Custom Tokenization Dictionary, CUSTODI: A General, Fast, and Reversible Data-Driven Representation and Regressor.

Area of Science:

• Computational chemistry
• Machine learning in chemistry

Background:

• Molecular representations are crucial for predicting chemical properties.
• Existing methods face challenges, especially with limited data.

Purpose of the Study:

• Introduce Custom Tokenization Dictionary (CUSTODI) as a novel molecular representation.
• Evaluate CUSTODI's performance in molecular property prediction.
• Explore CUSTODI's applicability to small training sets.

Main Methods:

• Developed a novel tokenization approach for molecular structures.
• Implemented a machine learning model utilizing the CUSTODI representation.
• Benchmarked performance against established methodologies.

Main Results:

• CUSTODI achieves performance comparable to existing benchmark methods.
• Demonstrated unique effectiveness when applied to small training datasets.
• Theoretical framework developed for a-priori fit quality estimation.

Conclusions:

• CUSTODI presents a viable and effective method for molecular representation and property prediction.
• Its strength with small datasets addresses a key limitation in cheminformatics.
• The theoretical underpinnings suggest broad applicability for assessing predictive model performance.