EZHit is a lightweight enzyme–reaction retrieval model for predicting potential catalytic compatibility between enzyme sequences and biochemical reactions.
Given an enzyme amino-acid sequence and a reaction SMILES, EZHit estimates whether the enzyme is likely to catalyze the reaction. The released checkpoints can be used for enzyme–reaction pair prediction, custom fine-tuning, and uncertainty-aware inference with Mahalanobis-distance-based distribution assessment.
The Colab notebook allows users to fine-tune EZHit on their own enzyme–reaction datasets and export a fine-tuned checkpoint together with train_distribution_stat.pt for Mahalanobis-distance inference.
Model variants
Model group
File pattern
Description
General model
binarycls_best_val_seed*.pt
General enzyme–reaction compatibility model
Cytochrome P450 model
ft_p450_best_seed*.pt
Fine-tuned model for cytochrome P450-related prediction
Phosphatase model
ft_phosphatase_best_seed*.pt
Fine-tuned model for phosphatase-related prediction
Terpene synthase model
ft_terpene_best_seed*.pt
Fine-tuned model for terpene synthase-related prediction
Each model group may contain multiple seed checkpoints for ensemble prediction.
Binary label. 1 for compatible enzyme–reaction pairs and 0 for negative pairs
An optional split column can be provided with values train, val, and test.
Output interpretation
EZHit can report the following outputs:
Output
Description
Match probability
Predicted enzyme–reaction compatibility probability
Ensemble uncertainty
Model-disagreement-based uncertainty estimate
Mahalanobis distance
Latent-space distance from the learned training distribution
A typical interpretation is:
Probability
Mahalanobis distance
Interpretation
High
Low
High-priority candidate
High
High
Potentially useful but less reliable or out-of-distribution
Low
Low
In-distribution but predicted as incompatible
Low
High
Low-priority candidate
Thresholds should be adjusted based on the model variant, dataset, and validation results.
Mahalanobis-distance statistics
Mahalanobis-distance inference requires a train_distribution_stat.pt file generated from the same model architecture and latent dimension as the checkpoint used for prediction.
The latent dimension of the statistics file must match the hidden dimension of the checkpoint. For example, if the model hidden dimension is 512, the expected shapes are:
text
1mean: [512]
2inv_cov: [512, 512]
If a checkpoint is fine-tuned with a different hidden dimension, the corresponding Mahalanobis statistics must be regenerated.
For very small fine-tuning datasets, covariance estimation may be unstable. In such cases, Mahalanobis distance should be interpreted cautiously together with probability and ensemble uncertainty.
Fine-tuning
Users can fine-tune EZHit using the Colab notebook:
