Limitations of Current Machine-Learning Models in Predicting Enzymatic Functions for Uncharacterized Proteins.

Area of Science:

• Genomics
• Proteomics
• Bioinformatics

Background:

• A significant portion of proteins across genomes, termed the 'protein unknome', lack assigned functions, hindering biological research.
• Machine learning (ML) methods offer potential for inferring protein functions but face challenges in predicting novel enzymatic activities.

Purpose of the Study:

• To evaluate the efficacy of ML in predicting enzymatic functions for uncharacterized proteins.
• To identify limitations in current ML approaches for functional annotation.

Main Methods:

• Utilized the DeepECTransformer platform to analyze over 450 enzymes of unknown function from *Escherichia coli*.
• Dissected ML prediction failures to understand shortcomings in functional knowledge propagation.

Main Results:

• ML models demonstrated success in propagating known functional information but largely failed to predict novel enzymatic functions.
• Analysis revealed specific failure modes of ML in predicting unseen enzymatic activities.

Conclusions:

• Current ML methods require refinement to accurately predict novel enzymatic functions for uncharacterized proteins.
• Addressing these limitations will enhance the utility of genomic data and aid in annotating the 'unknome'.