FragNet, a Contrastive Learning-Based Transformer Model for Clustering, Interpreting, Visualizing, and Navigating...

Area of Science:

• Cheminformatics
• Computational Chemistry
• Machine Learning

Background:

• Molecular similarity is crucial in cheminformatics, typically assessed via pairwise comparisons.
• Previous methods using variational autoencoders did not optimize latent space for similarity-based clustering.

Purpose of the Study:

• To develop a novel, disentangled latent space for molecular representation.
• To improve the clustering of similar molecules using advanced deep learning techniques.

Main Methods:

• Integration of transformers, contrastive learning, and an embedded autoencoder.
• Variational autoencoder (VAE) embedding of 160,000 biologically relevant molecules.
• Systematic variation of latent space dimensionality to observe molecular separation.

Main Results:

• Achieved a disentangled latent space where similar molecules cluster together.
• Demonstrated effective separation of molecule types across different latent dimensions.
• Validated the representation by analyzing clozapine and flucloxacillin neighbors.

Conclusions:

• Transformers coupled with contrastive learning enable effective one-shot learning for molecular representations.
• The novel approach yields a successful and interpretable disentangled latent space.
• This method enhances molecular similarity assessment by clustering related molecules within the entire training set.