Gabriele Corso

Gabriele Corso is the co-founder and CEO of Boltz and a contributor to molecular docking research. His work includes DiffDock, which predicts the three-dimensional structure of a protein–ligand complex, and DockGen, which tests how docking models generalize across protein-binding domains. At Boltz, the published model family extends biomolecular structure prediction to jointly modeling structures and binding affinities.[2][5][8][4]

The Boltz repository provides its code and model weights under the MIT license for both academic and commercial use. Boltz-2 jointly predicts biomolecular complex structures and binding affinities, adding an affinity task beyond the structural predictions of Corso’s earlier DiffDock work.[4][5]

In their January 8, 2026 manifesto, Corso, Wohlwend and Passaro expressed a shared ambition: “We believe in a future where every scientist can iterate at the speed of inference.” This was a vision for scientific iteration, rather than a claim that computational predictions had replaced physical experiments.[7]

Corso and Wohlwend also appeared together on Latent Space to discuss the evolution of structural biology models from AlphaFold to Boltz-1 and Boltz-2.[1]

Corso pursued a Computer Science PhD at MIT’s CSAIL from 2021 to 2025. His earlier docking research supplies context for Boltz-2’s broader prediction task: predicting a binding structure and predicting binding affinity are distinct problems.[2][5][4]

DiffDock credits Corso alongside Hannes Stark, Bowen Jing, Regina Barzilay and Tommi Jaakkola. It predicts a complex’s three-dimensional structure and outputs a confidence score for that structural prediction; it does not predict ligand–protein binding affinity. The repository credits Jacob Silterra with leading code maintenance and improvement since 2024.[5]

In the 2024 paper Deep Confident Steps to New Pockets, Corso, Arthur Deng, Benjamin Fry, Nicholas Polizzi, Regina Barzilay and Tommi Jaakkola addressed docking generalization. They introduced DockGen, a benchmark based on proteins’ ligand-binding domains, and reported that existing machine-learning docking models generalized weakly under this evaluation.[8][6]

The team studied larger datasets and models, synthetic data strategies, and Confidence Bootstrapping, a training approach built around interaction between diffusion and confidence models. They reported improved docking to unseen protein classes, making generalization an explicit target alongside performance on established benchmarks.[6][8]