One class classification as a practical approach for accelerating π–π co-crystal discovery

Aikaterini Vriza, Angelos B. Canaj, Rebecca Vismara, Laurence J. Kershaw Cook, Troy D. Manning, Michael W. Gaultois, P.a. Wood, Vitaliy Kurlin, Neil G. Berry, Matthew S. Dyer, Matthew J. RosseinskyChemical 2020 · DOI

TL;DR

This paper introduces a novel machine learning approach, "one-class classification," to accelerate the discovery of new organic co-crystals, specifically focusing on weakly bound π-π stacking systems. By training on existing co-crystal data and identifying patterns of "normal" behavior, the method effectively predicts promising new combinations from a vast chemical space, overcoming the common challenge of missing negative experimental data.

What Problem Does This Paper Address?

Discovering new materials with desired properties is a major goal in materials science. Machine learning (ML) offers powerful tools for this, but its effectiveness is often hampered by imbalanced datasets. In materials science, this imbalance is particularly acute because negative results (failed synthesis attempts) are rarely published, leading to datasets that are heavily skewed towards successful outcomes. This "positive examples only" problem makes it difficult for traditional ML models to accurately predict new materials, especially for classes that are inherently rare, like weakly bound polyaromatic hydrocarbon (PAH) co-crystals. These co-crystals are of interest for electronic applications but are challenging to design due to the weak nature of their primary interactions (π-π stacking).

Methods

The authors propose and evaluate a "one-class classification" (OCC) methodology. Unlike traditional ML that requires both positive and negative examples, OCC learns the characteristics of a single, well-defined class (in this case, known PAH co-crystals) and then identifies deviations from this norm as potential outliers or new members of the class.

Two main workflows were explored:

  1. Traditional OCC: This involved extensive feature engineering to reduce the dimensionality of molecular descriptors, followed by applying eight different OCC algorithms (e.g., Gaussian Mixture Models, Isolation Forest, One-Class SVM) and combining their outputs into an ensemble for more robust predictions.
  2. Deep OCC: This approach utilized a deep learning architecture (DeepSVDD) that automates feature learning, allowing it to handle higher-dimensional data and capture complex interactions between molecular descriptors without manual feature engineering.

Both workflows were trained on known co-crystal data and then used to score potential new co-crystal combinations.

Datasets

  • Cambridge Structural Database (CSD): Used to create the "labelled dataset" of 1,722 known PAH co-crystal molecular combinations.
  • ZINC15 Database: Used to generate the "unlabelled dataset" of 21,736 possible molecular combinations of PAHs, representing the vast chemical space to be explored.

Key Insights and Ideas

  • One-Class Classification as a Solution to Data Imbalance: The core idea is that by focusing on what constitutes a "normal" co-crystal, the model can effectively identify potential new candidates even without explicit examples of what doesn't form a co-crystal. This is crucial for rare material classes.
  • Deep Learning Outperforms Traditional Methods: The deep learning approach (DeepSVDD) demonstrated superior performance, achieving better separation between known and potential co-crystals and requiring less manual effort in feature engineering.
  • Feature Importance Reveals Co-crystallization Drivers: SHAP (SHapley Additive explanations) analysis revealed that molecular size, shape, and polarity are important factors, but their interplay is complex. Notably, for PAH co-crystals, dissimilar property values between co-formers can also lead to successful co-crystallization, a trend less observed in other co-crystal classes.
  • PAH Co-crystals Have Unique Rules: The study suggests that PAH co-crystals deviate from general co-crystallization rules, particularly those driven by hydrogen bonding. Heteroatoms (like oxygen and nitrogen) and specific structural motifs play a more significant role in stabilizing these π-π systems.
  • Pareto Optimization for Experimental Prioritization: Combining ML scores with similarity to known functional molecules (like TCNQ) via Pareto optimization effectively guides experimental efforts towards the most promising candidates.

Key Results

  • The developed OCC framework successfully predicted new co-crystal candidates.
  • Two novel co-crystals were synthesized and experimentally verified: pyrene-6H-benzo[c]chromen-6-one (1) and pyrene-9,10-dicyanoanthracene (2).
  • Co-crystal 1 was found to have a 1:2 stoichiometry (pyrene:co-former), which was predicted by the model to be a ratio higher than 1:1.
  • Co-crystal 2 was found to have a 1:1 stoichiometry, as predicted.
  • The deep learning model achieved higher accuracy and better separation of data compared to traditional OCC methods.
  • The study identified pyrene as a particularly versatile co-former, appearing frequently in high-scoring predicted pairs.

Limitations and Future Work

  • Limited Scope of Descriptors: The current ML model does not consider all possible chemical factors, such as the solubility of co-formers, which can influence experimental outcomes.
  • Need for Rigorous Screening: For predicted candidates that did not readily form co-crystals under initial conditions (e.g., pyrene-triphenylene), more rigorous screening of crystallization conditions is suggested.
  • Experimental Validation: While two novel co-crystals were synthesized, the authors acknowledge that six examples (including predicted but not synthesized ones) do not provide statistically significant validation of the model.
  • Incorporating Negative Data: While OCC addresses the lack of negative data, future work could explore methods to incorporate more nuanced negative information or experimental failures to further refine the models.
  • Expanding to Other Material Classes: The success of OCC in this specific domain suggests its potential applicability to other challenging materials discovery problems with imbalanced datasets.

Summary updated

Abstract

The implementation of machine learning models has brought major changes in the decision-making process for materials design. One matter of concern for the data-driven approaches is the lack of negative data from unsuccessful synthetic attempts, which might generate inherently imbalanced datasets. We propose the application of the one-class classification methodology as an effective tool for tackling these limitations on the materials design problems. This is a concept of learning based only on a well-defined class without counter examples. An extensive study on the different one-class classification algorithms is performed until the most appropriate workflow is identified for guiding the discovery of emerging materials belonging to a relatively small class, that being the weakly bound polyaromatic hydrocarbon co-crystals. The two-step approach presented in this study first trains the model using all the known molecular combinations that form this class of co-crystals extracted from the Cambridge Structural Database (1722 molecular combinations), followed by scoring possible yet unknown pairs from the ZINC15 database (21 736 possible molecular combinations). Focusing on the highest-ranking pairs predicted to have higher probability of forming co-crystals, materials discovery can be accelerated by reducing the vast molecular space and directing the synthetic efforts of chemists. Further on, using interpretability techniques a more detailed understanding of the molecular properties causing co-crystallization is sought after. The applicability of the current methodology is demonstrated with the discovery of two novel co-crystals, namely pyrene-6<i>H</i>-benzo[<i>c</i>]chromen-6-one (<b>1</b>) and pyrene-9,10-dicyanoanthracene (<b>2</b>).
BibTeX
@inproceedings{vriza2020one,
  author = {Aikaterini Vriza and Angelos B. Canaj and Rebecca Vismara and Laurence J. Kershaw Cook and Troy D. Manning and Michael W. Gaultois and P.A. Wood and Vitaliy Kurlin and Neil G. Berry and Matthew S. Dyer and Matthew J. Rosseinsky},
  title = {{One class classification as a practical approach for accelerating π–π co-crystal discovery}},
  booktitle = {Chemical},
  year = {2020},
  doi = {10.1039/d0sc04263c},
}
Figure