By Kenneth Bonett
One of the biggest bottlenecks in medicine is designing molecules that can grab onto the right protein and change what it does. Proteins run almost everything in your body. They drive inflammation, let cancer grow, transmit pain. Design something that sticks to the right protein in the right spot, and you can potentially treat the disease.
The problem? The best computer methods we had could design maybe one candidate every few hours. Some took days.
A new AI system called LigandForge just changed the math entirely. Instead of hours per molecule, it generates over 1,000 per second. In 3.4 minutes, it produced 150,000 candidates across 5 disease-relevant targets. What used to take months now takes minutes.
How? Traditional methods sculpt a 3D molecular shape, then check if it fits the target. Think of it like carving a key and testing it in a lock, one at a time. LigandForge learned the physics of binding during training, so at runtime it goes straight from “here’s the lock” to “here are thousands of keys” in a single pass. No sculpting step.
The results across 150 protein targets were remarkable. On targets where other leading AI methods completely failed (proteins involved in cancer, immune suppression, inflammation), LigandForge found promising candidates. It even generated hits against KRAS, the most commonly mutated oncogene, which researchers have called “undruggable” for decades.
As a medical student, this genuinely excites me. But it’s important to be honest about what still stands between a computer prediction and an actual treatment.
None of these results have been tested in a lab yet. Every binding affinity is a prediction, and what looks perfect on screen can fall apart in a test tube. Beyond that, peptides are notoriously hard to turn into drugs. Your body breaks them down fast, they usually can’t be taken as pills, and they struggle to reach the right tissues. Manufacturing them at pharmaceutical scale is expensive and technically demanding. And even with a perfect candidate, you’re still years away from clinical trials and regulatory approval.
So why am I still excited?
Because when you can screen hundreds of thousands of candidates in minutes, you can go after targets that were previously too difficult or too expensive to attempt. Combine that with improving delivery platforms (like lipid nanoparticles, the same tech behind mRNA vaccines), chemical modifications that extend peptide stability, and AI tools that predict drug behavior before you touch an animal model, and the path from computational hit to real therapeutic is getting shorter every year.
The finish line is still far away, but the race just got a lot faster.
Koukuntla R, Arya K, Rincon M (2026). LigandForge: Generative AI for Accelerated Ligand Design Across the Druggable Proteome. bioRxiv. doi: 10.64898/2026.03.14.711748
