Discovering H-Bonding Rules in Crystals with Inductive Logic Programming

Howard Y. Ando, Luc De Raedt, Walter Luyten, Elke Van Craenenbroeck, Henk Vandecasteele, Luc Van MeerveltMolecular 2006 · DOI

TL;DR

This paper introduces a novel approach using Inductive Logic Programming (ILP) to automatically discover rules predicting hydrogen bonding (H-bonding) in crystal structures, solely from 2D molecular information. By leveraging the Cambridge Structural Database and a specialized ILP system (DMAX), the researchers developed an expert system that significantly outperforms traditional methods, offering a more adaptable and precise way to understand and predict these crucial molecular interactions.

What Problem Does This Paper Address?

Understanding hydrogen bonding (H-bonding) is critical in many areas of chemistry, particularly in crystal engineering and drug design. H-bonding influences a molecule's crystal structure, cohesive energy, and solubility, impacting its behavior in biological systems. While existing methods classify H-bonding patterns in 3D structures or rely on general rules like Lipinski's "Rule of 5," there's a need for a system that can predict H-bonding directly from simpler, 2D structural information. This is challenging because H-bonding is highly context-dependent, influenced by the local environment of potential donor and acceptor atoms, which is difficult to capture with traditional global molecular descriptors. The authors aim to bridge this gap by developing an automated system that can extract specific, predictive H-bonding rules from large datasets.

Methods

The core of this research lies in the application of Inductive Logic Programming (ILP), a subfield of machine learning that learns logical rules from data. Specifically, the authors employed DMAX, an ILP system that generates hierarchical rules through a recursive partitioning strategy, similar to decision trees.

Here's a breakdown of their approach:

  • Data Preparation: A meticulous process was undertaken to extract potential H-bonding atom pairs from the Cambridge Structural Database (CSD). This involved searching for specific substructures ([ONS, ONS] and [ONS-H]), defining criteria for H-bond formation (distance and angle constraints), and filtering out malformed or ambiguous cases.
  • Hierarchical Rule Generation: DMAX was used to learn rules from this prepared data. The system builds rules in a hierarchical manner, starting with general rules and progressively refining them into more specific ones. This allows for a nuanced classification of H-bonding scenarios.
  • Background Knowledge Integration: A key innovation is the use of "background knowledge" (BK) to enhance the ILP system's learning capability. They tested three levels of BK:
  • BK-Level-0: Utilized a set of 95 elementary molecular substructures (functional groups, rings) and their frequencies. This level is accessible to non-ILP methods.
  • BK-Level-1: Incorporated relational information between moieties and the candidate donor/acceptor, enabling the system to consider the local structural context. This level requires ILP.
  • BK-Level-2: Added HB-specific concepts, such as conjugated paths and distances to functional groups, further enriching the context.
  • Expert System Development: The learned rules were compiled into an expert system. This system can then predict H-bonding probabilities for new, unseen molecules by navigating the rule hierarchy.

Datasets

The primary dataset used was the Cambridge Structural Database (CSD), specifically version 5.24 (November 2002). The data preparation steps resulted in approximately 62,781 labeled "ideal liquid state" atom pairs, which were then split into:

  • Training Set: 47,086 atom pairs (75%) for rule derivation.
  • Validation Set: 15,695 atom pairs (25%) for evaluating the performance of the learned rules.

The paper does not introduce a new benchmark dataset but rather utilizes a well-established one for a novel application.

Key Insights and Ideas

  • ILP's Power for Contextual Rules: The study powerfully demonstrates that ILP, by its nature, can learn rules that consider the local structural environment of potential H-bond donors and acceptors. This is a significant advantage over traditional methods that rely on global molecular properties, which treat all atom pairs within a molecule similarly.
  • Hierarchy of Rules for Nuance: The hierarchical rule generation approach allows for a graded prediction of H-bonding. Instead of a binary classification, the system provides scores, enabling a ranking of confidence and a more detailed understanding of why a particular prediction is made.
  • Value of Background Knowledge: The performance gains observed with increasing levels of background knowledge (especially the relational information in BK-Level-1) highlight the importance of providing the ILP system with relevant contextual information. This suggests that richer, more structured background knowledge can unlock more sophisticated predictive capabilities.
  • Intramolecular vs. Intermolecular Complexity: The results suggest that predicting intramolecular H-bonds is inherently easier than predicting intermolecular ones, as evidenced by the greater performance improvements seen with HB-specific background knowledge (BK-Level-2) for intramolecular predictions.

Key Results

  • Significant Performance Improvement: The ILP-based expert system consistently outperformed a trivial "ideal liquid state" baseline model across all tested background knowledge levels.
  • ILP Outperforms Non-ILP: Moving from BK-Level-0 (non-ILP accessible) to BK-Level-1 (ILP required) led to substantial improvements in prediction quality for both intra- and intermolecular H-bonds. For example, some cases that were poorly predicted at BK-Level-0 achieved 100% scores at BK-Level-1.
  • BK-Level-2 Enhances Intramolecular Prediction: While BK-Level-1 provided a significant boost, BK-Level-2 further improved the prediction of intramolecular H-bonds, demonstrating the value of HB-specific contextual information.
  • Accurate Classification of "No HB" Cases: The system was effective in identifying cases where H-bonding is unlikely, often assigning higher "no HB" scores to examples that were previously misclassified.

Limitations and Future Work

  • Reliance on 2D Information: While a strength, the reliance solely on 2D structural information might limit the prediction of H-bonding in highly complex 3D scenarios where subtle spatial arrangements are paramount.
  • Computational Cost: ILP can be computationally intensive, especially with large datasets and complex background knowledge.
  • Future Work: The authors suggest that this ILP approach could be integrated into an "artificial scientist" framework, enabling machines to autonomously generate hypotheses, design experiments, and refine knowledge bases. This points towards a future where computational tools play an even more active role in scientific discovery.

Summary updated

Abstract

In the domain of crystal engineering, various schemes have been proposed for the classification of hydrogen bonding (H-bonding) patterns observed in 3D crystal structures. In this study, the aim is to complement these schemes with rules that predict H-bonding in crystals from 2D structural information only. Modern computational power and the advances in inductive logic programming (ILP) can now provide computational chemistry with the opportunity for extracting structure-specific rules from large databases that can be incorporated into expert systems. ILP technology is here applied to H-bonding in crystals to develop a self-extracting expert system utilizing data in the Cambridge Structural Database of small molecule crystal structures. A clear increase in performance was observed when the ILP system DMax was allowed to refer to the local structural environment of the possible H-bond donor/acceptor pairs. This ability distinguishes ILP from more traditional approaches that build rules on the basis of global molecular properties.
BibTeX
@inproceedings{ando2006discovering,
  author = {Howard Y. Ando and Luc De Raedt and Walter Luyten and Elke Van Craenenbroeck and Henk Vandecasteele and Luc Van Meervelt},
  title = {{Discovering H-Bonding Rules in Crystals with Inductive Logic Programming}},
  booktitle = {Molecular},
  year = {2006},
  doi = {10.1021/mp060034z},
}
Figure