DTITR

DTITR predicts quantitative drug–target binding affinity using a Transformer-based architecture that models proteins and compounds from 1D sequential and structural data.

Key Features:

• End-to-End DTA prediction: Predicts continuous drug–target binding affinity values rather than binary interaction labels.
• Transformer-Based Architecture: Applies Transformer encoders to 1D raw sequential and structural representations of proteins and compounds with self-attention layers to capture biological and chemical context.
• Cross-Attention Mechanism: Uses cross-attention layers to enable information exchange between protein and compound representations, capturing pharmacological interaction context.
• Multiple Transformer-Encoders and Robust Representations: Integrates multiple Transformer-Encoders, including a Cross-Attention Transformer-Encoder, to produce discriminative aggregate representations of proteins and compounds.
• Explainability and Interpretability: Emphasizes attention-based mechanisms to provide interpretable signals reflecting the joint contribution of interacting substructures.
• Affinity Prediction and Ranking: Produces predicted interaction strength values and supports ranking of binding strengths across compound–target pairs.

Scientific Applications:

• Drug discovery and lead identification: Supports identification and prioritization of potential therapeutic candidates by predicting binding affinities.
• Off-target assessment: Aids distinction between primary interactions and off-target effects by modeling interaction strength.
• Experimental prioritization: Enables ranking of compound–target pairs to prioritize follow-up biochemical or cellular assays.

Methodology:

An end-to-end Transformer-based architecture processes 1D sequential and structural data with self-attention layers to capture molecular context, cross-attention layers to exchange information between proteins and compounds, and multiple Transformer-Encoders to aggregate discriminative representations.