models/README.md

AbMelt Trained Models

This directory contains the trained machine learning models for predicting antibody thermostability properties.

Model Files

T_agg (Aggregation Temperature)

• Model: tagg/efs_best_knn.pkl
• Algorithm: K-Nearest Neighbors (KNN)
• Features: tagg/rf_efs.csv
  • rmsf_cdrs_mu_400
  • gyr_cdrs_Rg_std_400
  • all-temp_lamda_b=25_eq=20

T_m (Melting Temperature)

• Model: tm/efs_best_randomforest.pkl
• Algorithm: Random Forest
• Features: tm/rf_efs.csv
  • gyr_cdrs_Rg_std_350
  • bonds_contacts_std_350
  • rmsf_cdrl1_std_350

T_m,onset (Onset Melting Temperature)

• Model: tmon/efs_best_elasticnet.pkl
• Algorithm: ElasticNet
• Features: tmon/rf_efs.csv
  • bonds_contacts_std_350
  • all-temp-sasa_core_mean_k=20_eq=20
  • all-temp-sasa_core_std_k=20_eq=20
  • r-lamda_b=2.5_eq=20

Model Training Details

These models were trained using:

• Feature Selection: Exhaustive Feature Selection (EFS) with Random Forest
• Optimization: Bayesian optimization (skopt) for hyperparameter tuning
• Cross-validation: Repeated K-Fold cross-validation
• Training Data: Internal antibody thermostability dataset

Usage

Models are automatically loaded by the AbMeltPredictor class in src/model_inference.py:

from src.model_inference import AbMeltPredictor

predictor = AbMeltPredictor()
predictions = predictor.predict_all(descriptors_df)

Source

These models and feature definitions were copied from the original AbMelt paper implementation:

• Original location: ../AbMelt/models/
• Paper: Rollins, Z.A., Widatalla, T., Cheng, A.C., & Metwally, E. (2024). AbMelt: Learning antibody thermostability from molecular dynamics.

File Sizes

• tagg/efs_best_knn.pkl: ~1-5 KB (KNN is lightweight)
• tm/efs_best_randomforest.pkl: ~100-500 KB (depends on tree depth/count)
• tmon/efs_best_elasticnet.pkl: ~1-5 KB (linear model is lightweight)

Notes

• Models expect features in the exact order specified in the rf_efs.csv files
• Feature values should be normalized/scaled consistently with training data
• The rf_efs.csv files include both feature columns and target columns (tagg, tm, tmonset)
• These are the best models selected from exhaustive feature selection experiments