BioEngine ethos — a model you can fine-tune to your own chemistry, now for molecular interactions.

🧭 The wiring of a cell — and the honest frontier

Here is why this belongs on a lab building toward a virtual cell. A cell is not a bag of isolated parts; it is a network of interactions — which protein binds which partner, which ligand touches which pocket, and how strongly. Co-folding predicts the edges of that network; affinity puts weights on them. If yesterday’s ensembles put motion under the individual parts, this puts connections and their strengths between them — the layer a Human Cell Simulator needs to reason about signalling and perturbation. It also closes a loop we opened last week: pair a fast affinity model with a generative molecule designer and an autonomous research agent can propose a candidate, score it, and send only the best to the bench — exactly the workflow the Boltz-2 team demonstrated, coupling their model with a generator to find synthesizable, high-affinity TYK2 binders. But the frontier stays honest, and that’s what keeps it useful. Boltz-2 approaches FEP — a Pearson of 0.62 is strong for ranking candidates, not a substitute for a measured binding constant; it still lags AlphaFold3 on antibody structures; and in that TYK2 demonstration every promising binder was checked by a full FEP simulation before anyone believed it. That is the same prove-it discipline we keep returning to: a predicted complex, like a generated stain or a virtual cell that shows its work, is a hypothesis until physics or the bench agrees. The fold told us what a protein is. The handshake is starting to tell us what it will do to the molecule across from it — fast enough, and openly enough, to screen at the scale biology actually needs.

Sources linked inline. Compiled by Happy Agent; the lab footer notes our AI-assisted content. (X/Twitter sweep was skipped today — our news API is out of credits; a Grok-based replacement is wired and awaiting credits.) Have lab news to share — a talk, paper, conference or release? Message me on Slack.