Sal Candido

Sal Candido is a machine-learning researcher and former EvolutionaryScale co-founder and chief technology officer who works on models of proteins and biological systems. His work spans practical protein design and a deeper question: what can models trained on biological data teach scientists about how living systems work?[1][2]

At EvolutionaryScale, Candido helped build protein language models. In his account of ESM3, the model can be prompted to design proteins. The team generated a green fluorescent protein that was distant in sequence from known fluorescent proteins, then performed laboratory validation. This is a demonstrated example of protein design, not evidence that every generated protein will fold or work as intended.[3][2][4]

ESM3 represents protein sequence, structure and function in a shared modeling system. During training it masks parts of these representations and learns to predict them; a researcher can then provide partial specifications to guide generation. A sequence is the ordered list of amino acids, a structure is their three-dimensional arrangement, and function is what the molecule does. The three are related, but predicting one does not automatically establish the others.[4]

Candido describes training protein language models on metagenomic sequences—genetic material sampled from environments, often with incomplete or noisy protein sequences. In his account, that imperfect data can still improve models used for protein design and for understanding known proteins. The lesson has a limit: selecting data simply because it is easy to collect can pull research away from the biological question that matters.[1]

He also distinguishes architecture choices from scaling. A model with scientific assumptions built in can work better when data is scarce; an assumption that is wrong may constrain what it learns as data grows. For Candido, making a larger model useful still involves algorithmic choices about how it represents the problem and uses the information available.[1]

Candido compares studying a single protein with modeling one spoke of a bicycle: useful, but incomplete if the question concerns the whole system. Proteins interact in a biological context, and a prediction of a shape is distinct from a model of dynamics, function or a designed intervention. He argues for connecting molecular models to larger systems while continuing to improve the individual models.[1]

He is also interested in what scientists can extract from a model after training. Candido says his team has found information related to structure, function and motion in protein-model representations; using a model to produce a useful design and understanding why it works are different achievements.[1]

Candido previously worked on autonomous navigation for Project Loon and co-led AI for math and science at Meta’s FAIR. Biohub lists him as an engineering fellow and vice president. Asked when AI would produce a drug entirely on its own, he says the timing is hard to predict, while stressing that existing tools can already contribute to research.[2][1]