Methods for Machine-Learning Guided Protein Design
This technology is related to case #26345.
The software algorithm and pending US patent application provide an iterative, machine-learning method for identifying protein variants having optimized performances for any type of binding activities between the two molecules (e.g. nucleic acids; sugars; proteins; small molecules). The software uses a protein large language model (LLM) to generate embeddings for candidate protein sequences and a predictive model to rank the variants based on experimentally measured activity. By directing experimental testing toward promising variants, the approach enables a faster, more efficient and comprehensive exploration of protein sequence space than conventional protein-engineering methods.
Researchers
-
using protein large language models to improve protein activity
United States of America | Pending
Technology
An initial protein sequence is modified, for example through amino acid substitutions, to generate a collection of candidate variants. Each sequence is processed by the LLM to produce a corresponding embedding in a latent space. The activity of these variants is measured experimentally, and the corresponding embeddings and activity measurements are used to train a predictive top-layer model. The model then estimates the activity of the remaining protein variants. The highest-ranking variants are selected for further experimental testing and used as the starting sequences for the next design cycle. This iterative process continues until a predetermined termination criterion is reached, typically 3-4 rounds of 5-8 candidates per round.
Problem Addressed
Evaluating the activity of genome-editing proteins such as Cas9 and Cas12 traditionally requires multiple labor-intensive steps, including cell culture, transfection, incubation, targeted DNA cleavage, and downstream analysis. Because only a limited number of variants can be tested simultaneously, these workflows make large-scale protein optimization time-consuming and resource intensive. Directed evolution can explore mutations without prior assumptions, but samples only a small fraction of the vast protein sequence space, while rational design is constrained by current knowledge of protein structure and function. This technology combines experimental measurements with machine-learning predictions to prioritize promising variants, expanding sequence-space exploration while reducing the experimental burden required for each design cycle. Proof of concept engineered protein variants for activity optimization include CRISPR Cas9 –RT prime editor, serine recombinase Bxb1 integrase, T7 RNA polymerase, CRISPR Cas12 enzyme, and monoclonal antibodies against C143 and CD71.
Advantages
Requires only protein sequences as model inputs, without necessarily relying on protein structures, extensive prior data, or expert knowledge
Uses an iterative active learning approach to reduce the number of variants requiring experimental testing
Supports model-guided exploration of large protein sequence spaces
Can optimize multiple protein activities and properties simultaneously
Publications
Jiang, Kaiyi, Zhaoqing Yan, Matteo Di Bernardo, Samantha R. Sgrizzi, Lukas Villiger, Alisan Kayabolen, Byungji Kim, et al. “Rapid In Silico Directed Evolution by a Protein Language Model with EVOLVEpro.” Science 387, no. 6732 (2025): eadr6006. https://doi.org/10.1126/science.adr6006
License this technology
Interested in this technology? Connect with our experienced licensing team to initiate the process.
Sign up for technology updates
Sign up now to receive the latest updates on cutting-edge technologies and innovations.