Figure 1: The overall architecture of the EMMA multi-task learning framework. Self, Cross and FF are abbreviations for self-attention and cross-attention and feed-forward, respectively. Red arrows represent residual connection followed by normalization (summation sign). (a) Modality-specific encoders. (b) Projection block. (c) Dual-stream transformer mechanisms. (d) Concatenation. (e) Task-specific prediction heads.

Description:

Predicting enzyme–small molecule interactions is critical for drug discovery. While recent deep learning approaches have shown promising results, several challenges remain: the lack of a comprehensive dataset, the need to capture enzyme–small molecule interaction signals, the tendency of previous models to simplify the problem as enzyme–substrate vs. enzyme--non-interacting, thereby, misclassify enzyme–inhibitor pairs as substrates, and the true impact of data leakage on model performance. To address these issues, we present EMMA, a transformer-based multi-task learning framework designed to learn pairwise interaction signals between enzymes and small molecules while adapting to out-of-distribution data. EMMA operates directly on SMILES strings and enzyme sequences, with two classification heads that distinguish enzyme–substrate, enzyme–non-interacting, and enzyme–inhibitor pairs. By evaluating EMMA under four distinct data-splitting regimes that control for different types of leakage, we demonstrate that EMMA achieves strong and robust performance, particularly on novel enzyme–small molecule combinations.

Table of Contents:

• Setup Instructions
  • Folder Structure
  • Setting up EMMA Environment
• Creating EMMA Dataset
  • 1-Search_GO_annotation.py
  • .
  • .
  • .
  • 12-Pooling_All_Data.py
• Embedding
  • 13-1-MoLFormer-XL.py
  • 13-2-ESM2_t30.py
• Data Splitting
  • 14-1-SplitByDataSAIL.py
  • 14-2-SplitByDataSAIL_LabelBased.py
• Map Embedding To Splits
  • 15-MapEmbeddingsOutputsToIDs.py
• Training EMMA
  • 16-SelfCrossAttentionDualStreamTransformer.py
• Results and Analysis
  • 17-1-Data_leakage_calculation.py
  • 17-2-Results_and_Analysis.ipynb

Setup Instructions:

Folder structure:

├── EMMA/
│ ├── data/
│ │ ├── Embedded_sequences_t30/
│ │ ├── Embedded_smiles/
│ │ ├── processed_data/
│ │ ├── raw_data/
│ │ ├── splits/
│ │ ├── trained_models/
│ │ └── training_test_results/
│ ├── code/
│ ├── additional_code/
│ ├── utilities/
│ ├── EMMA.png
│ ├── EMMA_dataset.png
│ └── README.md

Setting up EMMA Environment:

  conda create --name EMMA python=3.12.0
  conda activate EMMA
  conda config --set channel_priority strict
  conda config --set solver libmamba
  conda install mamba -n EMMA -c conda-forge
  mamba clean -a
  mamba install -c conda-forge -c bioconda -c kalininalab datasail
  conda install conda-forge::biopython
  conda install pytorch torchvision torchaudio pytorch-cuda=12.1 -c pytorch -c nvidia
  conda install numpy=1.26.4
  conda install pandas=2.2.3
  conda install conda-forge::rdkit
  conda install -c conda-forge seaborn
  pip:
      - transformers==4.48.1
      - scipy==1.15.2
      - scikit-learn==1.6.1
      - rdkit-pypi==2025.3.2
      - chemdataextractor==1.3.0
      - chembl_webresource_client==0.10.9
      - libchebipy==1.0.10
      - matplotlib==3.10.1
      - seaborn==0.13.2
      - fair-esm==2.0.0
      - py-cd-hit==1.1.4
      - tqdm==4.67.1
      - colorama==0.4.6
      - requests==2.32.3
      - grakel
      - pytorch-toolbelt
      - wandb==0.19.8

Creating EMMA Dataset:

Figure 1: Overview of EMMA dataset processing steps and its sources

Embedding:

Run:

    python 13-1-MolFormer_XL.py --input-path ./../data/processed_data/Final_Dataset.pkl --output-path ./../data/Embedded_smiles/
    python 13-2-ESM2_t30.py  --input-path ./../data/processed_data/Final_Dataset.pkl --output-path ./../data/Embedded_sequences_t30/

Data Splitting

This table outlines an overview of all different split strategies used in this project, including the number of training and test samples as well as the ratios of enzyme–inhibitor, enzyme–substrate, and enzyme–non-interacting pairs.

Split Method	Train Samples	Test Samples	Enz–Inh / Enz–Sub (Train)	Enz–Inh / Enz–Sub (Test)	Enz–Sub / Enz–NI (Train)	Enz–Sub / Enz–NI (Test)	Enz–Inh / Enz–NI (Train)	Enz–Inh / Enz–NI (Test)
Enzyme-based	117,771	29,453	0.985	0.932	0.424	0.454	0.417	0.423
Molecule-based	117,863	29,361	1.033	0.786	0.403	0.547	0.416	0.430
Label-based	117,594	29,630	0.973	0.977	0.433	0.417	0.421	0.408
Two-dimensional	64,881	16,083	1.193	0.763	0.324	0.666	0.387	0.508

Notes:

• Enz–Inh: Enzyme–inhibitor
• Enz–Sub: Enzyme–substrate
• Enz–NI: Enzyme–non-interacting

• DataSAIL can split data in 1 and 2 dimensions (1D, 2D). The 1D splits are [S1ₗ (or C1e), S1ₚ (or C1f), I1ₗ (or I1e), I1ₚ (or I1f)] and the 2D splits are S2 (or C2) and I2. For more information please check the DataSAIL's documentation.
• In this project we refer to the split method that was used in the ESP paper as ESP split.

14-1-SplitByDataSAIL.py:

• This script aims to split the dataESP by DataSAIL and generate negative data for each split.

    python 14-1-SplitByDataSAIL.py --split-method C2 --split-size 8 2 --strat True --input-path ./../data/processed_data/Final_Dataset.pkl --output-path ./../data/splits --epsilon-value 0.01 --delta-value 0.01
    python 14-1-SplitByDataSAIL.py --split-method C1f --split-size 8 2 --strat True --input-path ./../data/processed_data/Final_Dataset.pkl --output-path ./../data/splits --epsilon-value 0.01 --delta-value 0.01
    python 14-1-SplitByDataSAIL.py --split-method C1e --split-size 8 2 --strat True --input-path ./../data/processed_data/Final_Dataset.pkl --output-path ./../data/splits --epsilon-value 0.003 --delta-value 0.003
  

14-2-SplitByDataSAIL_LabelBased.py:

    python 14-2-SplitByDataSAIL_LabelBased.py --split-method C1 --split-size 8 2 --strat True --input-path ./../data/processed_data/Final_Dataset.pkl --output-path ./../data/splits --epsilon-value 0.01 --delta-value 0.01

Map Embedding To Splits

15-MapEmbeddingsOutputsToIDs.py:

    python 15-MapEmbeddingsOutputsToIDs.py  --path-split-folder ./../data/splits/C2_epsilon0.01_delta0.01_2S  --path-embedding-smiles ./../data/Embedded_smiles --path-embedded-sequence ./../data/Embedded_sequences_t30
    python 15-MapEmbeddingsOutputsToIDs.py  --path-split-folder ./../data/splits/C1f_epsilon0.01_delta0.01_2S  --path-embedding-smiles ./../data/Embedded_smiles --path-embedded-sequence ./../data/Embedded_sequences_t30
    python 15-MapEmbeddingsOutputsToIDs.py  --path-split-folder ./../data/splits/C1e_epsilon0.003_delta0.003_2S  --path-embedding-smiles ./../data/Embedded_smiles --path-embedded-sequence ./../data/Embedded_sequences_t30
    python 15-MapEmbeddingsOutputsToIDs.py  --path-split-folder ./../data/splits/C1_epsilon0.01_delta0.01_2S  --path-embedding-smiles ./../data/Embedded_smiles --path-embedded-sequence ./../data/Embedded_sequences_t30

Training EMMA:

16-SelfCrossAttentionDualStreamTransformer.py:

    python 16-SelfCrossAttentionDualStreamTransformer.py --used_split_tech C2 --molecule_column_name MolFormer --protein_column_name ESM2t30
    python 16-SelfCrossAttentionDualStreamTransformer.py --used_split_tech C1f --molecule_column_name MolFormer --protein_column_name ESM2t30
    python 16-SelfCrossAttentionDualStreamTransformer.py --used_split_tech C1e --molecule_column_name MolFormer --protein_column_name ESM2t30 
    python 16-SelfCrossAttentionDualStreamTransformer.py --used_split_tech C1 --molecule_column_name MolFormer --protein_column_name ESM2t30

• --used_split_tech It automatically reload the related train and test sets

Results and Analysis

17-1-Data_leakage_calculation.py:

17-2-Results_and_Analysis.ipynb: