Andre Watson has been trying to design peptides for thirteen years. He founded Ligandal in 2014 to work on precision genetic medicines, ran it quietly through the CRISPR boom, and watched as the field that was supposed to unlock peptide drugs — structure-based design — kept running into the same wall. TNF-α, PD-L1, HER2. Targets that matter enormously. Peptides that just wouldn't bind. Iterative refinement that ate GPU hours and produced nothing useful.
Now he has a preprint that says the wall is gone.
LigandForge, posted to bioRxiv on March 14, 2026 (doi: 10.64898/2026.03.14.711748), is a discrete diffusion model that takes a 3D receptor pocket geometry — a 48-dimensional feature vector per pocket residue describing physicochemical class, charge, solvent exposure, secondary structure, local geometry — and outputs a peptide sequence in a single forward pass. No structure prediction. No inverse folding. No gradient descent through a folding network at inference. Just geometry in, sequence out. Watson's team at Ligandal, a 12-year-old biotech with no prior public preprint record, is calling it a paradigm shift. They're also calling it a commercial product: the LigandAI platform, live at ligandai.com, with patents pending.
The speed numbers are the thing you'll hear first. LigandForge generated 150,000 candidates in 3.4 minutes on a single B200 GPU — 732 sequences per second average, peak 1,190. On a consumer NVIDIA RTX A2000 with 12GB of memory: 84 sequences per second, the same 150,000 candidates in just under 30 minutes. Against BindCraft, which requires approximately 2.5 hours per hallucination trajectory, the preprint claims a 1,000,000-fold throughput advantage. Against BoltzGen, a more structurally honest comparison: 10,000-fold.
But read the fine print on that million-times-faster number. BindCraft runs iterative AF2 hallucination plus five MPNN redesigns per candidate — it is solving a different and harder problem. The preprint acknowledges this. The 1M× figure is real arithmetic applied to incomparable things. What LigandForge is actually faster than is other single-pass methods, and the 10,000× figure against BoltzGen is the defensible version of the claim.
The number that matters is the head-to-head benchmark on five historically hard targets: TNF-α, PD-L1, VEGF-A, IL-7Rα, and HER2. LigandForge got sub-100nM predicted binders on all five — 23 hits from 576 folded structures. BoltzGen got 1/5 (2 hits from 100 designs). BindCraft got 0/5: zero pipeline-accepted designs. Against TNF-α specifically, where Google DeepMind's AlphaProteo failed entirely, LigandForge's production-scale run from 30,000 candidates produced three sub-100nM predicted binders including one sub-10nM, with a best predicted Kd of 1.7nM. Against PD-L1, where BindCraft reported 0% peptide success, LigandForge generated 62 good-tier binders — 39.2% of 158 folded structures.
The structural diversity story is worth sitting with. LigandForge produced 69% helical peptides, 9% beta-sheet, 8% multi-domain, 10% coil, 4% mixed. BoltzGen: 77% helical. BindCraft: 93% helical. If you're trying to hit a pocket that doesn't want a helix — and plenty of them don't — the competitor methods are not even in the game.
Here is the part that should make you read this paper twice before drawing conclusions: all of those binding affinities are predicted. Not measured. DeltaForge, Ligandal's thermodynamic scoring engine, achieves a Pearson r of 0.83 on the PPB-Affinity curated peptide benchmark — but that benchmark is 77 high-quality peptide complexes. On the full heterogeneous PPB-Affinity dataset of 4,347 structures, the correlation drops to r=0.36–0.41. The preprint includes both numbers. DeltaForge's thermodynamic floor is approximately -15 kcal/mol, which means any sub-nanomolar Kd claim is an upper bound, not a precise prediction. The paper is transparent about this. The reader's job is to take them at their word and adjust accordingly.
There is also the BindCraft paradox. Watson's team took the 36 MPNN redesigns that BindCraft's own pipeline rejected, re-folded them with Boltz-2, and found that 31 of 36 — 86% — achieved elite iPSAE scores of 0.8 or higher. These were not low-quality designs. They were functional binders that BindCraft's clash detection discarded, apparently because the pipeline was calibrated for full-length proteins, not short peptides. BindCraft isn't failing on these targets because the physics is wrong. It's failing because its quality control is measuring the wrong thing.
Watson is listed as founder, chairman, and CEO of Ligandal on the company org chart. The paper discloses his commercial interest explicitly. Patent applications have been filed covering the discrete diffusion architecture, the DeltaForge scoring engine, and related methods. Ligandal's platform is live. This is both a scientific preprint and a product announcement, and the preprint knows it.
For VCs and founders evaluating what this means for the field, a few frames:
If the predicted affinities hold in wet validation — a large if — the benchmark is significant precisely because TNF-α, PD-L1, HER2 are not cherry-picked failures. They are canonical hard targets. A method that produces sub-100nM predicted binders on all five of them, where established approaches produce zero, changes what "druggable" means for peptide therapeutics. That is not a marginal improvement. That is a category claim.
If the affinities don't hold — if DeltaForge's r=0.36 on heterogeneous data is the honest number — then the absolute Kd values are unreliable and the comparative benchmark tells you something useful but not definitive. The structural diversity claim survives that uncertainty. The BindCraft paradox is a real insight regardless of whether any individual LigandForge hit is a real binder.
The preprint is also, quietly, a verdict on the iterative hallucination paradigm. BindCraft has been one of the more respected methods in the space. Its failure on these five targets, and the discovery that its own pipeline was discarding designs that Boltz-2 called excellent, suggests that running AF2 repeatedly through a redesign loop is not automatically superior to a single-pass statistical model. The field's assumption that "more structure awareness = better" deserves to be tested, not presumed.
What LigandForge cannot tell you yet is whether any of its 23 hits on those five targets will survive contact with a real binding assay. The gap between a predicted Kd of 1.7nM and a measured Kd of 1.7nM is the entire remaining work of drug discovery. Watson knows this. He's been doing this for thirteen years. The preprint is not the end of the story. It is the point where the story becomes expensive.
What Ligandal has that other computational peptide companies have not is a concrete, falsifiable benchmark. Twenty-three predicted hits across five hard targets is a number that can be checked. If Watson's team synthesizes those peptides and they bind, this paper will look like a inflection point. If they don't, the preprint will look like an impressive piece of engineering built on a model that generalized poorly to physical chemistry. Either way, the experiment is runnable. That's more than most preprints in this space can claim.
