ProMol_Func: A Structure-Free Deep Learning Model for Virtual Screening.

Area of Science:

• Computational drug discovery
• Deep learning in pharmacology
• Bioinformatics and cheminformatics

Background:

• Structure-based drug design (SBDD) is computationally intensive and limited by protein structure availability.
• Existing models often have high false-positive rates due to insufficient negative binding data.
• Scalability and generalization are key challenges in computational drug discovery.

Purpose of the Study:

• To develop a structure-free deep learning framework for efficient and scalable drug discovery.
• To improve the screening power and generalization of drug candidate identification models.
• To address the limitations of SBDD by utilizing only protein sequences and small molecule structures.

Main Methods:

• Developed ProMol_Func, a deep learning framework integrating graph-based small molecule encodings with protein function embeddings from amino acid sequences.
• Augmented training datasets with experimentally validated inactives and random decoys to improve model robustness.
• Evaluated performance on the LIT-PCBA benchmark and conducted zero-shot prospective testing on E. coli DnaK.

Main Results:

• Achieved state-of-the-art performance on the LIT-PCBA benchmark with an enrichment factor (EF1%) of 10.9.
• Demonstrated robust screening power in realistic assay settings.
• Successfully identified novel inhibitors for E. coli DnaK in a zero-shot setting, validating its potential for new targets.

Conclusions:

• ProMol_Func offers an efficient and scalable alternative to structure-dependent methods in early-stage drug discovery.
• The structure-free approach enhances model generalization and reduces false positives.
• ProMol_Func shows promise for discovering binders against novel protein targets.