Experimental Validation of Protein Large Language Models Using Generated PETase Variants

Overview

Protein design has allowed researchers to harness the properties of natural protein families through the creation of improved proteins that tackle specific biomedical and environmental issues. Large language models (LLMs) trained on natural protein sequences have allowed researchers to design de novo sequences within unexplored regions of protein space with potentially enhanced properties. However, the proteins these LLMs produce lack experimental validation and therefore require synthesis, translation, and assay to determine proper functionality. With this information, conclusions can be drawn about how useful these models are for protein design applications. Here, we use ZymCTRL, an LLM 66 trained on the BRENDA database of enzymes. This LLM can be focused on a specific family of enzymes using an EC control tag and then further fine tuned using specific training data. We chose the PETase family of proteins to test the quality of the sequences produced by ZymCTRL within a specific protein family. The structure of PETase is well documented and its expected functionality is easy to detect. Furthermore, because of its ability to degrade plastics, the employment and improvement of PETase pose a potential solution to the worldwide ecological challenge of plastic waste, giving a prime example of the value of protein design. We finetuned ZymCTRL using data from the PETase family of proteins. We then experimentally assayed the PETase variants produced by ZymCTRL to determine the LLM's proficiency in producing functional proteins.

Competition history

• AJAS 2025 Category not listed

Related projects

Closest projects by meaning, across every fair and year in the corpus.

Browse more like this

• AJAS None projects
• Every AJAS 2025 project