Distributed Representation of Chemical Fragments.

Area of Science:

• Computational chemistry
• Machine learning
• Cheminformatics

Background:

• Traditional molecular representations like sparse "one-hot" vectors lack rich relational information.
• Computing similarity for small molecular fragments can be challenging.

Purpose of the Study:

• To develop an unsupervised machine learning method for generating distributed vector representations of molecular fragments.
• To enable efficient similarity computations between molecular fragments of varying sizes.
• To apply these vectors to tasks such as clustering, ligand recall, and QSAR modeling.

Main Methods:

• Adapted a word embedding algorithm from natural language processing for molecular fragments.
• Trained the model on approximately 6 million unlabeled chemical structures from PubChem.
• Averaged fragment vectors to create representations for larger molecules.

Main Results:

• Generated dense, high-dimensional fragment vectors capturing rich structural relationships.
• Demonstrated superior performance in clustering ring systems and recalling kinase ligands compared to standard binary fingerprints.
• Showcased the utility of unsupervised learning for fragment chemistry.

Conclusions:

• Unsupervised learning of fragment chemistry from large unlabeled datasets is feasible and effective.
• Distributed fragment vectors offer advantages over traditional representations for various cheminformatics tasks.
• This approach facilitates subsequent supervised learning on smaller labeled datasets.