SAMPLE: a Self-Driving Laboratory System for Protein Engineering

Tuesday April 25th, 4-5 pm EST | Jacob Rapp — PhD Candidate, UW-Madison

Abstract: Current advances in machine learning enable more rapid understanding of the relationship between protein sequence and protein function than previously possible, but these computer models rely on access to either high-quality or highly numerous data. If an AI agent had access to a robot capable of executing experiments, the combined machine could intentionally perform specific experiments to efficiently build its understanding with the aforementioned high-quality data without any need for the ongoing involvement of a researcher. We have assembled such a self-driving laboratory system with the help of cloud laboratory company Strateos. The self driving lab, named SAMPLE (Self-driving Autonomous Machines for Protein Landscape Exploration) is capable of selecting sequences from a library of possibilities, assembling and expressing those selected sequences, and interpreting the experimental results to better inform future selections. The machine learning portion of SAMPLE chooses which sequence to test next based on predicted values for the function of interest, the uncertainty of that prediction, and how likely the protein is to be functional at all. Sequences that score highly for all three are more likely to be tested. Four independent iterations of SAMPLE were tasked with finding highly thermostable sequences within a 1352-member chimeric library made from 6 naturally occurring beta glucosidases. No information was provided to the agents other than the protein sequence of each library member and the melting points of each of the 6 parental sequences. Each agent was permitted 60 experiments, during which time all four agents independently observed multiple sequences with a melting temperature 10 degrees greater than the best natural sequence in the library. SAMPLE is robust to experimental errors, with all four agents reaching their goals despite two agents experiencing failed experiments on the path to their highly stable observations.

Preprint: https://www.biorxiv.org/content/10.1101/2023.05.20.541582v1

Recording Link: https://youtu.be/CylJ1p8jxXs