Molecular Informatics最新文献

Gaussian process regression (GPR) is a nonparametric probabilistic model capable of computing not only the predicted mean but also the predicted standard deviation, which represents the confidence level of predictions. It offers great flexibility as it can be non-linearized by designing the kernel function, made robust against outliers by altering the likelihood function, and extended to classification models. Recently, models combining deep learning with GPR, such as Deep Kernel Learning GPR, have been proposed and reported to achieve higher accuracy than GPR. However, due to its nonparametric nature, GPR is challenging to interpret. While Explainable AI (XAI) methods like LIME or kernel SHAP can interpret the predicted mean, interpreting the predicted standard deviation remains difficult. In this study, we propose a novel method to interpret the prediction of GPR by evaluating the importance of explanatory variables. We have incorporated the GPR model with the Integrated Gradients (IG) method to assess the contribution of each feature to the prediction. By evaluating the standard deviation of the posterior distribution, we show that the IG approach provides a detailed decomposition of the predictive uncertainty, attributing it to the uncertainty in individual feature contributions. This methodology not only highlights the variables that are most influential in the prediction but also provides insights into the reliability of the model by quantifying the uncertainty associated with each feature. Through this, we can obtain a deeper understanding of the model's behavior and foster trust in its predictions, especially in domains where interpretability is as crucial as accuracy.

The presence of Activity Cliffs (ACs) has been known to represent a challenge for QSAR modeling. With its high data dependency, Machine Learning QSAR models will be directly influenced by the activity landscape. We propose several extended similarity and extended SALI methods to study the implications of ACs distribution on the training and test sets on the model's errors. Ununiform ACs and chemical space distribution tend to lead to worse models than the proposed uniform methods. ML modeling on AC-rich sets needs to be analyzed case-by-case. Proposed methods can be used as a tool to study the datasets, but as far as generalization, random splitting was the better-performing data splitting alternative overall.

Dimensionality reduction is an important exploratory data analysis method that allows high-dimensional data to be represented in a human-interpretable lower-dimensional space. It is extensively applied in the analysis of chemical libraries, where chemical structure data - represented as high-dimensional feature vectors-are transformed into 2D or 3D chemical space maps. In this paper, commonly used dimensionality reduction techniques - Principal Component Analysis (PCA), t-Distributed Stochastic Neighbor Embedding (t-SNE), Uniform Manifold Approximation and Projection (UMAP), and Generative Topographic Mapping (GTM) - are evaluated in terms of neighborhood preservation and visualization capability of sets of small molecules from the ChEMBL database.

Background: Effective molecular feature representation is crucial for drug property prediction. Recent years have seen increased attention on graph neural networks (GNNs) that are pre-trained using self-supervised learning techniques, aiming to overcome the scarcity of labeled data in molecular property prediction. Traditional GNNs in self-supervised molecular property prediction typically perform a single masking operation on the nodes and edges of the input molecular graph, masking only local information and insufficient for thorough self-supervised training.

Method: Hence, we propose a model for molecular property prediction based on generative double-masking self-supervised learning, termed as GDMol. This integrates generative learning into the self-supervised learning framework for latent representation, and applies a second round of masking to these latent representations, enabling the model to better capture global information and semantic knowledge of the molecules for a richer, more informative representation, thereby achieving more accurate and robust molecular property prediction.

Results: Our experiments on 5 datasets demonstrated superior performance of GDMol in predicting molecular properties across different domains. Moreover, we used the masking operation to traverse through the gradient changes of each node, the magnitude and sign of which reflect the positive and negative contribution respectively of the local structure in the molecule to the prediction outcome. This in-depth interpretative analysis not only enhances the model's interpretability, but also provides more targeted insights and direction for optimizing drug molecules.

Conclusions: In summary, this research offers novel insights on improving molecular property prediction tasks, and paves the way for further research on the application of generative learning and self-supervised learning in the field of chemistry.

Interpretability and reliability of deep learning models are important for computer-based drug discovery. Aiming to understand feature perception by such a model, we investigate a graph neural network for affinity prediction of protein-ligand complexes. We assess a latent representation of ligand binding sites and investigate underlying geometric structure in this latent space and its relation to protein function. We introduce an automated computational pipeline for dimensionality reduction, clustering, hypothesis testing, and visualization of latent space. The results indicate that the learned protein latent space is inherently structured and not randomly distributed. Several of the identified protein binding site clusters in latent space correspond to functional protein families. Ligand size was found to be a determinant of cluster geometry. The computational pipeline proved applicable to latent space analysis and interpretation and can be adapted to work for different datasets and deep learning models.

Visualization and analysis of large chemical reaction networks become rather challenging when conventional graph-based approaches are used. As an alternative, we propose to use the chemical cartography ("chemography") approach, describing the data distribution on a 2-dimensional map. Here, the Generative Topographic Mapping (GTM) algorithm - an advanced chemography approach - has been applied to visualize the reaction path network of a simplified Wilkinson's catalyst-catalyzed hydrogenation containing some 10⁵ structures generated with the help of the Artificial Force Induced Reaction (AFIR) method using either Density Functional Theory or Neural Network Potential (NNP) for potential energy surface calculations. Using new atoms permutation invariant 3D descriptors for structure encoding, we've demonstrated that GTM possesses the abilities to cluster structures that share the same 2D representation, to visualize potential energy surface, to provide an insight on the reaction path exploration as a function of time and to compare reaction path networks obtained with different methods of energy assessment.

Recent advances in machine learning have significantly impacted molecular design, notably the molecular generation method combining the chemical variational autoencoder (VAE) with Gaussian mixture regression (GMR). In this method, a mathematical model is constructed with X as the latent variable of the molecule and Y as the target properties and activities. Through direct inverse analysis of this model, it is possible to generate molecules with the desired target properties. However, this approach outputs many strings that do not follow the simplified molecular input line entry system grammar and generates unrealistic chemical structures in which the properties and activity do not satisfy the target values. In this study, we focus on hierarchical VAE using molecular graphs to address these issues. We confirm that the combination of hierarchical VAE and GMR does not generate invalid outputs and returns molecules that simultaneously satisfy multiple target values. Moreover, we use this method to identify several molecules that are predicted to exhibit activity against drug targets.

Located in plasma membranes, ATP hydrolases are involved in several dynamic transport processes, helping to control the movement of ions across cell membranes. ATP hydrolase acts as a transport protein, converting energy from ATP hydrolysis into transport molecules against their concentration gradients. In addition to energy metabolism and active transport, ATP hydrolase is essential for maintaining cellular homeostasis and cell function. This study focused on the domain architecture model of P-type ATPases, which participate in the reaction cycles of ATP hydrolysis carried out by membrane transport systems - Na+, K+-ATPase and Ca2+, Mg2+-ATPase. Targeted modulation of Na+, K+-ATPase and Ca2+, Mg2+-ATPase by unnatural drugs is of greatest interest due to the lack of known effectors. This new discovery presents a convenient model based on our recent experimental studies of the membrane structures and myocytes of the uterine smooth muscle, the myometrium. This current study strongly supports the fact that nanosized calix[4]arenes functionalised on the upper rings of the macrocycle with biologically active phosphonic acid fragments can serve as selective and potent inhibitors of cation-transporting electroenzymes. This is how we discovered that calix[4]arene of methylenebisphosphonic acid C-97 and calix[4]arene of bis-aminophosphonic acid C-107 selectively and effectively (I0.5 <100 nM) inhibit the activity of Mg2+, ATP-dependent electrogenic Na+ K+ plasma membrane pump. As drug discovery in the field of Mg2+-ATPase inhibitors is uncharted territory, basic research holds the key to explaining and predicting the mechanism of interaction and action of different classes of compounds. In light of the presented results, new calix[4]arene compounds can be used as potent inhibitors of Mg2+, ATP-dependent electrogenic ion pumps.

Predicting Protein-Ligand Binding Affinity (PLBA) is pivotal in drug development, as accurate estimations of PLBA expedite the identification of promising drug candidates for specific targets, thereby accelerating the drug discovery process. Despite substantial advancements in PLBA prediction, developing an efficient and more accurate method remains non-trivial. Unlike previous computer-aid PLBA studies which primarily using ligand SMILES and protein sequences represented as strings, this research introduces a Deep Learning-based method, the Enhanced Representation Learning on Protein-Ligand Graph Structured data for Binding Affinity Prediction (ERL-ProLiGraph). The unique aspect of this method is the use of graph representations for both proteins and ligands, intending to learn structural information continued from both to enhance the accuracy of PLBA predictions. In these graphs, nodes represent atomic structures, while edges depict chemical bonds and spatial relationship. The proposed model, leveraging deep-learning algorithms, effectively learns to correlate these graphical representations with binding affinities. This graph-based representations approach enhances the model's ability to capture the complex molecular interactions critical in PLBA. This work represents a promising advancement in computational techniques for protein-ligand binding prediction, offering a potential path toward more efficient and accurate predictions in drug development. Comparative analysis indicates that the proposed ERL-ProLiGraph outperforms previous models, showcasing notable efficacy and providing a more suitable approach for accurate PLBA predictions.

The advent of high-performance virtual screening techniques nowadays allows drug designers to explore ultra-large sets of candidate compounds in search of molecules predicted to have desired properties. However, the success of such an endeavor heavily relies on the pertinence (drug-likeness and, foremost, chemical feasibility) of these candidates, or otherwise, virtual screening will return valueless "hits", by the garbage in/garbage out principle. The huge popularity of the judiciously enumerated Enamine REAL Space is clear proof of the strength of this Big Data trend in drug discovery. Here we describe a new dataset of make-on-demand compounds called the Freedom space. It follows the principles of Enamine REAL Space and contains highly feasible molecules (synthesis success rate over 75 percent). However, the scaffold and chemography analysis revealed significant differences to both the REAL and biologically annotated compounds from the ChEMBL database. The Freedom Space is a significant extension of the REAL Space and can be utilized for a more comprehensive exploration of the synthetically feasible chemical space in hit finding and hit-to-lead campaigns.