An accurate and interpretable deep learning model for environmental properties prediction using hybrid molecular representations

Abstract

Lipophilicity, as quantified by the decimal logarithm of the octanol–water partition coefficient (log KOW), is an essential environmental property. Deep neural networks (DNNs) based quantitative structure–property relationship (QSPR) studies have received more and more attention because of their excellent performance for prediction. However, the black‐box nature of DNNs limits the application range where interpretability is essential. Hence, this study aims to develop an accurate and interpretable deep neural network (AI‐DNN) model for log KOW prediction. A hybrid method of molecular representation was employed to guarantee the accuracy of the proposed AI‐DNN model. The hybrid molecular representations are able to integrate the directed message passing neural networks (D‐MPNNs) learned molecular representations and the fixed molecule‐level features of CDK descriptors, and can capture both the local and the global features of overall molecule. The performance analysis shows that the proposed QSPR model exhibits promising predictive accuracy and discriminative power in the structural isomers and stereoisomers. Moreover, the Monte Carlo Tree Search (MCTS) approach was used to interpret the proposed AI‐DNN model by identifying the molecular substructures contributed to the lipophilicity. This interpretability can be applied to critical fields where there is a high demand for interpretable deep networks, such as green solvent design and drug discovery.

Combining Group-Contribution Concept and Graph Neural Networks Toward Interpretable Molecular Property Models

“…In this work, we rely on the chemical diversity analysis presented earlier and the William plot. The William plots are visual representations of the hat values and are considered a distance-based approach for identifying the applicability domain of QSPR models. ,, The hat values, also referred to as leverages, are proportional to the Hotelling T 2 and to the Mahalanobis distance which corrects for potential colinearity in the descriptors by using the covariance matrix . The leverages ( h i ) for compound i are the diagonal elements of the matrix H denoted hat matrix, which is acquired through eq , where X is the latent representation of the molecule, i.e., the output of the GNN part used as input to the MLP. H = X false( X T X false) − 1 X T …”

“…Plotting the acquired leverages against the standardized residuals results in the William plot, which has also been applied to define the domain of applicability of a wide range of QSPR models. , More details about the underlying theory can be found in the review paper on the domain of applicability of QSPR models …”

“…Zhang et al combined learned features produced from the D-MPNN model by Yang et al with fixed global features calculated using a variety of cheminformatics packages to produce the molecular representation. They then used Monte Carlo Tree Search and by continuously trimming the molecule of substructures identified the substructures responsible for lipophilicity . Recently, a GNN model embedded with domain knowledge was used to predict a set of property data related to physical chemistry such as the lipophilicity and the aqueous solubility of organic compounds .…”

“…Using domain knowledge to identify the substructures, they performed the rank-sum test and local mean to compare the importance of the substructures that could potentially be redesigned to increase inhibition efficiency. Zhang et al combined learned features produced from the D-MPNN model by Yang et al with fixed global features calculated using a variety of cheminformatics packages to produce the molecular representation. They then used Monte Carlo Tree Search and by continuously trimming the molecule of substructures identified the substructures responsible for lipophilicity .…”

Multiscale Screening of Deep Eutectic Solvents for Efficient Extraction of m-Cresol from Model Coal Tar

“…Nevertheless, the UNIFAC model has not been reported in the screening and design of mixtures. Besides, the deep learning model can be used to predict the thermodynamic properties of solvents. − However, the deep learning model is restricted to investigate systems lacking experimental data.…”