Feature Reduction for Molecular Similarity Searching Based on Autoencoder Deep Learning.

Area of Science:

• Computational Chemistry
• Cheminformatics
• Drug Discovery

Background:

• Molecular similarity is vital in rational drug design for retrieving functionally similar molecules from databases.
• Conventional 2D fingerprint methods for molecular similarity assessment often contain redundant and irrelevant features, hindering search performance.

Purpose of the Study:

• To develop a novel approach for identifying essential molecular features by representing them using Autoencoders (AE).
• To enhance molecular similarity searching by effectively removing irrelevant and redundant molecular descriptors.

Main Methods:

• Utilized Autoencoder (AE) models for feature representation and dimensionality reduction of molecular data.
• Experimented with the MDL Data Drug Report (MDDR) dataset to evaluate the proposed method.
• Compared the AE-based approach against established similarity methods like Tanimoto Similarity (TAN), ASMTP, and SQB.

Main Results:

• The proposed Autoencoder (AE) approach demonstrated superior performance compared to benchmark methods (TAN, ASMTP, SQB).
• The method showed significant improvements in molecular similarity searching, particularly for structurally heterogeneous datasets.
• Effective identification and removal of redundant/irrelevant features contributed to enhanced search accuracy.

Conclusions:

• Autoencoder (AE) based feature identification offers a powerful strategy for improving molecular similarity searching.
• This approach provides a more robust and accurate method for analyzing chemical datasets, especially complex ones.
• The findings suggest a promising direction for advancing rational drug design through enhanced cheminformatics tools.