Convergent Enzyme Classification: Traditional Approaches vs. Foundation Models

Authors

• Mahbuba Tasmin
• Bryn Reimer

Course: CS 690U, University of Massachusetts Amherst (Spring 2025)

Project Overview

This project evaluates the effectiveness of traditional bioinformatics techniques on the Convergent Enzyme Classification task from the DGEB benchmark. We specifically assess how simple modeling pipelines — such as BLAST-based nearest neighbor search and logistic regression using basic encodings — compare to the performance of modern foundation models like ESM.

---

Biological Motivation

The convergent_enzymes dataset comprises protein sequences of enzymes annotated with Enzyme Commission (EC) numbers, where training and test sequences with the same EC number share little to no sequence similarity. This simulates convergent evolution, where similar function has evolved independently multiple times. Because of this, even foundation models struggle to generalize — making this an ideal case for re-evaluating traditional approaches.

---

Dataset

• Source: TattaBio Convergent Enzymes
• Train set: 2000 amino acid sequences
• Test set: 400 amino acid sequences
• Task: Predict EC number (multi-class classification)

---

Methods

Traditional Machine Learning (PyTorch & scikit-learn)

• Inputs: Amino acid sequences
• Encodings:
  • One-hot encoding (e.g., max-len x 21 )
  • k-mer count vectors
  • Count encoding
  • Amino acid properties encoding
• Model: Multinomial Logistic Regression (implemented in PyTorch)
• Evaluation:
  • Accuracy
  • Macro F1-score
  • Number correct

BLAST-based Classification

• Use blastp to search test sequences against the training set
• Predict EC label based on the top BLAST hit
• Evaluate using accuracy, macro F1, and number correct

---

Usage

All results available in the main Jupyter notebook