Journal of Cheminformatics

AO_SCPLOWBSTRACTC_SCPLOWComputationally-aided design of novel molecules has the potential to accelerate drug discovery. Several recent generative models aimed to create new molecules for specific protein targets. However, a rate limiting step in drug development is molecule optimization, which can take several years due to the challenge of optimizing multiple molecular properties at once. We developed a method to solve a specific molecular optimization problem in silico: expanding a small, fragment-like starting molecule bound to a protein pocket into a larger molecule that matches that physiochemical properties of known drugs. Using data-efficient E(3) equivariant based neural networks and a 3D atomic point cloud representation, our model learns how to attach new molecular fragments to a growing structure by recognizing realistic intermediates generated en route to a final ligand. This approach always generates chemically valid molecules and incorporates all relevant 3D spatial information from the protein pocket. This framework produces promising molecules as assessed by multiple properties that address binding affinity, ease of synthesis, and solubility. Overall, we demonstrate the feasibility of 3D molecular structure expansion conditioned on protein pockets while maintaining desirable drug-like physiochemical properties and developed a tool that could accelerate the work of medicinal chemists.

Predicting blood-brain barrier (BBB) penetration remains a central challenge in CNS drug discovery. Existing computational models rely on physicochemical descriptors and are blind to active transport biology - the efflux pumps and carrier proteins that dominate drug exclusion at the BBB in vivo. We present BBB-Nuke, a modular prediction pipeline that integrates physicochemical scoring with explicit efflux transporter substrate modeling. The system computes ten molecular descriptors, predicts ionization state via a graph convolutional network, scores CNS-MPO desirability, and estimates substrate probability for seven efflux transporters (P-gp/MDR1, BCRP/ABCG2, MRP1, MRP2, MRP4, MATE1, OAT3) using Random Forest classifiers trained on curated ChEMBL bioactivity data. A gradient-boosted classifier trained on 67 features - ten physicochemical, seven efflux transporter probabilities, and fifty fingerprint-derived principal components - achieves an area under the receiver operating characteristic curve (AUROC) of 0.933 {+/-} 0.006 under five-fold cross-validation on 9,262 labeled compounds, and 0.810 on a fully held-out benchmark of 470 clinically validated compounds. In head-to-head comparisons, BBB-Nuke outperforms CNS-MPO, LightBBB, ADMETlab 2.0, and BBB-Score on both cross-validation and external test sets. We apply the pipeline to screen over one billion commercially available compounds from the Enamine REAL library and PubChem, identifying enriched regions of BBB-penetrant chemical space and characterizing the structural features that distinguish permeable from excluded molecules. BBB-Nuke is freely available as a Python package, REST API, and Model Context Protocol server.

In recent years, several chemical language models have been developed, inspired by the success of protein language models and advancements in natural language processing. In this study, we explore whether pre-training a chemical language model on billion-scale compound datasets, such as Enamine and ZINC20, can lead to improved compound representation in the drug space. We compare the learned representations of these models with the de facto standard compound representation, and evaluate their potential application in drug discovery and development by benchmarking them on biophysics, physiology, and physical chemistry datasets. Our findings suggest that the conventional masked language modeling approach on these extensive pre-training datasets is insufficient in enhancing compound representations. This highlights the need for additional physicochemical inductive bias in the modeling beyond scaling the dataset size.

SummaryRecently, novel representation learning algorithms have shown potential for predicting molecular properties. However, unified frameworks have not yet emerged for fairly measuring algorithmic progress, and experimental procedures of different representation models often lack rigorousness and are hardly reproducible. Herein, we have developed MolRep by unifying 16 state-of-the-art models across 4 popular molecular representations for application and comparison. Furthermore, we ran more than 12.5 million experiments to optimize hyperparameters for each method on 12 common benchmark data sets. As a result, CMPNN achieves the best results ranked the 1st in 5 out of 12 tasks with an average rank of 1.75. Relatively, ECC has good performance in classification tasks and MAT good for regression (both ranked 1st for 3 tasks) with an average rank of 2.71 and 2.6, respectively. AvailabilityThe source code is available at: https://github.com/biomed-AI/MolRep Supplementary informationSupplementary data are available online.

BackgroundPredicting molecular activity against protein targets is difficult because of the paucity of experimental data. Approaches like multitask modeling and collaborative filtering seek to improve model accuracy by leveraging results from multiple targets, but are limited because different compounds are measured with different assays, leading to sparse data matrices. Profile-QSAR (pQSAR) 2.0 addresses this problem by fitting a series of partial least squares models for each target, using as features the predictions from single-task models on the remaining targets. This method has been shown to produce better results than single task and multitask models. However, the factors determining the success of pQSAR 2.0 have as yet not been characterized. In this paper we examine the experimental conditions that lead to better pQSAR models. We limit the amount of data available to the method by retraining with decreasing amounts of data and explore the models ability to generalize to compounds that have never been assayed. Finally, we look at the properties of training data needed to demonstrate pQSAR improvement. ResultsWe apply pQSAR 2.0 on a collection of GPCR and safety targets collected from Drug Target Commons, ExcapeDB, and ChEMBL. We found that pQSAR improved models on 34 of the 149 assays selected. In the other 115 assays, single task random forests offered better performance. There are many factors that contribute to an increase in performance, but the main factor is compound assay coverage. The pQSAR model improves when more compounds are measured in multiple assays. ConclusionIt is necessary to consider the available data before applying pQSAR. Successful pQSAR models require a profile made of correlated targets that share compounds with other assays. This technique is best used when experimental data is available as random forest regressors often do not generalize well enough for virtual drug search applications.

MotivationProtein-Protein Interactions (PPIs) are crucial in biological processes and disease mechanisms, underscoring the importance of discovering PPI inhibitors in drug development. Machine learning can expedite this discovery process. Although machine learning techniques for predicting general compound inhibition are available, we are not aware of any that accurately forecast the inhibitory effect of a compound on a specific protein complex, utilizing inputs from both the compound and the protein complex. MethodsWe present the first targeted machine learning based predictor of small molecule based inhibition of protein complexes. Our proposed graph neural network integrates the structure of a protein complex, its protein-protein binding site or interface features and a compounds SMILES representation to predict the potential of the given compound to inhibit the interaction between proteins in the given complex in a targeted manner. ResultsValidated on the 2p2i-DB-v2 database, encompassing 714 inhibitors across 23 complexes with over 12,000 instances, our model achieves superior predictive accuracy (cross-validation AUC-ROC of 0.86), outperforming established kernel methods and pre-trained neural networks. We further tested the predictive performance of our model on two independent external datasets - one collected from recent publications and another consisting of putative inhibitors of the SARS-CoV-2-Spike and Human-ACE2 protein complex with AUC-ROCs of 0.82 and 0.78, respectively. Our targeted predictor introduces a novel approach for PPI inhibitor discovery, laying foundational work for future advancements in addressing this complex and previously unexplored prediction challenge. AvailabilityCode/supplementary material available: https://github.com/adibayaseen/PPI-Inhibitors

Chemical toxicity assessment is critical for drug development and environmental safety. Computational models have emerged as a promising alternative to animal testing and now play a significant role in efficiently evaluating new chemicals. To address the urgent need for providing user-friendly machine learning tools in computational toxicology, we developed ToxiVerse, a public web-based platform. It provides curated toxicity datasets, automatic chemical bioprofiling, and a predictive modeling interface designed for researchers who lack programming expertise. The platform comprises three integrated modules: (i) the Bioprofiler module, which provides chemical descriptors by combining chemical-bioactivity data from PubChem assay with a machine learning-based data gap-filling procedure; (ii) the Database module, which hosts around 50,000 curated unique chemicals covering diverse toxicity endpoints; and (iii) the Cheminformatics module, which allows users to upload their own datasets, use datasets from ToxiVerse, or retrieve existing data from PubChem; perform chemical curation; and automatically generate Quantitative Structure-Activity Relationship (QSAR) models to predict chemicals of interest. ToxiVerse enables researchers to carry out bioprofiling, access curated toxicity datasets, and evaluate chemical toxicity through machine learning-based modeling and prediction. The platform is supported by sample files and a detailed tutorial, and it is freely accessible at www.toxiverse.com. GRAPHICAL ABSTRACT O_FIG O_LINKSMALLFIG WIDTH=200 HEIGHT=80 SRC="FIGDIR/small/708255v1_ufig1.gif" ALT="Figure 1"> View larger version (22K): org.highwire.dtl.DTLVardef@d92764org.highwire.dtl.DTLVardef@a92f4aorg.highwire.dtl.DTLVardef@15fa39corg.highwire.dtl.DTLVardef@1ee89bc_HPS_FORMAT_FIGEXP M_FIG C_FIG

Databases are indispensable in biological and biomedical research, hosting vast amounts of structured and unstructured data, facilitating the organization, retrieval, and analysis of complex data. Database access, however, remains a manual, tedious, and sometimes overwhelming, task. We investigate in this study the current state of using pre-trained, search-enabled LLMs for data retrieval from biological databases. Equipped with internet search and code generation capabilities, LLMs promise to streamline database access through natural language, expedite search and knowledge retrieval, and provide coherent analytical summaries. As an example database, we focus on evaluating a current search-enabled LLMs (GPT-4o) for retrieval from the PubChem database, a flagship, heavily used database that plays a critical role in biological and biomedical research. As PubChem is an open archival repository, it provides a well-documented programmatic interface that can be exploited through LLM code generation capabilities. We evaluate retrieval tasks for eight common PubChem access protocols that were previously documented. The tasks include identifying interacting genes and proteins, finding drug-like compounds based on structural similarity, retrieving bioactivity data, and locating stereoisomers and isotopomers. We develop a methodology for adopting the protocols into an LLM-prompt, where we supplement the prompt with additional context through iterative prompt refinement as needed. To further evaluate the LLM capabilities, we instruct the LLM to perform the retrieval with and without using programmatic access. We compare the results (referred to as gold and silver answers) when using these retrieval modalities with two traditional retrieval baselines that include running the manual search steps for each reference protocol through the PubChem database web interface, and through the provided PUG (Power-User Gateway) programmatic access. We quantitatively and qualitatively summarize our results, showing that generating programmatic access is more likely to yield the correct answers. We highlight the value and limitations of using current search-based LLMs for database retrieval. We also provide guidance for the future development that can improve the accuracy and reliability of search-based LLMs.

The discovery of novel small molecules is challenging because of the vastness of chemical space and the complexity of protein-ligand interactions, leading to low success rates and time-consuming workflows. Here, we present LLMsFold, a computational framework that combines Large Language Models (LLMs) and biophysical foundation tools to design and validate new small molecules targeting pathogenic proteins. The pipeline starts by identifying viable binding pockets on a target protein through geometry-based pocket detection. A 70-billion-parameter transformer model from the LlaMA family then generates candidate molecules as SMILES strings under prompt constraints that enforce drug-likeness. Each molecule is evaluated by Boltz-2, a diffusion-based model for protein-ligand co-folding that predicts bound 3D structure and binding affinity. Promising candidates are iteratively optimized through a reinforcement learning loop that prioritizes high predicted affinity and synthetic accessibility. We demonstrate the approach on two challenging targets: ACVR1 (Activin A Receptor Type 1), implicated in fibrodysplasia ossificans progressiva (FOP), and CD19, a surface antigen expressed on most B-cell lymphoma and leukemia cells. Top candidates show strong in silico binding predictions and favorable drug-like profiles. All code and models are made available to support reproducibility and further development.

A complex web of intermolecular interactions defines and regulates biological processes. Understanding this web has been particularly challenging because of the sheer number of actors in biological systems: [~]104 proteins in a typical human cell offer a plausible 108 interactions. This number grows rapidly if we consider metabolites, drugs, nutrients, and other biological molecules. The relative strength of interactions also critically affects these biological processes. However, the small and often incomplete datasets (103-104 protein-ligand interactions) traditionally used for binding affinity predictions limit the ability to capture the full complexity of these interactions. To overcome this challenge, we developed Yuel 2, a novel neural network-based approach that leverages transfer learning to address the limitations of small datasets. Yuel 2 is pre-trained on a large-scale dataset to learn intricate structural features and then fine-tuned on specialized datasets like PDBbind to enhance the predictive accuracy and robustness. We show that Yuel 2 predicts multiple binding affinity metrics - Kd, Ki, and IC50 - between proteins and small molecules, offering a comprehensive representation of molecular interactions crucial for drug design and development.

Recent advancements in large language models (LLMs) offer new opportunities for automating the manual curation of biochemical reaction databases from scientific literature. In this study, we present an integrated pipeline that enhances LLM-based extraction of enzymatic reactions with machine learning and cheminformatics-informed validation. Using BRENDA-linked PubMed articles, we evaluate GPT-4s ability to extract reactions and infer missing chemical entities in textual descriptions of enzymatic reactions. Extracted reactions are converted to SMILES and InChI notations before being encoded into molecular fingerprint similarity scores and atom mapping metrics. These cheminformatics metrics are then used to train machine learning classifiers that validate GPT extractions. We employ a Positive-Unlabeled learning approach with synthetic invalid reactions to train various classifiers and assess model performances. The best classifier is then benchmarked on GPT extractions. Our findings show that GPT can accurately infer incomplete reactions and cheminformatics tools can serve as effective predictors of reaction validity. This work demonstrates a scalable framework for automated and reliable curation of enzymatic reaction databases, highlighting the potential of combining LLMs with cheminformatics and machine learning for reliable scientific knowledge extraction. Author SummaryCurating databases of biochemical reactions is a time-consuming and manual task, yet it plays a vital role in advancing research in biology and chemistry. Many scientific articles describe important enzymatic reactions, but often do so in incomplete ways--such as mentioning only the starting molecule or the enzyme, and leaving out the rest. In this work, we explore how recent advancements in artificial intelligence, specifically large language models like GPT, can help extract such information automatically from scientific literature. We show that these models can not only find reactions in text, but also infer missing parts of reactions based on the surrounding context. To make sure these inferred reactions are chemically plausible, we use computational chemistry tools that analyze the structure of the molecules involved. We then train a machine learning model to help us automatically detect which reactions are likely to be valid. This combination of tools offers a new way to speed up and improve how biochemical knowledge is extracted from the growing body of scientific literature. Our study suggests that this kind of automation could help scientists keep biological databases up to date and reduce the burden of manual data entry.

Drug-target affinity (DTA) prediction is a key task in drug discovery, enabling the estimation of the interaction strength between candidate compounds and biological targets. However, current models rely on connectivity-based molecular representations and do not explicitly account for the spatial organization, also known as stereochemistry. This limitation becomes evident when considering chirality, where a drug can exist as enantiomers, i.e., molecules that share the same atoms and bonds but differ in their three-dimensional arrangement. Despite their chemical similarity, they can interact differently with the same target, leading to variations in binding affinity and biological activity. In this paper, we propose a stereochemistry-aware DTA prediction framework that incorporates this information into molecular representations. Drug representations are learned from chemical structure using a directed-bond message passing graph neural network that captures enantiomers configurations, while protein targets are represented through sequence-based embeddings. Experiments on the Davis dataset demonstrate that our model can improve affinity prediction. Importantly, a case study on a manually curated dataset of enantiomers with different biological action shows that the model is able to distinguish the affinities in the two forms consistent with their experimentally observed biological activity. These findings support the relevance of stereochemistry-aware molecular representation for more accurate and chemically faithful DTA prediction.

Drug development projects are getting increasingly more expensive while their success rate is stagnating. Safety issues attributed to off-target binding represent a major reason for the failure of new drugs. Besides desired on-target binding, small molecules may interact with off-targets, triggering adverse effects. Therefore, the development of novel methods for early recognition of such issues that are resource-efficient and cost-effective becomes vital. Here, we introduce PanScreen, an online platform for the automated assessment of off-target liabilities. PanScreen combines structure-based modeling techniques with state-of-the-art deep learning methods to not only predict accurate binding affinities but also give insight into potential modes of action. We show that the predictions are approaching experimental accuracy found in public datasets and that the same technology can also be used for other research areas, such as drug repurposing. Such fast and inexpensive methods allow researchers to test not only drug candidates, but all small molecules that might come into contact with a human organism for potential safety concerns very early in the development process. PanScreen is publicly available at www.panscreen.ch.

Prediction of a compounds toxicity is a key step toward realizing animal-free testing of chemical compounds. Recent advances have yielded significant progress in computational toxicity prediction, including machine learning methods that utilize chemical fingerprints and deep learning-based latent representations. However, challenges remain, primarily due to the lack of clean training datasets and the inconsistent model performance. To address these challenges, we curated a comprehensive dataset of aquatic toxicity from seven data sources, which contains 50,603 records for 5,889 compounds across 2,285 different species, much larger than similar datasets used in previous studies. We also developed tox-learn, a Python library featuring tools for automated dataset cleaning, machine learning methods and performance evaluation. The library places special emphasis on avoiding overestimation of prediction accuracy caused by improper train-test data splitting. Based on this toolbox, we benchmarked various predictive models using different train-test splitting strategies on the curated dataset. Our results showed that the choice of machine learning method, molecular fingerprint, and train-test splitting strategy all significantly affect performance. We demonstrated that incorporating species information generally improved predictions, although the degree of improvement depended on how this information was represented. In addition, we developed a new 3D structure-based deep-learning model, 3DMol-Tox, which achieves regression accuracy comparable to the best 2D-structure based model (GPBoost) while exhibiting consistently higher within-one-bin (W1B) classification accuracy. Finally, we analyzed the impact of different train-test splitting strategies and provide recommendations based on our benchmarking, such as using structure-aware splitting to mitigate information leakage, a common issue that inflates reported model performance.

Ab initio prediction of side effect frequencies is important for assessing the risk-benefit profile of drugs and for identifying potential adverse effects early in development. A key challenge is chirality: many drugs exist as enantiomers, pairs of molecules with the same atoms and bond connectivity but different three-dimensional arrangements. Although chemically similar, enantiomers can interact differently with biological targets and therefore exhibit distinct efficacy and adverse-effect profiles. Here we introduce F2S (Features to Signatures), a method to predict the frequencies of drug side effects while explicitly accounting for chirality. Drug representations are learned directly from chemical structure using a directed-bond message-passing graph neural network that captures stereochemical configurations. Side effect representations are derived from curated textual descriptions encoded with a frozen PubMedBERT model. Side effect frequencies are predicted from the dot product between drug and side effect signatures together with biases for drugs and side effects. We evaluated F2S extensively across multiple settings, including cold-start and warm-start prediction, prospective evaluation, and scenarios controlling for chemical similarity between training and test drugs. Across these evaluations, F2S achieves performance comparable to state-of-the-art methods for general side-effect frequency prediction while producing fewer false positives and substantially improves the prediction of frequency differences between enantiomer pairs. Finally, F2S learns compact 10-dimensional signatures that support interpretability: drug signatures reflect therapeutic class and shared targets, side-effect signatures capture phenotype similarity, and the learned bias terms correlate with the popularity of drugs and side effects.

Targeting cancer metabolism has emerged as a promising therapeutic strategy, particularly through the inhibition of Lactate Dehydrogenase A (LDHA), a key enzyme that supports the Warburg effect in tumor cells. In this study, we present a comprehensive and fully reproducible machine learning (ML) and artificial intelligence (AI)-driven pipeline for the discovery of small-molecule LDHA inhibitors. By integrating bioactivity datasets from ChEMBL and BindingDB, along with natural products from COCONUT and AI-generated compounds from a ChemGPT-based molecular language model, we constructed a diverse and chemically rich screening library. Molecular descriptors were computed using Mordred, followed by feature selection, dataset balancing using SMOTE, and extensive model benchmarking across 11 classifiers. LightGBM was selected as the top-performing model with an AUC of 0.97. SHAP analysis provided model interpretability, revealing key molecular features influencing LDHA inhibition. Additionally, we trained ChemGPT on LDHA-specific SMILES in SELFIES format to generate 1,000 novel molecules, of which over 100 passed stringent drug-likeness, toxicity, and solubility filters. A subset exhibited high LDHA inhibition probabilities (>0.90) and structural novelty. This work highlights the potential of combining predictive modeling and generative chemistry for accelerating the early stages of cancer drug discovery and provides an open-source platform for continued development and validation.

BackgroundLarge data sets of compound activity lend themselves to building predictive models based on compound and target structure. The simplest representation of structure is via vectorisation. Compound fingerprint vectorisation has been successfully employed in predicting compound activity classes. ResultsA vector representation of a protein-compound pair based on a concatenation of the compound fingerprint and the protein triplet vector has been used to train random forest and neural network models on multiple datasets of protein-compound interaction together with compound associated transcription and activity profiles. Results for compound-target predictability are comparable with more complex published methodologies. ConclusionA simple intuitive representation of a protein-compound pair can be employed in a variety of machine learning models to gain a predictive handle on the activity of compounds for which there is no activity data. It is hoped that this transparent approach will prove sufficiently portable and simple to implement that drug discovery will be opened up to the wider research community.

MotivationMolecular representation learning is central to computational drug discovery. However, most existing models rely on single-modality inputs, such as molecular sequences or graphs, which capture only limited aspects of molecular behaviour. Yet unifying these modalities with complementary resources such as textual descriptions and biological interaction networks into a coherent multimodal framework remains non-trivial, hindering more informative and biologically grounded representations. ResultsWe introduce SELFormerMM, a multimodal molecular representation learning framework that integrates SELFIES notations with structural graphs, textual descriptions, and knowledge graph- derived biological interaction data. By aligning these heterogeneous views, SELFormerMM effectively captures complementary signals that unimodal approaches often overlook. Our performance evaluation has revealed that SELFormerMM outperforms structure-, sequence-, and knowledge-based models on multiple molecular property prediction tasks. Ablation analyses further indicate that effective cross-modal alignment and modality coverage improve the models ability to exploit complementary information. Overall, integrating SELFIES with structural, textual, and biological context enables richer molecular representations and provides a promising framework for hypothesis-driven drug discovery. AvailabilitySELFormerMM is available as a programmatic tool, together with datasets, pretrained models, and precomputed embeddings at https://github.com/HUBioDataLab/SELFormerMM. Contacttuncadogan@gmail.com

Proteins interact through their interfaces, and dysfunction of protein-protein interactions (PPIs) has been associated with various diseases. Therefore, investigating the properties of the drug-modulated PPIs and interface-targeting drugs is critical. Here, we present a curated large dataset for drug-like molecules in protein interfaces. We further present DiPPI (Drugs in Protein-Protein Interfaces), a two-module website to facilitate the search for such molecules and their properties by exploiting our dataset in drug repurposing studies. In the interface module of the website, we extracted several properties of interfaces, such as amino acid properties, hotspots, evolutionary conservation of drug-binding amino acids, and post-translational modifications of these residues. On the drug-like molecule side, we curated a list of drug-like small molecules and FDA-approved drugs from various databases and extracted those that bind to the interfaces. We further clustered the drugs based on their molecular fingerprints to confine the search for an alternative drug to a smaller space. Drug properties, including Lipinskis rules and various molecular descriptors, are also calculated and made available on the website to guide the selection of drug molecules. Our dataset contains 534,203 interfaces for 98,632 proteins, of which 55,135 are detected to bind to a drug-like molecule. 2,214 drug-like molecules are deposited on our website, among which 335 are FDA-approved. DiPPI provides users with an easy-to-follow scheme for drug repurposing studies through its well-curated and clustered interface and drug data; and is freely available at http://interactome.ku.edu.tr:8501.

Natural compounds constitute a rich resource of potential small-molecule therapeutics. While experimental access to this resource is limited due to its vast diversity and difficulties in systematic purification, computational assessment of structural similarity with known therapeutic molecules offers a scalable approach. Here, we assessed functional similarity between natural compounds and approved drugs by combining multiple chemical similarity metrics and physicochemical properties through a random forest model. As a training set, we used pair-wise similarity between 1410 drugs in terms of their shared protein targets. The resulting model featured high performance metrics (matthews correlation coefficient of 0.81, and balanced accuracy of 0.91) suggesting that it well-captured the structure-activity relation. The model was then used to predict protein targets of circa 11k natural compounds by comparing them with the drugs. This revealed therapeutic potential of several natural compounds, including those with support from previously published sources as well as those hitherto unexplored. We experimentally validated one of the predicted links activities, viz., Cox-1 inhibition by 5-methoxysalicylic acid, a molecule commonly found in tea, herbs and spices. In contrast, another natural compound, 4-isopropylbenzoic acid, which showed a higher similarity when considering the most weighted similarity metric but was not picked by the random forest model, did not inhibit Cox-1. Our results demonstrate the utility of a machine-learning approach combining multiple chemical features for uncovering protein binding potential of natural compounds.