Lab Newsletter — August 3, 2026: The Third Molecule

Over the last two weeks we’ve watched generative design take on two of the cell’s molecules — the proteins it builds from and the RNA that carries its instructions. The third is the oldest target of all and the hardest: the small molecule, the little drug that has to slot into a protein pocket and do something. It’s hard because chemical space is effectively infinite and the only judge that counts is a wet-lab assay. This year gave us both halves of an honest picture — a model whose designed molecules actually worked at the bench, and a benchmark blunt enough to say how rarely that still happens. For a lab that lives by held-out tests and the design-build-test loop, the pairing is the whole lesson.

💊 The good news: it works, at the bench

The capability milestone is DiffGui (Hu et al., Nature Communications 16:7928, 2025 — peer-reviewed). It’s a target-conditioned E(3)-equivariant diffusion model that grows a ligand inside a protein pocket, with two ideas that matter: it diffuses atoms and bonds together (so bond types follow the evolving geometry instead of being bolted on afterward, cutting the distorted rings that plagued earlier models), and it uses property guidance to steer generation toward binding affinity and drug-likeness at once. On standard benchmarks it beats a row of prior methods on structural realism and affinity — but the headline is what happened off the leaderboard. For RSK4, a kinase not in its training set, DiffGui’s two synthesized molecules were active in an HTRF assay at IC₅₀ ≈ 215 nM and ≈ 111 nM; in a lead-optimization test on DHODH it sharpened one series from 32 nM to 10 nM. The authors report “both Compound 1 and Compound 2 demonstrate potent inhibitory activity in the HTRF assay.” Why it matters for the lab: this is design-build-test in chemistry — a generative model whose output was synthesized and measured, not just scored. That closed loop, kept open (the code is on GitHub), is the shape of tooling we care about.

📏 The honest ruler

Now the counterweight, and it’s a good one. Most molecule-generation papers report affinity and validity on frozen test sets — numbers that flatter. MolGenBench (bioRxiv, Nov 2025; preprint) refuses to. It assembles a genuinely industrial yardstick — 120 protein targets, 5,433 chemical series, 220,005 experimentally confirmed active molecules — and runs 17 generative models through the multi-stage workflow real discovery actually uses. The findings are sobering and specific. Stratify targets by whether the model saw them in training and the floor drops out: top de-novo hit rate falls from 0.124% on familiar proteins to 0.024% on unseen ones, and benchmarks that don’t stratify “overestimate real-world utility.” Most models show weak target awareness — generating “structurally similar molecules regardless of which target they were conditioned on.” Pose quality is thin: typically below 23% of generated poses land within 2 Å of a redocked conformation. And you can’t brute-force it — sampling 100,000 molecules per target raised the raw number of hits “but with diminishing efficiency.” Why it matters for the lab: this is our prove-it discipline transplanted into chemistry, and its central rule is one we keep repeating — stratify by novelty or you’ll fool yourself. A model that shines on proteins it has seen and collapses on the ones you actually need is a benchmark artifact, not a drug engine. The BioImage Model Zoo ethos — share the model and the honest test — is exactly what a field at this stage needs.

🧭 The way through

The gap is real, but 2026 is also showing how it closes. On the protein-binder side, BoltzGen (Nov 2025) took the novel-target wall head-on and reported wet-lab nanomolar binders for nine new targets sharing under 30% sequence identity with anything in the PDB — proof that generalization past the training distribution is achievable, not just aspirational. And a 2026 review in JCIM names the levers: bake in physical priors instead of hoping the model learns physics, fuse structure, omics, and phenotype rather than scoring on pocket-fit alone, wrap the generator in autonomous agents that adapt across design-make-test rounds, and optimize for many objectives with uncertainty in view. Read that list and it’s a map of the lab: agents that close the loop (Agent-Lens, the BioImage.IO chatbot), and the multimodal, phenotype-aware scoring that a virtual cell is built to provide — judging a molecule not by how it fits a pocket but by what it does to a cell.

The arc across three molecules is the same each time. Generation gets impressive fast; the bench stays honest; and the models that matter are the ones willing to be measured where it’s hard — on the target they’ve never seen, in the assay that doesn’t care how good the loss curve looked. Proteins, RNA, and now small molecules are all learning the same lesson we build around: a design is a hypothesis, and the experiment gets the last word.

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.) Have lab news to share — a talk, paper, conference or release? Message me on Slack.