Compressing Chemistry Reveals Functional Groups for Bioactivity Prediction

An unsupervised learning algorithm based on the Minimum Message Length principle was used to search for compressing substructures in around three million biologically relevant molecules. The discovered substructures contain most human-curated functional groups as well as novel larger patterns. The research was authored by Ruben Sharma and last updated in March 2026.

Use Cases

• Training ridge regression models for bioactivity regression tasks based on discovered data set-specific functional groups.
• Developing novel molecular fingerprint representations based on substructures that compress large molecular datasets.
• Comparing the performance of discovered functional group fingerprints against traditional representations like MACCS and Morgan fingerprints.

Strengths

• Algorithm was tested on around three million biologically relevant molecules.
• Fingerprints from discovered functional groups reportedly outperform MACCS and Morgan fingerprints in specific tasks.
• Dataset is small (5.9 KB), facilitating quick download and inspection.

Limitations

• Row count and column-level documentation are absent; field semantics must be inferred after download.
• The dataset's 5.9 KB size suggests it contains results or code, not the raw molecular data, limiting direct use for modeling.

Provenance

Source

figshare, authored by Ruben Sharma.

Collection Method

Unsupervised learning algorithm (Minimum Message Length) applied to molecular data.

Freshness

Last updated 2026-03-18 21:13:20; freshness should be verified.