Parameter Efficient Deep Learning Models for Multi-Target Binding Affinity and hERG Cardiotoxicity Prediction

Area of Science:

• Computational chemistry
• Drug discovery
• Toxicology

Background:

• Accurate prediction of binding affinity and toxicity is crucial for efficient drug discovery, aiming to lower development costs and improve drug safety.
• Traditional computational methods like molecular docking and QSAR models have limitations in scalability and generalizability due to reliance on handcrafted features.
• Chemical Language Models (CLMs) offer a scalable approach by learning molecular representations from Simplified Molecular Input Line Entry System (SMILES) strings.

Purpose of the Study:

• To develop and evaluate a novel dual-paradigm computational framework for predicting drug binding affinity and cardiotoxicity.
• To compare the performance of the proposed framework against existing computational methods in drug discovery.
• To demonstrate the efficiency of Low-Rank Adaptation (LoRA) for fine-tuning CLMs for toxicological predictions.

Main Methods:

• A hybrid framework integrating a Graph Neural Network (GNN) operating on molecular graphs and a fine-tuned Chemical Language Model (CLM) using Low-Rank Adaptation (LoRA).
• The GNN component processes molecular structures directly, treating atoms and bonds as nodes and edges.
• The CLM component is efficiently adapted for specific toxicological tasks, such as predicting cardiotoxicity for the human ether-á-go-go related gene (hERG).

Main Results:

• The hybrid framework achieved a high average Area Under the Receiver Operating Characteristic curve (AUROC) of 0.92 for predicting binding affinity across three protein targets.
• The LoRA-adapted CLM demonstrated a strong AUROC of 0.93 for cardiotoxicity prediction.
• The LoRA fine-tuning resulted in a significant 98% reduction in trainable parameters compared to traditional fine-tuning methods, while outperforming existing models.

Conclusions:

• The proposed dual-paradigm framework offers a powerful and efficient approach for predicting drug binding affinity and cardiotoxicity.
• The integration of GNNs and LoRA-adapted CLMs represents a significant advancement in computational drug discovery, enhancing predictive accuracy and reducing computational overhead.
• This methodology holds promise for accelerating the drug development pipeline by enabling more reliable early-stage safety and efficacy assessments.