license: mit

base_model: westlake-repl/SaProt_35M_AF2

Model Description

This model is fine-tuned to predict the mutation effects of *Bacillus subtilis* α-amylase. It takes amino acid sequences as input and performs protein-level regression to predict quantitative enzyme activity values, enabling accurate assessment of how mutations alter enzyme function.

Task type: protein-level regression

Model input type: Amino acid sequence

Dataset

The dataset is sourced from van der Flier et al. (2024), available at https://www.sciencedirect.com/science/article/pii/S2001037024002940. It contains a total of 3706 rows, with each sample carrying 1 to 8 mutation sites. The full dataset is randomly split into training, validation, and test sets following an 8:1:1 ratio. The target label is absorbance, ranging from -0.001 to 0.211. A higher absorbance value indicates greater starch degradation, corresponding to stronger detergent activity of the amylase enzyme.
Performance (on test set)

Spearman correlation: 0.76

LoRA config

r: 8

lora_dropout: 0.1

lora_alpha: 16

target_modules: ["query", "intermediate.dense", "value", "output.dense", "key"]

modules_to_save: ["classifier"]

Training config

optimizer:

class: AdamW

betas: (0.9, 0.98)

weight_decay: 0.01

learning rate: 0.0005

epoch: 30

batch size: 64

precision: 16-mixed