Protein Language Model-Based Fitness Estimates Facilitate Resistance Mutation Identification

Area of Science:

• Computational biology
• Drug discovery
• Genomics

Background:

• Drug resistance in cancer is a significant clinical challenge, leading to treatment failure and recurrence.
• Identifying resistance mutations early can guide treatment decisions and accelerate the development of new therapies.
• On-target mutations conferring drug resistance are key to understanding treatment failure.

Purpose of the Study:

• To evaluate the combination of physics-based free energy perturbation (FEP) and protein language model-based fitness estimates for improved in silico identification of cancer drug resistance mutations.
• To assess the utility of these computational methods in predicting mutations that confer resistance to targeted cancer therapies.

Main Methods:

• Utilized free energy perturbation (FEP) for affinity estimates and protein language models for protein fitness predictions.
• Validated computational predictions against deep mutational scanning (DMS) experimental data for resistance mutations.
• Tested the approach on public ERK2 inhibitor resistance data and internal EGFR exon 20 mutant data.

Main Results:

• Protein fitness estimates effectively filtered out mutations that were not experimentally validated as resistant.
• FEP flagged some mutations as potentially resistant due to decreased affinity, but these were correctly excluded by fitness estimates.
• The combined approach demonstrated improved accuracy in identifying true resistance mutations.

Conclusions:

• Protein language model-based fitness estimates offer a computationally efficient method for filtering potential resistance mutations.
• This approach can help avoid costly and error-prone modeling of protein function and stability.
• Accurate in silico prediction of drug resistance mutations can significantly impact clinical trial design and patient treatment.