Evaluating Pretrained Protein Language Model Embeddings as Proxies for Functional Similarity.

Area of Science:

• Computational Biology
• Bioinformatics
• Protein Science

Background:

• Protein Language Models (PLMs) are increasingly used for protein sequence representation.
• Pretrained PLM embeddings offer potential as direct numeric proxies for protein structure and function.
• No additional training or fine-tuning is required for using these embeddings.

Purpose of the Study:

• To explore the utility of pretrained PLM embeddings as direct numeric proxies for protein structure and function.
• To assess the ability of PLM embeddings to discriminate subtle functional differences using cross-species complementation triplets.
• To investigate the correlation between embedding similarity and functional conservation in orthology relationships.

Main Methods:

• Utilized embeddings from the ESM-C 600M PLM.
• Applied pooled sliced-Wasserstein embeddings to summarize amino acid embeddings.
• Employed optimal transport theory for comparing embedding distributions.
• Analyzed 22 cross-species complementation triplets for functional similarity.
• Explored orthology relationships and the Ortholog Conjecture.

Main Results:

• Pooled sliced-Wasserstein embeddings demonstrated high discrimination of subtle functional differences in complementation triplets.
• Preliminary results suggest embedding similarity may correlate with functional conservation in orthologs (approx. two-thirds of cases).
• No clear differences in embedding similarity were observed between one-to-one orthologs and inparalogs at equivalent sequence divergence.

Conclusions:

• PLM embeddings show promise for protein functional analysis and detecting functional divergence.
• A hybrid approach integrating PLMs with phylogenetic methods is proposed for enhanced homology detection and evolutionary precision.
• Embedding-tree versus gene-tree discordance is introduced as a metric for functional divergence detection.