The Work
Thiel reframed genomic language modelling as a design task: describe a plasmid in plain terms, get back a complete DNA sequence.
What Worked
On 1,000 held-out prompts, PlasmidLM produced a computationally useful plasmid in 48.5% of single samples and 89.7% of best-of-four.
What Has Been Achieved
PlasmidLM, a 19.3M-parameter model post-trained with GRPO, was the first DNA language model to generate complete plasmids from human-readable prompts, and a companion project showed reinforcement learning could shift generated sequences toward natural distributions, including traits it was not directly rewarded for. Validation stayed in silico on held-out benchmarks, with three sequences synthesised to show manufacturability.
What Was Learned
AI could not predict whether a design would work in living cells, because biology lacks the cheap, reliable verification that makes other AI loops fast.
About
McClain holds a BS in Data Science from Berkeley and MS in CS from Georgia Tech. He was a senior ML Engineer at SnorkelAI, transforming manual AI development processes into programmatic solutions. He previously worked as an ML Scientist at Tempus AI, where he conceived, developed and deployed several large-scale user-facing machine learning projects bringing the latest and greatest in LLMs to clinical and genomic applications.
Lab & Advisors
Chris Barnes
University College London
Professor of Systems and Synthetic Biology
Chris Barnes is a Professor in the Division of Biosciences at UCL. He has spanned multiple fields including particle physics, genetics, statistics, computational systems biology and synthetic biology. In addition to research, he is also actively engaged in training the next generation of researchers; he developed the BBSRC-funded e-learning resource SysMIC to train life scientists in computational and mathematical skills.
