One of the most relevant problems in computational biology is the annotation of protein function starting from the protein residue sequence. When methods based only on the comparison of protein sequences are adopted, a reliable transfer of annotation is possible only when two proteins share a high level of similarity. In recent years, the advent of Protein Language Models (PLMs) has provided new ways to encode protein sequences by means of embeddings, in which each residue in the sequence is represented with a high-dimensional vector conveying contextual information. Consequently, methods for comparing proteins based on their embeddings have been implemented, such as the Embedding-Based Alignment (EBA). In this thesis, we investigate how much the EBA method expands the possibility of functional annotation transfer as compared to the traditional sequence-based alignment. We restricted our analysis to a set of 256,769 enzymes, whose function is hierarchically encoded with the Enzyme Commission (EC) numbers. Different protein language models (ESM-2, ProtT5, and ProstT5) are adopted to generate residue-level embeddings for enzymes annotated with a single EC number. The methods are evaluated at different levels of the EC classification hierarchy using the MCC and F1-score. The study also examines the effect of protein length on EBA scores and introduces a self-alignment-based normalization procedure. Overall, the work shows that the EBA method significantly expands the possibility of function annotation transfer with respect to traditional sequence comparison methods (in particular BLAST), even if the accuracy of the annotation largely depends on the level of the classification hierarchy.
One of the most relevant problems in computational biology is the annotation of protein function starting from the protein residue sequence. When methods based only on the comparison of protein sequences are adopted, a reliable transfer of annotation is possible only when two proteins share a high level of similarity. In recent years, the advent of Protein Language Models (PLMs) has provided new ways to encode protein sequences by means of embeddings, in which each residue in the sequence is represented with a high-dimensional vector conveying contextual information. Consequently, methods for comparing proteins based on their embeddings have been implemented, such as the Embedding-Based Alignment (EBA). In this thesis, we investigate how much the EBA method expands the possibility of functional annotation transfer as compared to the traditional sequence-based alignment. We restricted our analysis to a set of 256,769 enzymes, whose function is hierarchically encoded with the Enzyme Commission (EC) numbers. Different protein language models (ESM-2, ProtT5, and ProstT5) are adopted to generate residue-level embeddings for enzymes annotated with a single EC number. The methods are evaluated at different levels of the EC classification hierarchy using the MCC and F1-score. The study also examines the effect of protein length on EBA scores and introduces a self-alignment-based normalization procedure. Overall, the work shows that the EBA method significantly expands the possibility of function annotation transfer with respect to traditional sequence comparison methods (in particular BLAST), even if the accuracy of the annotation largely depends on the level of the classification hierarchy.
Embedding-Based Alignment for Protein Function Annotation: a comparative study of Protein Language Models
BARO, GIADA
2025/2026
Abstract
One of the most relevant problems in computational biology is the annotation of protein function starting from the protein residue sequence. When methods based only on the comparison of protein sequences are adopted, a reliable transfer of annotation is possible only when two proteins share a high level of similarity. In recent years, the advent of Protein Language Models (PLMs) has provided new ways to encode protein sequences by means of embeddings, in which each residue in the sequence is represented with a high-dimensional vector conveying contextual information. Consequently, methods for comparing proteins based on their embeddings have been implemented, such as the Embedding-Based Alignment (EBA). In this thesis, we investigate how much the EBA method expands the possibility of functional annotation transfer as compared to the traditional sequence-based alignment. We restricted our analysis to a set of 256,769 enzymes, whose function is hierarchically encoded with the Enzyme Commission (EC) numbers. Different protein language models (ESM-2, ProtT5, and ProstT5) are adopted to generate residue-level embeddings for enzymes annotated with a single EC number. The methods are evaluated at different levels of the EC classification hierarchy using the MCC and F1-score. The study also examines the effect of protein length on EBA scores and introduces a self-alignment-based normalization procedure. Overall, the work shows that the EBA method significantly expands the possibility of function annotation transfer with respect to traditional sequence comparison methods (in particular BLAST), even if the accuracy of the annotation largely depends on the level of the classification hierarchy.| File | Dimensione | Formato | |
|---|---|---|---|
|
Baro_Giada.pdf
Accesso riservato
Dimensione
5.89 MB
Formato
Adobe PDF
|
5.89 MB | Adobe PDF |
The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License
https://hdl.handle.net/20.500.12608/115818