Improved Deep Learning Based Method for Molecular Similarity Searching Using Stack of Deep Belief Networks

Area of Science:

• Computational chemistry
• cheminformatics
• machine learning in drug discovery

Background:

• Virtual screening (VS) is crucial for identifying lead molecules in drug discovery.
• Current VS methods often use 2D fingerprints and assume equal feature importance, potentially limiting accuracy.
• Differentiating feature importance can enhance molecular similarity searching.

Purpose of the Study:

• To improve virtual screening performance by developing a novel multi-descriptor similarity searching method.
• To reweight molecular features using deep belief networks (DBN) to account for varying importance.
• To create a new descriptor by merging reweighted features for enhanced similarity searching.

Main Methods:

• Applied deep belief networks (DBN) to reweight molecular features for similarity searching.
• Utilized multiple descriptors on the MDL Drug Data Report (MDDR) dataset.
• Implemented a multi-descriptor approach, Stack of Deep Belief Networks (SDBN), by merging reweighted features.

Main Results:

• The proposed SDBN method demonstrated superior performance compared to benchmark methods like Bayesian Inference Networks (BIN) and Tanimoto similarity (TAN).
• The reweighting strategy using DBN led to a lower error rate in feature selection.
• The multi-descriptor approach achieved higher accuracy on structurally diverse datasets.

Conclusions:

• The Stack of Deep Belief Networks (SDBN) method significantly enhances virtual screening accuracy by effectively reweighting molecular features.
• This multi-descriptor approach offers a more robust and accurate alternative to traditional similarity searching methods.
• The findings highlight the potential of deep learning in optimizing computational drug discovery processes.