On November 15, 2024, Bioptic launched PLUMBER — Protein–Ligand Unseen Matching Benchmark for Evaluating Robustness — a new standard for testing how well AI models generalize in predicting protein-ligand binding. Created by Bioptic team member Vlad Vinogradov and released on the Polaris platform by Valence Labs, PLUMBER is the largest multi-target benchmark of its kind, with 1.8 million data points spanning small molecules and amino acid sequences.
Why We Built PLUMBER
Most structure-based drug discovery methods depend on 3D protein structures. But proteins are messy — some resist crystallization, others are disordered or unstable. That’s why Bioptic takes a sequence-first approach, training models to predict binding using only the protein’s amino acid sequence.
Existing benchmarks weren’t cutting it — so we built our own. PLUMBER combines aggressive filtering, high-quality train/test splits, and public accessibility under CC-BY-4.0. It’s not just a dataset — it’s a stress test for AI.
How We Use It
PLUMBER is already powering Bioptic’s next-gen model development, including our BiPELL protein-agnostic language model. It helps us test robustness across targets, validate generalization, and improve predictions — faster.
We’ve also made all the pre-processing code available on Polaris. We believe in open tools for shared progress.
Why It Matters
PLUMBER unlocks new territory for AI in drug discovery. By moving beyond reliance on protein structures, it opens the door to targets previously out of reach — and gets us closer to scalable, fast, AI-native pipelines for finding drugs that work.
This benchmark is one more step toward Bioptic’s mission: building better software so scientists can build better medicine.
Explore PLUMBER on Polaris
Ready to test your model? Dive into PLUMBER on Polaris and join a growing community pushing the boundaries of molecular AI.
And congrats to Vlad and the team for setting a new bar.
