Brian Naughton's new guide for Asimov Press is a practical walkthrough of computational antibody design — the kind of honest, unglamorous technical writing the field rarely produces. It doesn't oversell what the AI tools can do. It also captures, in real benchmark numbers, where they fall short.
The guide centers on VHH nanobodies: single heavy-chain antibody fragments naturally produced by camelids. They are roughly one-tenth the size of a conventional antibody, expressible in bacteria, and structurally simpler in ways that suit AI-driven generative design. The target Naughton selects is Nipah virus Glycoprotein G — a pandemic-relevant protein with available structural data, which matters because every tool in the pipeline depends on it.
The five-step workflow — target selection, structure preparation, running design campaigns, candidate filtering, and experimental validation — runs through Ariax, a cloud platform that handles compute infrastructure in the background. Until recently, this kind of work required specialist resources that most labs couldn't access.
The open-source ecosystem is anchored by BoltzGen, developed by Hannes Stärk and the Boltz team — the same researchers behind Boltz-2, MIT's structure prediction model that competes directly with AlphaFold 3. Released under the MIT license, BoltzGen has achieved sub-micromolar binding affinity across the majority of targets it has been tested on. That's the strong result. The more complicated one: on the Nipah G Adaptyv Bio competition dataset, BoltzGen's pass rate was 1%. Naughton treats this as a target-specific failure mode rather than a systemic one, and he's probably right that per-target variance is the more honest framing. But the gap between general performance claims and dataset-specific results is real, and it applies to every tool in this space.
BindCraft — which runs AlphaFold 2 in reverse, generating binders from a target structure — is the other open-source option with a meaningful track record. On the commercial side, Nabla Bio, Chai Discovery, Latent Labs, and Isomorphic Labs are all reporting high success rates, though those claims are harder to evaluate without transparent benchmarks.
Success in antibody design is measured by dissociation constant (Kd). Sub-micromolar affinity is the basic threshold; picomolar-to-nanomolar is what therapeutic utility actually requires. Those numbers cut through a lot of promotional noise.
The Ariax platform deserves more attention than it gets in the piece. It represents a design pattern becoming common across biotech: a managed workflow that turns a multi-step scientific process into something closer to a configurable pipeline. Target in, candidate list out, validation handled separately by humans. What Naughton has written is less a tutorial than a protocol — defined inputs, defined outputs, honest data at each stage. That pipeline structure, where agentic tooling handles the combinatorial search and human judgment comes in at the filtering and validation steps, is where the serious infrastructure investment in this space is heading. In a field where the gap between benchmark claims and experimental results tends to be wide, a guide that names the numbers is more useful than it sounds.