ConPLex

ConPLex predicts drug-target interactions by integrating pretrained protein language models with a protein-anchored contrastive coembedding approach to enable large-scale computational identification and characterization of binding partners.

Key Features:

• Deep Learning Architecture: ConPLex integrates pretrained protein language models (PLex) with a protein-anchored contrastive coembedding technique ("Con") to learn joint representations of drugs and proteins and analyze distances between their embeddings.
• High Accuracy and Adaptivity: The model demonstrates high accuracy in predicting binding affinities, maintains specificity against decoy compounds, and generalizes to unseen data as shown by experimental validation of kinase-drug predictions (12 of 19 validated, including four with subnanomolar affinity and an EPHB1 inhibitor with K_D = 1.3 nM).
• Scalability: ConPLex is designed to scale to extensive compound libraries and the entire human proteome, supporting genome-wide drug screening applications.
• Interpretability: The learned embeddings are interpretable and enable visualization of the drug–target embedding space to explore functional relationships among proteins, including human cell-surface proteins.

Scientific Applications:

• Drug Discovery Acceleration: Computational prediction of drug-target interactions to prioritize candidate compounds and complement experimental screens in early-stage drug development.
• Validation of Predictions: Provides predictions that have been experimentally tested and validated for kinase-drug interactions, including several high-affinity binders (12/19 validated; four subnanomolar; EPHB1 K_D = 1.3 nM).
• Functional Characterization: Enables exploration and characterization of human cell-surface protein functions based on their interaction profiles with potential drugs.

Methodology:

Pretrained protein language models (PLex) are combined with a protein-anchored contrastive coembedding method ("Con") to learn joint drug and protein representations and predict interactions by analyzing distances between learned embeddings within a contrastive learning framework.