RomeroLab-Duke/protein-fitness-datasets-for-benchmarking-ft-esm2-strategies

Protein Fitness Datasets for Benchmarking ESM-2 Fine-Tuning Strategies

Dataset description

This repository contains processed protein sequence–function datasets for CreiLOV, avGFP, and Ube4b.

The variants and experimental measurements were obtained in previously published deep mutational scanning studies:

Chen, Y. et al. Deep Mutational Scanning of an Oxygen-Independent Fluorescent Protein CreiLOV for Comprehensive Profiling of Mutational and Epistatic Effects. ACS Synthetic Biology 12, 1461–1473 (2023).

Sarkisyan, K. S. et al. Local fitness landscape of the green fluorescent protein. Nature 533, 397–401 (2016).

Starita, L. M. et al. Activity-enhancing mutations in an E3 ubiquitin ligase identified by high-throughput mutagenesis. Proceedings of the National Academy of Sciences 110, E1263–E1272 (2013).

The datasets are organized by the number of amino acid substitutions relative to each reference sequence. They include individual datasets for variants containing one through five substitutions and combined datasets containing variants with zero through five substitutions.

These data were assembled to benchmark strategies for fine-tuning ESM-2 across different proteins, experimental assays, training-set sizes, and mutational distances from a reference parent sequence.

The test set includes 50 sequence variants from each tested mutational regime for a total of 250 sequence variants (5x50=250)

Dataset organization

For each protein, the repository contains a combined dataset covering variants with zero through five amino acid substitutions:

0thru5_<protein>_mutants.csv

It also contains datasets restricted to exact mutation counts:

subset_single_<protein>_mutants.csv
subset_double_<protein>_mutants.csv
subset_triple_<protein>_mutants.csv
subset_quadruple_<protein>_mutants.csv
subset_five_<protein>_mutants.csv

Here, <protein> is one of:

avgfp
CreiLOV
ube4b

Intended uses

These datasets can be used for:

• Fine-tuning protein language models for sequence-to-function prediction
• Benchmarking ESM-2 fine-tuning strategies
• Evaluating generalization across mutational distances
• Comparing model performance across distinct proteins and assays
• Developing supervised protein property predictors

Important considerations

The experimental targets are not directly comparable across proteins:

• CreiLOV and avGFP were evaluated using fluorescence-based measurements.
• Ube4b variants were evaluated using a selection for ubiquitination activity.
• Each source study used a different library design, assay, and measurement scale.

Models should therefore be trained and evaluated using the appropriate target variable and normalization for each protein dataset.