Medicinal-chemistry review is often built from disconnected notebooks, spreadsheets, and slide decks that hide structure-normalization choices and blur the boundary between measured observations and heuristic prioritization. MolScope is an open-source Python workflow toolkit that converts molecular structure collections and optional assay joins into a canonical chemistry table, category-aware summaries, and decision-ready report and picklist artifacts. The workflow makes representation policy explicit, preserves evidence provenance, and packages outputs as portable HTML, Markdown, and CSV bundles. Using frozen example datasets, we show reproducible end-to-end execution, auditable handling of salts, tautomers, stereo ambiguity, charge state, and one bounded round-review example for comparative campaign support.
Deep neural network (DNN)-based in silico models show great promise in predicting the properties and bioactivities of novel compounds, including small molecules. Among traditional approaches, structure-based drug design (SBDD) remains a fundamental approach for drug discovery using molecular docking, scoring functions, and molecular dynamics simulations. However, these approaches are often constrained by limited flexibility, resolution, and generalizability. Geometric deep learning (GDL) offers a transformative alternative by enabling models to learn directly from non-Euclidean molecular representations, such as graphs, point clouds, and meshes, capturing critical 3D spatial relationships inherent to protein-ligand interactions. This review highlights the theoretical underpinnings and practical applications of GDL in small-molecule drug discovery, focusing on tasks including binding affinity prediction, virtual screening, de novo molecule generation, pose prediction, ADMET profiling, and protein flexibility modeling. We explore key GDL architectures, graph neural networks, SE(3)-equivariant networks, 3D convolutional neural networks, point cloud models, and geometric transformers, and assess their performance across various drug discovery benchmarks. The integration of geometry-aware AI models with experimental and computational workflows was also highlighted for its potential to streamline hit-to-lead optimization and advance rational drug design. Despite remarkable progress, the field faces challenges including limited high-quality 3D structural datasets, protein flexibility representation, and the interpretability of deep models. Addressing these issues through hybrid modeling approaches, multi-resolution learning, and self-supervised training could further elevate GDL's impact. Ultimately, GDL stands at the frontier of AI-enhanced pharmaceutical innovation, offering unprecedented precision, efficiency, and insight in the pursuit of next-generation therapeutics.
PairMap2 is a web application for testing intermediate-molecule insertion in relative binding free energy calculations. Transformations between structurally distant ligands can have insufficient overlap in free-energy space, leading to poor convergence and unreliable predictions. PairMap2 reimplements the PairMap intermediate-insertion workflow as a browser-based tool and supports two-molecule path analysis and multiligand perturbation-map construction. The interface visualizes perturbation networks, generated intermediates, and maximum common substructure-based atom mappings. In nine benchmark cases from the original PairMap study, PairMap2 completed faster in eight cases; for an additional low-similarity transformation (lead optimization mapper (LOMAP) similarity score 0.0111), it achieved an approximately 19.6-fold speedup. PairMap2 enables computational chemists to assess intermediate insertion before free energy perturbation FEB calculations without setting up a local execution environment.
Linkers play a central role in many areas of medicinal chemistry, including proximity inducers, small-molecule conjugates, and DNA-encoded libraries. However, little is known about the accessibility of molecules to linker attachment when bound to proteins. Here, we analyze linker accessibility across protein-ligand complexes in cocrystal structures. A computational workflow was developed to evaluate the linkerability of modifiable positions on molecules based on solvent accessibility, local steric space for introduction of a linker atom, and the geometry of solvent-directed escape paths approximated as conical frustums. Analysis of 8,228 protein-ligand cocrystal structures with 131 431 modifiable positions shows that approximately 22% of positions can accommodate linkers without significant geometric restriction. Limited linkerability of positions influences DEL data and may confound efforts to use such data for lead prediction.
Drafting chemical compound patents, particularly those involving Markush structures, is a complex task often hindered by manual workflows that are time-consuming and error-prone and may result in incomplete protection. To address these challenges, we introduce SpaceExpander, an open-source tool designed to automate the generation of Markush patent claims with improved accuracy and efficiency. The tool systematically extracts molecular scaffolds, identifies R-groups, and constructs comprehensive Markush structures to expand patent coverage. To enhance accessibility and usability, we provide a publicly available web server for online usage, along with standalone executables compatible with Windows, macOS, and Linux to support offline access.
Venom-derived peptides are structurally diverse bioactive scaffolds with growing relevance for computational discovery and molecular design. However, their organization within chemical space remains poorly characterized in integrated frameworks. Here, we present a curated, structure-informed venom peptide atlas supported by an integrated descriptor-based framework combining sequence- and structure-derived features across 3,423 peptides. High-confidence AlphaFold models were integrated with physicochemical and structural descriptors to generate a unified molecular representation. Multivariate analyses reveal that venom peptides occupy constrained and interpretable regions of chemical space shaped by disulfide-mediated stabilization, structural compactness, and electrostatic-hydrophobic balance. Functional classes show clear separation, whereas taxonomic origin explains less variance, supporting convergent design principles. This atlas provides a reference resource for comparative analyses, scaffold prioritization, and hypothesis generation in peptide science.
The discovery of novel compounds for protein targets is an important step in the drug discovery process. Target-specific molecules can be designed based on protein structures, but molecule prediction based on protein sequences alone remains a challenge. We hypothesize that including information on noninteracting pairs can improve the specificity and generalizability of molecule prediction models. To this end, we introduce ConGen, a sequence-conditioned framework for targeted molecule generation from protein sequence alone. To our knowledge, it is the first model that combines contrastive protein-molecule representation learning with interacting and noninteracting molecules and latent optimization (LO) in a shared pretrained space, addressing the limitations of interacting-only training. ConGen follows two stages: (i) contrastive learning (CL) maps proteins and molecules into a joint space where interacting pairs are pulled together and noninteracting pairs are pushed apart and (ii) LO starts from random molecule latents, moves them toward the target protein's embedding, and decodes a local neighborhood to propose candidates. This approach enables the generation of molecules specifically designed for a given target by efficiently exploiting the structure of the shared latent space. Our results show that ConGen achieves performance comparable to EncDecLM, a state-of-the-art sequence-based model for targeted drug generation. An ablation study further demonstrates the contributions of both the CL and LO stages.
De novo molecular generation remains a central challenge in drug discovery due to the vastness of chemical space and the difficulty of generating molecules that are simultaneously valid, diverse, and property-aware. While graph-based generative models have improved structural validity over sequence-based approaches, many existing methods suffer from mode collapse, limited controllability, and unstable reinforcement learning dynamics. In this work, we propose NovMolG-GAN, a graph-based generative adversarial framework that integrates graph attention mechanisms, a learnable reward network, mode-seeking regularization, and proximal policy optimization-based reinforcement learning to address these limitations. The proposed model enables stable training while explicitly encouraging molecular validity, novelty, diversity, and drug-likeness. Experiments on the ChEMBL-35 dataset demonstrate that NovMolG-GAN achieves high validity (99.6%), high novelty (99.4%) strong uniqueness (88.5%), and competitive quantitative drug-likeness scores under unconstrained generation. Furthermore, a conditional generation ablation incorporating synthetic accessibility as an additional reward illustrates the controllability of the framework and its ability to steer generation toward synthesis-feasible regions of chemical space. These results highlight the potential of NovMolG-GAN as a flexible and extensible foundation for goal-directed molecular design.
Learning expressive representations for complex, size-varied molecular graphs remains a fundamental challenge in toxic molecular property prediction and regression. The inherent nonlinearity of atomic interactions, together with intricate structural dependencies and latent high-order structural characteristics, makes accurate graph-based learning particularly difficult. Although recent deep learning (DL)/graph neural network (GNN) approaches have leveraged message passing, attention mechanisms, and graph transformer architectures, these techniques still suffer from limited nonlinear expressiveness, insufficient modeling of high-order structural information, and a lack of explicit uncertainty handling. To address these limitations, we propose a novel AKAGTL model, which is an attention-driven Kolmogorov-Arnold network (KAN)-based graph transformer framework for toxic molecular graph embedding and regression learning. Unlike existing approaches that rely on linear or shallow nonlinear transformations, our proposed AKAGTL model introduces a structured KAN-based functional transformation to explicitly model complex high-order atomic interactions within an attention-driven graph transformer backbone. In addition, high-order structural representations are systematically incorporated to complement structural encoding, while a Gaussian neuro-fuzzy fusion mechanism enables uncertainty-aware aggregation of heterogeneous feature spaces. The proposed framework is evaluated on multiple molecular toxicity benchmarks under a unified experimental protocol with repeated runs. Comprehensive experiments within benchmark molecular graph datasets demonstrate that our AKAGTL model can consistently improve regression accuracy compared to representative GNN and graph transformer baselines. These findings suggest that jointly modeling nonlinear functional interactions, structural dependencies, and uncertainty-aware fusion provides a more expressive and robust solution for toxic molecular graph embedding and regression learning.
Efficient exploration of chemical space is an essential component of modern generative drug design. Herein, we introduce ChemBang, a computational engine that grows small molecules based on chemical transformations extracted by matched molecular pair analysis of all structures available in catalogues of synthesized molecules. Each chemical transformation is mapped onto its associated atomic environment defined as the substructure within a three-atom radius from the transformation site. Unsupervised chemical evolution is then performed in cycles by systematically applying chemical transformations to all exposed atomic environments present in a seed structure. Multiple physicochemical properties and substructural alerts are incorporated to effectively guide the generation of drug-like synthetically accessible molecules. As a use case, the generation of the Erdafitinib structure from any of its three ring systems (pyrazole, benzene and quinoxaline), and the evolution of the property distributions from all molecules generated in each cycle, are discussed in detail. The ability to explore the chemical space of pharmaceutical relevance is shown by successfully generating the exact chemical structure of 95.3% of all 2,809 small-molecule ATC drugs from their constituting fragments.
Drug-target interaction represents a critical focus area in computational drug discovery and pharmaceutical research. However, the process of identifying and analyzing these interactions is often resource-intensive, requiring extensive experimentation to evaluate the binding relationships between numerous drugs and their respective targets. This complexity is further compounded by the fact that a drug can inhibit multiple targets, and a target may also bind to various drugs. To address these issues, advanced deep learning models have been introduced as promising tools, offering the ability to accurately predict binding affinity and other bioactivity values to distinguish between potential drug-target interactions. The proposed model, named the Augmented Deep Meta Ensemble for Drug-Target Interaction (ADME-DTI), leverages multiple descriptors and fingerprint representations to extract meaningful insights from drug and protein data. These submodels from each representation are then combined with a deep learning architecture along with the metadata of the drug and target entries. The proposed approach has proved competitive performance against state-of-the-art models across diverse datasets and evaluation metrics, as evidenced by key metrics, namely r m 2 $r_m^2$ , concordance index, and mean squared error. Specifically, the model achieved mean squared error values of 0.186 (Davis), 0.118 (Kiba), 0.134 (DTC), 0.262 (Metz), 0.300 (ToxCast), and 0.791 (STITCH), all of which are publicly available benchmark datasets commonly used in drug-target interaction prediction tasks. The results highlight the model's effectiveness in reducing prediction errors and improving accuracy. Previous drug-target interaction research often relied on limited, non-diverse datasets, reducing generalizability. Few studies addressed drug-target interaction prediction as a regression task. Our novel deep learning metamodel integrates multiple models and representations, surpassing benchmarks and delivering greater prediction accuracy across varied datasets and evaluation metrics.
Chemical space analysis is extensively used in different chemistry areas, ranging from the study of natural products to drug discovery projects. Its versatility stems from the ability to integrate continuous properties with molecular representations. This data is used to generate visualizations through dimensionality reduction algorithms. Constellation Plots have been proposed as a general approach to the visual representation of chemical space by encoding structural similarity, scaffold contents, frequency, and continuous properties into a single coordinate-based map. Thus, Constellation Plots provide a high-density visual representation of the chemical space of compound datasets with complex relations. Despite the versatility of Constellation Plots, there remains a significant lack of intuitive, user-friendly, or low-code protocols to automate the generation of these plots for non-computational experts. Herein, we present an interactive and automated scaffold-based Constellation Plot workflow developed within the open-source platform KNIME, facilitating chemical space visualization and analysis. To illustrate the application of the workflow, we used a dataset of 5,211 compounds that inhibit Tau protein, a key therapeutic target for Alzheimer's disease. The KNIME workflow is a general resource that can be used to analyze virtually any data set annotated with a property, including biological activity. The workflow is freely available at: https://github.com/Daniphantom99/KNIME_Constellation_plots.
Skin permeability is a critical factor in pharmaceuticals, cosmetics, and occupational safety. Experimental determination of skin permeability coefficients (Kp) is time-intensive and resource-intensive, highlighting the importance of computational predictions. This study presents a quantitative structure-property relationship (QSPR) model developed using the recently published SkinPiX dataset and the HuskinDB skin permeability database, comprising 209 curated compounds with associated Kp values and metadata. The model performance was assessed on three new experimental data points. The datasets and models are freely available, providing a valuable tool to enhance decision-making and support the development of safer and more effective products.
Membrane technologies for the separation of gases, such as CO2/CH4 mixtures, have attracted attention because of their high energy efficiency. Polyimides are considered promising membrane materials for CO2 separation, and there is a growing demand for materials with even higher performance. In the screening of candidate materials, it is essential to consider not only separation performance but also solubility and polymerizability during the synthesis process. Low solubility or polymerizability can inhibit membrane fabrication and the evaluation of separation performance, potentially leading to wasted resources and effort. In this study, we developed machine learning models to predict the solubility and polymerizability of polyimides. Mixture features derived from molecular descriptors of multiple monomers and mixing ratios were used as inputs for the classification models. The models were then applied to novel candidates, and their effectiveness was validated experimentally.
Drug-target interaction (DTI) prediction is a fundamental task in the field of drug discovery, with direct implications for the identification of novel therapeutic candidates and the repositioning of existing drugs. However, the practical application of DTI prediction remains hindered by several persistent challenges, including data scarcity, the absence of reliable negative samples, and limited model generalization across diverse biological contexts. Addressing these limitations is crucial for developing robust and generalizable predictive frameworks. In this study, we present LapGAT, a semi-supervised framework combining graph-enhanced Laplacian regularized least squares (LapRLS) with a graph attention network (GAT) to address these issues. In the upstream stage, LapRLS fuses multiple drug-target similarity matrices, applies Laplacian regularization, and selects top- and bottom-scoring pairs as high-confidence positive and negative samples. In the downstream stage, a multilayer GAT learns from these pseudo-labeled interactions, capturing both local graph structures and nonlinear dependencies. We validate LapGAT on four target categories (enzymes, G-protein-coupled receptors (GPCRs), ion channels, nuclear receptors), demonstrating robust performance in computational and experimental validation. Molecular docking (AutoDock Vina) confirms the physical plausibility of top-ranked predictions, with binding affinities ranging from -4.5 to -7.3 kcal/mol. Literature-based validation achieves accuracies of 81.3% (enzymes), 100% (GPCRs), 71.4% (ion channels), and 88.9% (nuclear receptors). This work offers a scalable and flexible computational tool for accelerating drug discovery efforts, with the potential for broad applicability across various therapeutic domains.
The growing number and size of DNA-encoded libraries (DELs), together with the vast space of possible DEL designs, demand interpretable and scalable criteria for selecting which libraries to construct and screen against a given target. An ideal target-focused DEL shows both strong similarity with an active reference compound collection and high intra-DEL diversity. Chemography with Generative Topographic Mapping (GTM) was shown to be a promising approach for selecting DELs, offering both intuitive visualization and fast quantitative analysis scalable to thousands of DEL designs. This is achieved by defining each library by a "stand-alone" vector, the comparison of which precludes costly pairwise inter-molecular similarity calculations. However, the extent to which such "stand-alone" (SA) approaches in general, and GTM-derived SA metrics in particular, recover DELs that are reference-proximal and chemically diverse as evaluated by conventional compound pair-matching (CP) metrics in the initial descriptor space remains insufficiently characterized. In this article, the comparative analysis of the Morgan count fingerprint-based chemical-library similarity versus GTM-derived metrics, using 100 diverse DEL subsets and a reference set of compounds tested against cyclin-dependent kinase 2 (CDK2) from ChEMBL, was performed. GTM-based SA metrics provide robust approximations for "gold standard" molecular descriptor space CP metrics for DEL selection: Spearman rank correlations fall in the 0.6-0.7 range. Our results demonstrate that GTM helps to identify DELs that best span the reference space according to same "gold standard" molecular descriptor space metrics: SA GTM-driven rankings of libraries achieve enrichment factors at 5% (EF5%) of 4-12 (in terms of finding "gold standard" top libraries within the 5% best ranked by GTM)-always picking 2 out of the top 3 libraries. The accompanying two-dimensional landscapes make intra- and interlibrary diversity visually accessible, supporting rapid, interpretable screening of alternative DEL designs. Collectively, these results position GTM as an efficient tool for chemical-library similarity assessment and target-focused DEL selection.
Human 5-lipoxygenase (5-LOX) plays a crucial role in the biosynthesis of leukotrienes (LTs). Therefore, 5-LOX inhibitors are designed as effective agents for the treatment of several diseases such as asthma, cardiovascular disorders, allergies, and cancer. Insights into crystal structures of several 5-LOX isoforms have revealed that this protein adopts two different conformations (open/closed) through modulation of its Hα2 and arched helix regions, which are conditioned by the presence or absence of ligand in the active site; moreover, these structures are incomplete in regions critical for ligand binding. To advance the design of 5-LOX inhibitors, we developed a computational procedure to reconstruct the first full-length open conformation structure of 5-LOX complexed with chelating inhibitor within the active site. Dynamic simulations and protein model validation confirmed the quality of our model, which was subsequently used for docking analyses and culminated in the development of a structure-based pharmacophore model. These computational studies might constitute powerful tools for rationally designing and identifying novel 5-LOX iron chelator inhibitors.
Foundation and multimodal models are rapidly becoming a core methodology in molecular informatics, particularly for drug discovery, by leveraging large-scale pretraining across sequences, graphs, 3D structures, and text. This mini-review provides practical guidance on when these models help, how to choose representations and data, and how to design pretraining and adaptation pipelines for real-world use. We clarify what qualifies as a foundation model in chemistry; compare chemical language models, graph-based architectures, and 3D equivariant networks; review multimodal strategies that connect molecules with proteins, pockets, and natural language; and summarize diffusion-based generative modeling. We also emphasize rigorous evaluation, discussing realistic splitting protocols, distribution shift, activity cliffs, uncertainty calibration, and conformal prediction in the context of widely used benchmarks.
Rapidly progressive Alzheimer's disease (rpAD) is a rare but severe form of Alzheimer's disease characterized by accelerated cognitive decline and limited therapeutic options. Conventional anti-amyloid-β interventions have shown little success due to poor target specificity, neurotoxicity, and lack of efficacy, underscoring the need for novel therapeutic strategies. This study aimed to identify and prioritize molecular targets associated with rpAD by investigating the protein interactome of amyloid-β (Aβ42) using integrative computational approaches. Functional enrichment, protein-protein interaction network analysis, and community clustering revealed that rpAD-specific Aβ42 interactors were predominantly involved in mitochondrial bioenergetics, redox regulation, and cytoskeletal stability, pathways central to neuronal survival and synaptic function. Molecular docking identified fumarate hydratase, carbonyl reductase 1, and the F-actin capping protein as high-affinity interactors of Aβ42, linking these proteins to energy failure, oxidative stress, and synaptic dysfunction. Virtual screening of a therapeutic drug library against fumarate hydratase identified several compounds with strong binding affinities, among which quinestrol, estradiol benzoate, norethindrone, tamibarotene, drospirenone, and ketanserin emerged as lead candidates. Pharmacokinetic profiling, including ADMET modeling, confirmed their blood-brain barrier permeability and drug-likeness, supporting their potential as central nervous system active agents. Together, this work highlights key molecular targets in rpAD and proposes repurposed, pharmacologically diverse compounds with multitarget neuroprotective potential. By utilizing in silico analysis, the study provides a rational framework for target discovery and drug prioritization in rpAD, offering a foundation for future experimental validation and the development of translational research.
