Neil Thomas

I work at the intersection of AI and biology. I am currently at Biohub (fka EvolutionaryScale) building AI tools for biological design.

Previously, I was a Research Scientist at Google X. I completed my PhD in Computer Science at UC Berkeley in 2022, advised by Professor Yun S. Song. Prior to that, I was a Software Engineer at 23andMe. I received my BS in Engineering Mathematics and Statistics from UC Berkeley.

When I'm not being humbled by biology, I like to be humbled by a variety of hobbies. I like cooking recipes from Alison Roman, climbing rocks, skiing, playing ultimate frisbee, cycling, playing piano, watching comedy, and watering my plants.

email  /  twitter  /  bsky  /  github  /  scholar  /  linkedin

Research Highlights

My research focuses on learning meaningful representations of proteins, with the aim of enabling applications in protein design, functional annotation, and structure prediction. Check out my thesis talk "Browsing in the Library of Babel" for an accessible introduction.

Language Modeling Materializes a World Model of Protein Biology

Introduces ESM Cambrian (ESMC), a 6B parameter protein language model trained on over 2B sequences. Introduces ESMFold2, a state of the art single sequence folding model which uses an ESMC backbone. Validates inversion of ESMFold2 for the design of nanomminiprotein and scFv binders with nanomolar affinities. Introduces the ESM Atlas, a map of 6.8B sequences and 1.1B predicted structures connected through interpretable features derived from ESMC’s latent space.

Engineering highly active and diverse nuclease enzymes by combining machine learning and ultra-high-throughput screening

Introduces TeleProt, a framework for guiding protein library design with machine learning that leverages evolutionary and experimental data. Validated in an enzyme engineering campaign to optimize the endonuclease NucB. Across four rounds, ML-designed libraries show improved hit rate and diversity compared to directed evolution - the first head-to-head comparison of its kind.

Simulating 500 million years of evolution with a language model

A multi-modal generative protein model that reasons flexibly across protein sequence, structure, and function.

Tuned Fitness Landscapes for Benchmarking Model-Guided Protein Design

Tunable, realistic, synthetic fitness landscapes for benchmarking protein design.

Interpreting Potts and Transformer Protein Models Through the Lens of Simplified Attention

Introduces “factored attention,” a simplified attention layer that we use to compare and contrast Potts models and Transformers.

Evaluating Protein Transfer Learning with TAPE

A suite of benchmarking tasks for protein language models.

Research

For an up-to-date list, see Google Scholar

Whole-genome sequencing reveals a complex African population demographic history and signatures of local adaptation

End-to-end learning of multiple sequence alignments with differentiable Smith-Waterman

Functional genomics of OCTN2 variants informs protein-specific variant effect predictor for Carnitine Transporter Deficiency

Minding the gaps: The importance of navigating holes in protein fitness landscapes

Teaching

During my graduate studies at Berkeley I had the privilege of teaching:

• Summer 2022 CS 188: Introduction to Artificial Intelligence
• Fall 2020 Stat 135: Concepts of Statistics

My teaching statement.

Built on Leonid Keselman's Jekyll fork of Jon Barron's website