In March 2026, Andre Watson at Ligandal dropped LigandForge on bioRxiv — a discrete diffusion model that designs high-affinity peptide binders directly from receptor pocket geometry in a single forward pass. No docking. No iterative refinement. No waiting weeks for physics simulations. Over 700 sequences per second on a single GPU, peaking above 1,000. That’s 10,000 to one million times faster than BindCraft or BoltzGen.
The old drug discovery pipeline just became obsolete overnight. This isn’t incremental. It’s the moment artificial intelligence learned to speak the native language of proteins — and write the code that controls them.
Peptides Are Biology’s Executable Code
Proteins run the cell. Peptides are the precise instructions that turn them on, off, or rewrite their behavior entirely. They offer a level of specificity small molecules can only dream of — and a directness antibodies lack in many intracellular contexts.
Until now, designing them was painfully slow. LigandForge changes that by treating sequence generation as a learned diffusion process conditioned purely on pocket structure. The model has internalized the thermodynamics. Inference becomes pure, blazing-fast creativity.
The Million-Fold Compression of Discovery Time
Benchmarked across 150 protein targets, LigandForge produced hundreds of thousands of candidates. DeltaForge, their thermodynamic scorer, hit a Pearson r of 0.83 against experimental binding affinity data. On tough targets — TNF-α, PD-L1, VEGF-A, HER2 — it delivered predicted sub-100 nM binders where competitors struggled or failed.
It doesn’t just copy known motifs. DSSP analysis of over 8,500 folded peptides shows rich structural diversity: helices, sheets, coils, multi-domain folds. Real exploration, not regurgitation.
One GPU. 150,000 candidates in 3.4 minutes. This is what scalable biological programming looks like.
From Undruggable Targets to Precision Cellular Control
Many of the most important disease drivers have been considered “undruggable” precisely because we couldn’t reliably design binders fast enough or diversely enough. LigandForge opens the floodgates.
Peptide-based therapeutics, targeted delivery vehicles, even synthetic gene circuits — all now accelerate from concept to candidate at unprecedented speed. When you can program protein interfaces this fluidly, the proteome becomes a programmable interface.
This is synthetic biology meeting computation at the deepest level.
The Biocomputer Stack Gains Its Software Layer
Biology was always computational. We just lacked the tools to program it at will. LigandForge isn’t merely an AI tool for pharma. It’s a foundational compiler for the emerging biocomputer — where wetware runs the code we design in silicon.
Every advance like this brings us closer to fully realizing biology as programmable computation. The interface between digital design and molecular execution just got orders of magnitude thinner.
What happens when designing a molecular controller for any protein target takes seconds instead of years? The boundary between engineer and cell dissolves.
The biocomputer isn’t a future device. It’s the living world itself — and we’re finally learning to code it.
References
- Watson, A. (2026). Single-Pass Discrete Diffusion Predicts High-Affinity Peptide Binders at >1,000 Sequences per Second across 150 Receptor Targets. bioRxiv. https://www.biorxiv.org/content/10.64898/2026.03.14.711748v1
Related: What Is a Biocomputer in 2026? · Programmable Biology: When Cells Become Living Software
Feature image: AI-generated using Grok.