DeltaCata-AL: Active Learning for Enzyme Engineering

Overview

This repository provides the implementation and reproduction code for the DeltaCata-AL active learning experiments, a downstream application of DeltaCata. DeltaCata-AL leverages DeltaCata's hidden embeddings to guide iterative mutant selection through an active learning framework, demonstrating efficient identification of fitness-enhanced enzyme variants.

Requirements

• Same environment as DeltaCata
• Additional dependency: scikit-learn

pip install scikit-learn

Quick Start: Reproduce Results

Pre-computed DeltaCata embeddings are already provided. Simply run:

python reproduce.py

This automatically executes the complete pipeline:

1. Loads embeddings from the pre-trained DeltaCata
2. Performs 10 rounds of active learning (20 mutants per round)
3. Generates visualization figures in fig/ directory:
   • line_plot_{dms_id}.svg: Cumulative numbers of fitness-enhanced mutants per round
   • box_plot_{dms_id}.svg: Fitness distributions by different methods after 10 rounds

Results are logged to results.log and intermediate data saved to plot_data.pkl.

Extract Embeddings for New DMS Datasets

To apply DeltaCata-AL to new datasets:

1. Prepare your DMS CSV file following the format in ./DeltaCata/data/DMS_data/
2. Place corresponding PDB structure file (same name) in ./DeltaCata/data/pdbs/
3. Run the embedding extraction script: Example:

cd ./DeltaCata/
python predict_and_embed.py --DMS_id AMIE_PSEAE_Wrenbeck_2017 --wt_seq MRHGDISSSNDTVGV...EGLEKEA

Experimental Setup

Datasets: Three benchmark DMS datasets are included:

DMS ID	UniProt ID	Reference DOI	Enzyme name	Substrate	Fitness determination	Wild-type fitness	% of beneficial mutations
BLAT_ECOLX_Firnberg_2014	P62593	10.1093/molbev/msu081	Beta-lactamase TEM	Ampicillin	Antibiotic resistance	1.0	19.5
LGK_LIPST_Klesmith_2015	B3VI55	10.1021/acssynbio.5b00131	Levoglucosan kinase	Levoglucosan	Growth rate	0.0	5.5
AMIE_PSEAE_Wrenbeck_2017	P11436	10.1038/ncomms15695	Aliphatic amidase	Isobutyramide	Growth rate	0.0	22.4

Baselines: DeltaCata-AL is compared against the zero-shot DeltaCata and other methods (EITLEM-Kinetics, CatPred, xTrimoPGLM, MSA Transformer, and ProSST)

Citation

If you use this code, please cite XXX (under review).

Contact

For questions or issues, please contact Qianmu Yuan (yuanqm3@mail3.sysu.edu.cn) or Yuedong Yang (yangyd25@mail.sysu.edu.cn).