Compression of molecular fingerprints with autoencoder networks

Area of Science:

• Computational chemistry
• Machine learning in cheminformatics

Background:

• Molecular fingerprints are crucial for representing chemical structures in machine learning.
• Autoencoder neural networks offer a method for dimensionality reduction and feature learning.

Purpose of the Study:

• To evaluate the impact of autoencoder compression on molecular fingerprint performance in downstream tasks.
• To investigate the role of property co-learning in enhancing compressed fingerprint utility.

Main Methods:

• Binary molecular fingerprints were compressed using an autoencoder neural network.
• The performance of compressed fingerprints was assessed in classification and regression tasks.
• Property co-learning was integrated to evaluate its influence on predictive accuracy.

Main Results:

• Classification models showed negligible performance changes with compressed fingerprints.
• Regression models, particularly for longer fingerprints (e.g., Morgan, RDK), benefited from compression.
• Performance degradation in regression was observed beyond 90% compression levels.
• Property co-learning improved compressed fingerprint predictive power by up to 20%.

Conclusions:

• Autoencoder compression is a viable technique for molecular fingerprints, especially for regression tasks.
• Property co-learning effectively enhances the predictive capabilities of compressed molecular representations.
• Tailored compression strategies are necessary to avoid performance loss at high compression ratios.