vedatonuryilmaz/dark-proteins-model

Dark Proteins PCI/LIFA Prototype

This repository contains a working prototype for genome-scale protein-chemical interaction (PCI) and ligand-induced functional activity (LIFA) prediction for dark proteins.

Architecture: frozen ESM-2 t6 8M protein encoder + hashed SMILES ligand encoder + multitask heads for binding, activity, and affinity.

OOD dark-protein evaluation: {"binding_accuracy": 0.8888888888888888, "activity_accuracy": 0.8888888888888888, "affinity_rmse": 1.5511236557645622}

Artifacts

• dark_protein_model.pt: trained PyTorch checkpoint
• avil_screen.csv: ranked AVIL inhibitor/ligand candidates
• drd2_drd3_dual_antagonists.csv: ranked DRD2/DRD3 dual antagonist candidates

Top AVIL candidates

target	compound	smiles	binding_probability	predicted_pAffinity	activity	activity_confidence
AVIL	benzamide	CC(C)N1CCC(C(=O)Nc2ccccc2N2CCCCC2)CC1	0.417894	6.50255	inactive	0.613166
AVIL	tyramine	CN(C)CCc1ccc(O)cc1	0.417845	6.48015	inactive	0.612231
AVIL	haloperidol-like	Clc1ccc(C2CCNCC2)cc1	0.416993	6.52308	inactive	0.614537
AVIL	naltrexone-like	C=CCN1CC(C(=O)OC)C=C2c3cc4c(cc3CC21)OCO4	0.416627	6.53851	inactive	0.615551
AVIL	arylpiperazine	CCOc1ccccc1N1CCN(C)CC1	0.41661	6.52149	inactive	0.614814

Top DRD2/DRD3 dual candidates

target_DRD2	compound	smiles	binding_probability_DRD2	predicted_pAffinity_DRD2	activity_DRD2	activity_confidence_DRD2	target_DRD3	binding_probability_DRD3	predicted_pAffinity_DRD3	activity_DRD3	activity_confidence_DRD3	dual_score
DRD2	benzamide	CC(C)N1CCC(C(=O)Nc2ccccc2N2CCCCC2)CC1	0.589879	6.9194	inactive	0.392142	DRD3	0.677743	7.30553	antagonist	0.409369	0.633811
DRD2	tyramine	CN(C)CCc1ccc(O)cc1	0.589791	6.89811	inactive	0.391389	DRD3	0.67738	7.28465	antagonist	0.409597	0.633586
DRD2	haloperidol-like	Clc1ccc(C2CCNCC2)cc1	0.589155	6.93862	inactive	0.393289	DRD3	0.677397	7.32427	antagonist	0.408889	0.633276
DRD2	naltrexone-like	C=CCN1CC(C(=O)OC)C=C2c3cc4c(cc3CC21)OCO4	0.588766	6.95339	inactive	0.394221	DRD3	0.677305	7.33871	antagonist	0.408469	0.633035
DRD2	arylpiperazine	CCOc1ccccc1N1CCN(C)CC1	0.58875	6.9371	inactive	0.393589	DRD3	0.677092	7.32269	antagonist	0.408721	0.632921

Generated by ML Intern

This model repository was generated by ML Intern, an agent for machine learning research and development on the Hugging Face Hub.

• Try ML Intern: https://smolagents-ml-intern.hf.space
• Source code: https://github.com/huggingface/ml-intern

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = 'vedatonuryilmaz/dark-proteins-model'
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)

For non-causal architectures, replace AutoModelForCausalLM with the appropriate AutoModel class.