Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Compressing Chemistry Reveals Functional Groups

Ruben Sharma, Ross D. King

TL;DR

This work tackles whether traditional chemical functional groups arise as compressive substructures in large-scale molecular data. It introduces FGCompress, a Minimum Message Length (MML)-based unsupervised algorithm, to identify substructures that best compress nearly 3 million SMILES strings, recovering known functional groups and dataset-specific patterns. It further derives MML87 fingerprints from these substructures and shows they significantly improve ridge-regression bioactivity predictions across 24 datasets compared with MACCS and Morgan fingerprints. The findings provide computational support for the functional-group partitioning of chemistry and offer a practical, compression-driven fingerprint representation with real-world predictive benefits.

Abstract

We introduce the first formal large-scale assessment of the utility of traditional chemical functional groups as used in chemical explanations. Our assessment employs a fundamental principle from computational learning theory: a good explanation of data should also compress the data. We introduce an unsupervised learning algorithm based on the Minimum Message Length (MML) principle that searches for substructures that compress around three million biologically relevant molecules. We demonstrate that the discovered substructures contain most human-curated functional groups as well as novel larger patterns with more specific functions. We also run our algorithm on 24 specific bioactivity prediction datasets to discover dataset-specific functional groups. Fingerprints constructed from dataset-specific functional groups are shown to significantly outperform other fingerprint representations, including the MACCS and Morgan fingerprint, when training ridge regression models on bioactivity regression tasks.

Compressing Chemistry Reveals Functional Groups

TL;DR

This work tackles whether traditional chemical functional groups arise as compressive substructures in large-scale molecular data. It introduces FGCompress, a Minimum Message Length (MML)-based unsupervised algorithm, to identify substructures that best compress nearly 3 million SMILES strings, recovering known functional groups and dataset-specific patterns. It further derives MML87 fingerprints from these substructures and shows they significantly improve ridge-regression bioactivity predictions across 24 datasets compared with MACCS and Morgan fingerprints. The findings provide computational support for the functional-group partitioning of chemistry and offer a practical, compression-driven fingerprint representation with real-world predictive benefits.

Abstract

We introduce the first formal large-scale assessment of the utility of traditional chemical functional groups as used in chemical explanations. Our assessment employs a fundamental principle from computational learning theory: a good explanation of data should also compress the data. We introduce an unsupervised learning algorithm based on the Minimum Message Length (MML) principle that searches for substructures that compress around three million biologically relevant molecules. We demonstrate that the discovered substructures contain most human-curated functional groups as well as novel larger patterns with more specific functions. We also run our algorithm on 24 specific bioactivity prediction datasets to discover dataset-specific functional groups. Fingerprints constructed from dataset-specific functional groups are shown to significantly outperform other fingerprint representations, including the MACCS and Morgan fingerprint, when training ridge regression models on bioactivity regression tasks.

Paper Structure

This paper contains 19 sections, 8 equations, 7 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related Work
  3. Algorithms
  4. FGCompress
  5. FGFingerprinter
  6. Definitions
  7. The Message
  8. Message Length Calculation
  9. Part 1: Communicating the substrings
  10. Part 2: Sending Vocabulary Probabilities and Counts
  11. Part 3: Specifying the Sequence of the Symbols
  12. Full Message Length
  13. Experiments
  14. Experimental Choices
  15. Results and Discussion
  16. ...and 4 more sections

Figures (7)

  • Figure 1: Example drug molecules and necessary functional groups
  • Figure 2: An illustration of an example run of the FGCompress algorithm. At iteration 2, adding the best substring C=O does not further compress the message. The algorithm is then terminated and the codebook at step 2 is taken as final. Length values are invented for illustrative purposes.
  • Figure 3: Example Message Contents for the Original Dataset $\texttt{\{C(=O), N(=O), S(=O)\}}$
  • Figure 4: Top 15 substructures derived by the algorithm. Substructures are numbered in order of relative importance.
  • Figure 5: Top 40 substructures extracted as complete branches discovered by FGCompress. Substructures are numbered in order of relative importance
  • ...and 2 more figures

Theorems & Definitions (4)

  • Definition 1: SMILES symbol
  • Definition 2: Valid Substring
  • Definition 3: Codebook
  • Definition 4: Optimal Coding Scheme