Integrating convolutional layers and biformer network with forward-forward and backpropagation training.

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Accurate molecular property prediction is vital for accelerating drug discovery and computational chemistry.
• Traditional machine learning methods struggle with high-dimensional data and manual feature engineering.
• Existing deep learning models may not fully capture complex molecular structures, creating a research gap.

Purpose of the Study:

• Introduce Deep-CBN, a novel framework to improve molecular property prediction accuracy and efficiency.
• Capture intricate molecular representations directly from raw data.
• Address limitations of current methods in handling complex molecular data.

Main Methods:

• Combine convolutional neural networks (CNNs) for feature learning from SMILES strings.
• Utilize a BiFormer attention mechanism with the forward-forward algorithm for global context refinement.
• Employ backpropagation for fine-tuning the prediction subnetwork.

Main Results:

• Deep-CBN achieved near-perfect ROC-AUC scores on benchmark datasets (Tox21, BBBP, SIDER, ClinTox, BACE, HIV, MUV).
• Significantly outperformed state-of-the-art methods in molecular property prediction.
• Demonstrated effectiveness in capturing complex molecular patterns.

Conclusions:

• Deep-CBN offers a robust and efficient tool for molecular property prediction.
• The framework accelerates drug discovery processes by improving accuracy.
• Highlights the potential of combining CNNs, BiFormer, and novel training algorithms for molecular modeling.