TLDR: A new study introduces an unsupervised learning algorithm, FGCompress, based on the Minimum Message Length principle, to identify substructures that compress large chemical datasets. The algorithm successfully rediscovers most human-curated functional groups and identifies novel, biologically relevant patterns. Furthermore, molecular fingerprints generated from these compression-derived substructures significantly improve the performance of machine learning models in bioactivity prediction tasks compared to traditional fingerprints.
A recent study delves into the fundamental building blocks of chemistry, known as functional groups, by applying a principle from computational learning theory: compression. Functional groups are specific arrangements of atoms within molecules that are responsible for the molecule’s characteristic chemical reactions and properties. While chemists have long used these groups to explain molecular behavior, there hasn’t been an objective, large-scale assessment of their utility.
The research, titled Compressing Chemistry Reveals Functional Groups, introduces a novel unsupervised learning algorithm that searches for substructures within molecules that can effectively ‘compress’ chemical data. The core idea is that a good explanation of data should also be able to compress it, meaning it can represent the data more concisely. This concept is rooted in the Minimum Message Length (MML) principle, which seeks the most concise explanation for data by balancing the complexity of a hypothesis with how well it fits the data.
The algorithm, named FGCompress, works by identifying repeating substrings (which correspond to chemical substructures) in a large dataset of molecules. It iteratively selects substructures that provide the greatest compression, effectively building a ‘codebook’ of these important chemical patterns. The process continues until no further compression can be achieved. To ensure chemical validity, the algorithm filters out substrings that do not represent actual chemical substructures.
The researchers applied FGCompress to the ChEMBL dataset, which contains nearly three million biologically relevant molecules. The findings were striking: the algorithm successfully discovered substructures that largely correspond to the well-known, human-curated functional groups that chemists have used for decades. For instance, among the top discovered substructures were the carbonyl group, trifluoromethyl group, methyl group, amide group, and benzene ring. This provides strong computational validation for the traditional understanding of functional groups.
Beyond validating known functional groups, FGCompress also uncovered novel, larger patterns with more specific biological functions. These included substructures found in important drug classes, such as a fragment of desosamine (central to macrolide antibiotics), a substructure present in triterpene compounds (precursors to cholesterol), and patterns found in antiviral therapies like Zidovudine and antifungal medications like Fluconazole. The algorithm even identified specific amino acid sequences, highlighting its ability to find patterns relevant to biological activity.
The study also investigated the practical utility of these discovered substructures in machine learning tasks. A second algorithm, FGFingerprinter, was developed to generate molecular fingerprints from the substructures identified by FGCompress. These MML87 fingerprints were then used to train ridge regression models for 24 different bioactivity prediction datasets.
The results demonstrated a significant advantage: models trained with the MML87 fingerprints consistently outperformed those using conventional MACCS and Morgan fingerprints. This suggests that the substructures identified through data compression are not only chemically meaningful but also highly effective features for predicting molecular bioactivity, particularly in linear regression models. The choice of ridge regression was deliberate, as it allows for a direct assessment of the quality of the molecular representations without the added complexity of more advanced machine learning models.
Also Read:
- ReaDISH: A New Approach to Chemical Reaction Prediction with Permutation Invariance and Interaction Awareness
- S2Drug: Enhancing Drug Discovery by Combining Protein Sequence and 3D Structure Data
In conclusion, this research offers a compelling computational perspective on the nature of functional groups. By demonstrating that these groups naturally emerge from the principle of data compression, the study provides an objective basis for their importance in chemistry. Furthermore, the ability of these compression-derived substructures to enhance machine learning performance underscores their value as a powerful tool for drug discovery and chemical research.
