AUKCAT-CVAE

AUKCAT-CVAE is a conditional variational autoencoder designed to generate high-quality synthetic data to enhance training and generalization of AI models for kcat prediction. Its underlying algorithm, however, is domain-agnostic and can be applied to any scenario requiring data augmentation for model generalization.

Dependencies

1. pytorch 1.10.0
2. pandas 1.4.2
3. numpy 1.24.3
4. sklearn 1.0.2
5. scipy 1.5.3
6. matplotlib 3.7.2
7. CUDA 11.1

Input Data to the Model

The CVAE model uses substrate–EC number–species triples paired with experimentally measured kcat values as input instances. Substrates are embedded via Mol2Vec, EC numbers via EC2Vec, and species via Node2Vec.

The file ./Datasets/Original_data.csv contains all input instances used to train the CVAE model described in the paper. The file ./Datasets/example_data.csv contains example data after embedding, which is used as input to the CVAE model.

Train the CVAE Model

To train the CVAE model for synthetic data generation, run CVAE.py. The trained model will be saved to the ./Trained_model directory.

Generate Synthetic Data Using the Trained CVAE Model

To generate synthetic data points using the trained CVAE model, run CVAE_data_generation.py. The number of synthetic instances per original data point can be configured via the n_replicates parameter. The generated dataset will be saved in the ./Synthetic_data folder.