apalania/Polaris-ASAP-OpenADMET2025

Models from experiments with the ASAP-Polaris-OpenADMET Antiviral Challenge 2025

ADME models:

1. LogD ensemble:

• LogD_Augmented_55Desc_random_split1.pth
• LogD_Polaris_25Desc_random_split6.pth
• LogD_Polaris_55Desc_random_split6.pth

2. KSOL: single best model

• KSOL_Polaris_55Desc_scaffold_split6.pth
• KSOL_Polaris_55_scaffold_model_6_retrained.pth

The model before retraining as well as after retraining is provided

3. HLM ensemble:

• HLM_Augmented_25Desc_random_split8.pth
• HLM_Augmented_55Desc_random_split4.pth
• HLM_Polaris_55Desc_random_split5.pth

4. MLM ensemble:

• MLM_Augmented_55Desc_scaffold_split2.pth
• MLM_Polaris_25Desc_random_split9.pth
• MLM_Polaris_55Desc_random_split1.pth

5. MDR1-MDCKII: single best model

• MDR1-MDCK2_Polaris_25Desc_random_split1.pth
• MDR1-MDCK2_Polaris_25_scaffold_model_1_retrained.pth

Both the scaffold and random split for the same model configuration are provided.

Potency models:

1. Base ChEMBL-UniProt model:

• Potency_Base-model_EnhancedFeats.pt
• Potency_BaseModel_tokenizer_enhanced-features.pkl
• Potency_BaseModel_vocab_enhanced-features.pkl

2. Finetuned models:

• Potency_BaseModel_GPFT-Polaris_target-strats_random_p1-10-best.pt
• Potency_BaseModel_GPFT-Polaris_target-strats_scaffold_p1-5-best.pt

3. Domain-specific model GraphML trained from scratch:

• Potency_AttFPGNN-Transformer_Polaris-only_target-strats_random.pt
• Potency_AttFPGNN-Transformer_Polaris-only_target-strats_scaffold.pt
• Potency_AttFPGNN-Transformer_Polaris-only_tokenizer_enhanced_scaffold.pkl
• Potency_AttFPGNN-Transformer_Polaris-only_vocab_enhanced_scaffold.pkl

4. XGBoost models (enhanced with docking scores):

• Potency_xgb_25D_wDock_target-strat_random-split.json
• Potency_xgb_25D_wDock_target-strat_random-split.pkl

5. Stacking meta-learner (overall best):

• Potency_Stacking_overall-best_Polaris-only_random-split.pth

Inference codes:

• adme_inference.py
  • Usage notes
• potency_inference.py
  • Usage notes

References

• Preprint: Experiments with data-augmented modeling of ADME and potency endpoints in the ASAP-Polaris-OpenADMET Antiviral Challenge