FunSiNN: Predicting Functional Similarity of Protein Pairs Open Access

Abstract

The rapid growth of unannotated protein sequence data presents a major challenge

for functional annotation. This thesis proposes a scalable approach for predicting

functional similarity between protein pairs using ProtT5 language model embeddings

and a Siamese Neural Network (SNN). We first generate a similarity network by

thresholding cosine similarity scores between embeddings, followed by subsampled

Louvain clustering to produce functionally similar groups. A refinement step further

improves cluster granularity. From these clusters, we sample labeled protein pairs to

train the SNN, which learns to classify whether two proteins share function.

Experimental results show that refined clusters yield stronger labels and improved

predictive performance (F1 = 0.7670, AUC = 0.8472). Our findings demonstrate the

potential of combining pLM embeddings, unsupervised clustering, and deep learning

to enable large-scale protein function prediction in the absence of curated annotations.

Table of Contents

1 Introduction

1 2 Background 3

2.1 Protein Function Prediction Methods . . . . . . . . . . . . . . . . . . 3

2.2 Protein Database . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

2.3 Transfer Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

3 Approach 8 3.1 Protein Sequence Dataset . . . . . . . . . . . . . . . . . . . . . . . . 8

3.2 ProtT5 Embeddings . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

3.3 Similarity Metric . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

3.3.1 Cosine Similarity Threshold Calculation . . . . . . . . . . . . 9

3.3.2 Threshold Search . . . . . . . . . . . . . . . . . . . . . . . . . 11

3.3.3 Clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

3.3.4 Subsampled Clustering Method . . . . . . . . . . . . . . . . . 12

3.4 SNN Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

3.4.1 Evaluation Metrics . . . . . . . . . . . . . . . . . . . . . . . . 17

4 Experiments and Analysis 19

4.1 Working Environment . . . . . . . . . . . . . . . . . . . . . . . . . . 19

4.2 Results and Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

4.2.1 Theshold For Cosine Similarity . . . . . . . . . . . . . . . . . 19

4.2.2 Clustering Results . . . . . . . . . . . . . . . . . . . . . . . . 21

4.2.3 Assessing Clustering Quality . . . . . . . . . . . . . . . . . . . 23

4.2.4 SNN Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

5 Analysis 26

5.1 Limitations and Future Directions . . . . . . . . . . . . . . . . . . . . 26

5.1.1 Overfitting and Model Stability . . . . . . . . . . . . . . . . . 26

5.1.2 Clustering Label Improvements . . . . . . . . . . . . . . . . . 26

5.1.3 Architectural Improvements for Prediction . . . . . . . . . . . 27

6 Conclusion 28

A Appendix 29

A.1 Cosine Similarity Threshold Experiments . . . . . . . . . . . . . . . . 30

A.2 Clustering Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . 33

A.3 SNN Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34

A.3.1 SNN Experiment 1 . . . . . . . . . . . . . . . . . . . . . . . . 35

A.3.2 SNN Experiment 2 . . . . . . . . . . . . . . . . . . . . . . . . 37

Bibliography 39

About this Honors Thesis

Rights statement

• Permission granted by the author to include this thesis or dissertation in this repository. All rights reserved by the author. Please contact the author for information regarding the reproduction and use of this thesis or dissertation.

School	Emory College
Department	Mathematics and Computer Science
Degree	B.S.
Submission	Honors Thesis
Language	English
Research Field	Biology, Bioinformatics Computer Science
Keyword	proteins siamese neural network large language model
Committee Chair / Thesis Advisor	Yana Bromberg, Emory University
Committee Members	Gordon Berman, Emory University Steven La Fleur, Emory University

Last modified

Primary PDF

Thumbnail	Title	Date Uploaded	Actions
	FunSiNN: Predicting Functional Similarity of Protein Pairs ()	2025-04-08 14:35:44 -0400	Press to Select an action Download

Supplemental Files

Thumbnail	Title	Date Uploaded	Actions