Representing molecular and materials data for unsupervised machine learning

Predicting Antimicrobial Activity of Conjugated Oligoelectrolyte Molecules via Machine Learning

New antibiotics are needed to battle growing antibiotic resistance, but the development process from hit, to lead, and ultimately to a useful drug takes decades. Although progress in molecular property prediction using machine-learning methods has opened up new pathways for aiding the antibiotics development process, many existing solutions rely on large data sets and finding structural similarities to existing antibiotics. Challenges remain in modeling unconventional antibiotic classes that are drawing increasing research attention. In response, we developed an antimicrobial activity prediction model for conjugated oligoelectrolyte molecules, a new class of antibiotics that lacks extensive prior structure-activity relationship studies. Our approach enables us to predict the minimum inhibitory concentration for E. coli K12, with 21 molecular descriptors selected by recursive elimination from a set of 5305 descriptors. This predictive model achieves an R² of 0.65 with no prior knowledge of the underlying mechanism. We find the molecular representation optimum for the domain is the key to good predictions of antimicrobial activity. In the case of conjugated oligoelectrolytes, a representation reflecting the three-dimensional shape of the molecules is most critical. Although it is demonstrated with a specific example of conjugated oligoelectrolytes, our proposed approach for creating the predictive model can be readily adapted to other novel antibiotic candidate domains.

Using molecular embeddings in QSAR modeling: does it make a difference?

With the consolidation of deep learning in drug discovery, several novel algorithms for learning molecular representations have been proposed. Despite the interest of the community in developing new methods for learning molecular embeddings and their theoretical benefits, comparing molecular embeddings with each other and with traditional representations is not straightforward, which in turn hinders the process of choosing a suitable representation for Quantitative Structure-Activity Relationship (QSAR) modeling. A reason behind this issue is the difficulty of conducting a fair and thorough comparison of the different existing embedding approaches, which requires numerous experiments on various datasets and training scenarios. To close this gap, we reviewed the literature on methods for molecular embeddings and reproduced three unsupervised and two supervised molecular embedding techniques recently proposed in the literature. We compared these five methods concerning their performance in QSAR scenarios using different classification and regression datasets. We also compared these representations to traditional molecular representations, namely molecular descriptors and fingerprints. As opposed to the expected outcome, our experimental setup consisting of over $25 000$ trained models and statistical tests revealed that the predictive performance using molecular embeddings did not significantly surpass that of traditional representations. Although supervised embeddings yielded competitive results compared with those using traditional molecular representations, unsupervised embeddings tended to perform worse than traditional representations. Our results highlight the need for conducting a careful comparison and analysis of the different embedding techniques prior to using them in drug design tasks and motivate a discussion about the potential of molecular embeddings in computer-aided drug design.

Nanoinformatics, and the big challenges for the science of small things

The combination of computational chemistry and computational materials science with machine learning and artificial intelligence provides a powerful way of relating structural features of nanomaterials with functional properties. However, combining these fundamentally different scientific approaches is not as straightforward as it seems. Machine learning methods were developed for large data sets with small numbers of consistent features. Typically nanomaterials data sets are small, with high dimensionality and high variance in the feature space, and suffer from numerous destructive biases. None of the established data science or machine learning methods in widespread use today were devised with (nano)materials data sets in mind, but there are ways to overcome these challenges and use them reliably. In this review we will discuss domain-specific constraints on data-driven nanomaterials design, and explore the differences between nanomaterials simulation and nanoinformatics that can be leveraged for greater impact.

Convolutional Neural Networks for Crystal Material Property Prediction Using Hybrid Orbital-Field Matrix and Magpie Descriptors

Computational prediction of crystal materials properties can help to do large-scale insiliconscreening. Recent studies of material informatics have focused on expert design of multidimensionalinterpretable material descriptors/features. However, successes of deep learning suchas Convolutional Neural Networks (CNN) in image recognition and speech recognition havedemonstrated their automated feature extraction capability to effectively capture the characteristicsof the data and achieve superior prediction performance. Here, we propose CNN-OFM-Magpie, aCNN model with OFM (Orbital-field Matrix) and Magpie descriptors to predict the formationenergy of 4030 crystal material by exploiting the complementarity of two-dimensional OFM featuresand Magpie features. Experiments showed that our method achieves better performance thanconventional regression algorithms such as support vector machines and Random Forest. It is alsobetter than CNN models using only the OFM features, the Magpie features, or the basic one-hotencodings. This demonstrates the advantages of CNN and feature fusion for materials propertyprediction. Finally, we visualized the two-dimensional OFM descriptors and analyzed the featuresextracted by the CNN to obtain greater understanding of the CNN-OFM model.

Predicting archetypal nanoparticle shapes using a combination of thermodynamic theory and machine learning

Machine learning is a useful way of identifying representative or pure nanoparticle shapes as part of a larger ensemble, but its predictive capabilities can be limited when a large dataset of candidate structures must already exist. Ideally one would like to use machine learning to define the ideal dataset for future, more computationally intensive, studies before a significant amount of resources are consumed. In this work we combine an established analytical phenomenological model and statistical machine learning to predict the archetypes and prototypes of a diverse ensemble of 2380 platinum nanoparticle morphologies developed with less than twenty input electronic structure simulations. By parameterising a size- and shape-dependent thermodynamic model, probabilities are assigned to seventeen different shapes between three and thirty nanometres, which together with structural features such as nanoparticle diameter, surface area, sphericity and facet configuration form the basis for archetypal analysis and K-means clustering. Using this approach we rapidly identify six "pure" archetypes and twelve "representative" prototypes that can be used in future computational studies of properties such as catalysis.