Matthew McPartlon

Matthew McPartlon is a co-founder of Chai Discovery, which builds AI models and software for molecular design. As of July 2026, he led research alongside fellow co-founders Joshua Meier, Jack Dent and Jacques Boitreaud, who led applied science. McPartlon’s public account of the company’s ambition goes beyond accelerating existing discovery methods: first match what biological screening can accomplish, then design molecules those methods cannot produce.[3][1]

In his August 2026 Latent Space appearance, McPartlon recalled that he needed months of conversations with Meier before believing the thesis behind Chai. He framed the challenge as “can I beat a mouse, and then can I do what mice can’t do?” The comparison expresses his ambition for computational design, rather than a claim that Chai had demonstrated every such capability.[1]

In an undated Training Data interview, McPartlon described simplicity as a guiding principle for scaling data, models and compute. He estimated that Chai-1 had 23 distinct submodules, qualifying the count with “I think.” Understanding each module’s behavior and interactions made iteration difficult; simplifying the model, he argued, made it easier to identify the research directions that mattered.[2]

McPartlon described weighing a target-specific extra module against an improvement that could generalize across a broader class of targets. His example concerned glycosylation sites: sequence motifs might allow a sufficiently capable model to infer where those sites belong without a separate explicit representation. He presented this as a modeling hypothesis and retained the requirement for rigorous validation.[2]

In the same interview, McPartlon endorsed Chai’s partnership model because better models could produce better customer results while opening new opportunities to generate training data. He said the company’s science team also used the models internally to understand their capabilities. His account described successive model generations: a newly usable capability enables new experiments and data, which can support the next model.[2]

A technical goal McPartlon highlighted was controllability. Rather than merely hoping a generated molecule would bind well, he wanted users to specify a property such as exactly a 10-nanomolar binder. He described that degree of control as work still to be solved.[2]

McPartlon’s earlier work in protein structure prediction supplies context for these design questions. In his Training Data account, he traced his entry into the field to a change of direction during his PhD.[2]

McPartlon recalled studying pure mathematics and beginning a PhD in theoretical computer science, then switching after his third year into deep learning and protein structure prediction. In his advisor’s lab, he said, the team worked on an early deep-learning system that predicted distances between amino acids; a separate renderer converted those noisy, incomplete predictions into a protein structure. He described the resulting structures as broadly correct but less detailed than later models.[2]

McPartlon characterized his early approach as focused on predicting three-dimensional structure, contrasting it with Meier’s work on language models for protein sequences. He explained the connection between them: predicting a missing amino acid requires understanding its local environment, which depends on the protein’s shape. Better structure predictors could then be applied to large sequence databases to create additional training structures.[2]

Meier recalled an early company discussion at McPartlon’s house, where McPartlon showed him antibody–antigen structure-prediction results. In Meier’s account, these results mattered because predicting an antibody’s structure was a prerequisite for designing one, and limited antibody data had made that prospect seem difficult.[2]

McPartlon also recalled the practical jump in compute: four people shared eight GPUs in his PhD group, whereas Chai began with 128. He described the subsequent engineering challenges as keeping clusters healthy, sustaining training runs for months, automatically resuming failures and using compute efficiently. His emphasis on a high-quality codebase followed from how he defined the team’s deliverable: software, even when its builders were AI researchers and protein designers.[2]