Multitask Bayesian Neural Networks for Multiparameter Protein Engineering
Zenodo Dataset Description This repository contains a curated benchmark collection of 27 multiparameter protein mutational datasets for evaluating machine learning models in protein engineering, with a particular focus on multi-task learning. The benchmark includes both single-point and mixed mutagenesis experiments across diverse protein systems. Dataset sizes range from 41 to 12,294 protein variants, providing challenging evaluation scenarios spanning both low-data and high-data regimes. The datasets contain between 2 and 16 experimentally measured functional properties, enabling the assessment of multi-objective prediction methods across varying levels of task complexity. Data Curation and Processing The datasets were collected from peer-reviewed publications and publicly available databases. During preprocessing: Non-informative and redundant columns were removed. Entries lacking complete sequence information or one or more experimental labels were excluded. Experimental measurements were normalized relative to their corresponding wild-type references using assay-specific transformations (e.g., relative differences and logarithmic scaling) to reduce skewness and stabilize variance. Full-length mutant protein sequences were reconstructed from mutation annotations using the corresponding wild-type sequence. Standardized mutant sequences and unique mutation identifiers were generated for every variant. All preprocessing operations were programmatically logged to ensure complete traceability and reproducibility. Repository Contents This repository includes: Curated benchmark datasets for 27 protein systems. Standardized mutant protein sequences. Normalized experimental measurements. Processing logs summarizing all preprocessing operations, removed entries, label transformations, and sequence reconstruction steps. A machine-readable JSON metadata catalog describing every dataset, including: UniProt or PDB identifiers Original data source Associated publication Experimental property descriptions Additional dataset metadata supporting FAIR data principles Intended Use This benchmark is designed for developing and evaluating machine learning methods for protein engineering, including: Protein fitness prediction Multi-objective optimization Multi-task learning Benchmarking of predictive models on diverse experimental protein mutational datasets The standardized preprocessing pipeline and comprehensive metadata facilitate reproducible benchmarking and straightforward integration into new computational workflows.