Journal of Chemical Information and Modeling 最新文献

Drug discovery and medicinal chemistry efforts are increasingly influenced by machine learning (ML), with compound property prediction as a central application. ML models have demonstrated strong performance in predicting various compound properties from chemical structure. However, these models can exhibit varying levels of prediction error, making uncertainty quantification (UQ) essential for informed decisions. Standard UQ metrics include the distance to the molecules in the training set and prediction variance, obtained through methods such as model ensembles or Bayesian modeling. Although several UQ methodologies have been developed in recent years, no single approach consistently outperformed others. Herein, we present a comprehensive benchmark of UQ strategies for ML-based prediction of absorption, distribution, metabolism, and excretion (ADME) properties, using both in-house and public data sets. We employed the recently introduced UNIQUE (UNcertaInty QUantification bEnchmarking) framework and evaluated UQ method performance under data shifts. Our findings indicate data-based UQ metrics (e.g., chemical distance), and model-based UQ metrics (e.g., predicted value and variance) may capture complementary aspects of uncertainty. Their combination through error models, designed to predict the original ML model’s error, yielded higher-quality uncertainty estimates. These error models emerged as a promising strategy for enhancing UQ, showing robustness in under various degrees and types of data shift. Taken together, our work highlights the potential of combining diverse UQ metrics and error modeling to improve reliability in molecular property prediction. By establishing standardized evaluation setups and assessing UQ under data shifts, we provide a foundation for future UQ method development and benchmarking in the field.

Pancreatic cancer remains a formidable health challenge due to its late-stage diagnosis and limited therapeutic options, underscoring the need for novel targets and modalities. Our previous work revealed that the natural product, BE-43547A₂, could effectively inhibit the progression of pancreatic cancer by the covalent binding to eukaryotic translation elongation factor 1 α 1 (eEF1A1) at Cys234 (C234). Considering the critical role in protein synthesis and the association with pancreatic cancer progression, eEF1A1 is a novel promising target for pancreatic cancer. However, the rational drug design methods for eEF1A1 are extremely lacking. Herein, using microsecond-scale molecular dynamics (MD) simulations, we identify a suitable eEF1A1 conformation for structure-based virtual screening (SBVS) by targeting the residue of C234. Through a tailored SBVS pipeline, we identified AKOS-04 as a novel small-molecule covalent inhibitor with nanomolar-level potency (IC₅₀ = 28.5 ± 2.86 nM in the PATU8988T cell line). Notably, cellular thermal shift assays (CETSA), with the treatments of dithiothreitol (DTT) and iodoacetamide (IAM), confirmed the covalent Cys-involved interaction of AKOS-04 and eEF1A1. Further structural modification validated the critical contribution of a double bond in the acrylamide group of AKOS-04 for its covalent binding with eEF1A1, manifested by the abolished inhibitory activity of compound 9 with the changed single bond in the acrylamide group. MST experiments confirmed direct binding of the compounds to eEF1A1 protein. AKOS-04 exhibited the strongest binding among the tested compounds, consistent with effective covalent target. Finally, MD simulations and pair-interaction energy analyses highlighted Lys84, Arg218, and Glu230 of eEF1A1 as key residues for driving its binding interactions to AKOS-04. These results reveal that AKOS-04, screened by SBVS against C234 of eEF1A1, represents a promising lead for eEF1A1-targeted pancreatic cancer therapy, highlighting the power of computational approaches in covalent drug discovery.

Peptides featuring synthetic modifications, such as noncanonical amino acids, backbone modifications, cyclic structures, and polymer units have become central to modern drug design due to their enhanced stability and functional diversity. However, current machine learning (ML) approaches are restricted by challenges associated with transforming peptide sequences into atom-level representations, leading ML efforts to focus largely on datasets containing linear peptides comprised of standard residues. Here, we present PepFoundry, a Python package that handles peptide sequences beyond canonical amino acids and linear topologies by using SMILES strings in the CHUCKLES format. PepFoundry generates atom-mapped RDKit molecule objects, enabling the extraction of atom-level features, such as Morgan fingerprints and graph representations. We demonstrate its utility by processing a dataset of peptide sequences containing noncanonical amino acids and generating atomic level features for downstream property prediction. We show that atomic-level representations of peptides containing noncanonical amino acids consistently outperform sequence-level representations, regardless of model type. We additionally explore the representation of noncanonical peptides through latent space visualization and show that models with atomic-level information can effectively learn relationships between analogous sequences of l-peptides, d-peptides, and peptoids. This framework allows for the flexible incorporation of new amino acid chemistries, enabling existing ML methods to be straightforwardly applied to datasets of peptides containing nonstandard features. It also facilitates the rapid construction of customized peptide libraries and provides a scalable platform to accelerate ML-driven peptide discovery and optimization.

DNA barcoding is a powerful tool for exploring biodiversity, and DNA language models have significantly facilitated its construction and identification. However, since DNA barcodes come from a specific region of mitochondrial DNA and there are structural differences between DNA barcodes and reference genomes used to train existing DNA language models, it is difficult to directly apply the existing DNA language models to the DNA barcoding task. To address this, this paper introduces DNACSE (DNA Contrastive Learning for Sequence Embeddings), an unsupervised noise-contrastive learning framework designed to fine-tune the DNA language foundation model while enhancing the distribution of the embedding space. The results demonstrate that DNACSE outperforms the direct usage of DNA language models in DNA barcoding-related tasks. Specifically, in fine-tuning and linear probe tasks, it achieves accuracy rates of 99.17 and 98.31%, respectively, surpassing the current state-of-the-art BarcodeBERT by 6.44 and 6.44%. In zero-shot clustering tasks, it raises the adjusted mutual information (AMI) score to 92.25%, an improvement of 8.36%. In addition, zero-shot benchmarking and genomic benchmarking tests are evaluated, indicating that DNACSE enhances the performance of DNA language models in generalized genomic tasks. In summary, DNACSE has demonstrated excellent performance in DNA barcode species classification by making full use of multispecies information and DNA barcode information, providing a feasible way to further explore and protect biodiversity. The code repository is available at https://github.com/Kavicy/DNACSE.

While contiguous subsequences of hydrophobic residues are essential to protein structure and function, as in the hydrophobic core and transmembrane regions, there are no current bioinformatics tools for module identification focused on hydrophobicity. To fill this gap, we created the blobulator toolkit for detecting, visualizing, and characterizing hydrophobic modules in protein sequences. This toolkit uses our previously developed algorithm, blobulation, which was critical in both interpreting intraprotein contacts in a series of intrinsically disordered protein simulations (Lohia et al., 2019) and defining the “local context” around disease-associated mutations across the human proteome (Lohia et al., 2022). The blobulator toolkit provides accessible, interactive, and scalable implementations of blobulation. These are available via a webtool, a visual molecular dynamics (VMD) plugin, and a command line interface. We highlight use cases for visualization, interaction analysis, and modular annotation through three example applications: a globular protein, two orthologous membrane proteins, and an intrinsically disordered protein. The blobulator webtool can be found at www.blobulator.branniganlab.org, and the source code with pip installable command line tool, as well as the VMD plugin with installation instructions, can be found on GitHub at www.GitHub.com/BranniganLab/blobulator.

This study presents a quantum mechanical (QM)-based workflow utilizing semiempirical methods for estimating ligand binding poses in protein–ligand complexes, focusing on kinases of pharmaceutical relevance: CDK2, CK2, p38α, and CAMK1DA. The protocol integrates xTB-based docking with PM6-D3H4X rescoring within a QM/MM framework, aiming to enhance the structural accuracy over traditional docking methods. Drug-like and fragment-like ligands were tested across different receptor structures to assess robustness for CDK2 and CK2. Results show that for drug-like ligands the QM approach reproduces experimental poses. However, performance is less consistent for the fragment-like ligands we have considered, perhaps due to structural ambiguity and weak binding interactions limiting accuracy.

Predicting cytochrome P450 (CYP450) ligand binding is critical in early-stage drug discovery as CYP450-mediated metabolism profoundly influences drug efficacy, safety, and adverse reaction risks. However, experimental determination of CYP450-ligand interactions remains resource- and time-intensive, underscoring the need for robust computational alternatives. While ligand-based methods are commonly employed, they often fail to fully account for structural intricacies governing protein–ligand interactions. To address this gap, we developed a hybrid machine learning framework integrating ligand descriptors, protein descriptors, and protein–ligand interaction descriptors that include molecular docking-derived parameters, rescoring function components from multiple algorithms, and structural interaction fingerprints (SIFt). Evaluated on CYP1A2 and CYP17A1 isoforms, our model demonstrated superior predictive accuracy in cross-validation compared with stand-alone molecular docking and ligand-based approaches. Furthermore, benchmarking against state-of-the-art tools (SwissADME and ADMETlab 3.0) revealed enhanced performance in binding prediction. This work establishes a versatile framework for advancing computational tools to prioritize CYP450 binding assessments during drug discovery.

Reduction potentials of redox-active molecules and materials are essential descriptors of their performance as catalysts, antioxidants, electrode materials, etc. For a given species, its practical applications often span a range of solvent environments, which profoundly impact its redox properties. In this work, we present a message passing graph neural network architecture with a Set Transformer readout trained on ca. 20,000 reduction potentials of chemically diverse closed- and open-shell organic redox-active molecules (the “ReSolved” data set), computed using a rigorously benchmarked density functional theory procedure. The predictor model affords high accuracy with mean absolute errors of ca. 0.2 eV and is uniquely able to generalize to previously unseen solvents. We couple this architecture with an evolutionary algorithm to inverse-design synthetically accessible candidate molecules with target reduction potentials for several battery-related practical applications.

The accurate prediction of drug-induced side effects remains a significant challenge in pharmaceutical development, particularly in early development, as drug programs often fail due to unforeseen adverse reactions. Conventional approaches, such as preclinical animal testing and in vitro assays, are limited by high costs, ethical concerns, and reduced translatability to human biology. Predictive algorithms, including protein–protein interaction network models, have emerged as a computational approach with the potential to predict adverse drug effects, but these models often suffer from limited performance, specifically underprediction. Further, the documented drug–protein interactions are inconsistently reported, affecting our ability to select data for predicting drug-induced side effects or building more performant models. We integrated drug-binding targets from six sources: DrugBank, ChEMBL, PubChem, Search Tool for Interacting Chemicals, Therapeutic Target Database, and PocketFEATURE into our existing platform, PathFX, to understand their impact on the prediction of drug side effects. We observed unique drug–target interactions and target-associated protein classes and functions across sources. Integrating new drug targets predicted previously unrecognized side effects and revealed a trade-off between sensitivity and specificity. Sensitivity generally improved using large exploratory databases or the union of all targets, at the cost of reduced specificity. Databases with smaller numbers of curated targets or structurally predicted targets improved specificity. This quantitative analysis lays the foundation for improvement of drug-side effect prediction, where sophisticated machine learning approaches may better leverage large exploratory databases when balanced and performant analysis is required, and smaller, curated data sources could be integrated with simple but explainable platforms, like PathFX, for hypothesis generation.

Generative models such as diffusion-based approaches have transformed de novo drug design by enabling rapid generation of novel molecular structures in both 2D and 3D formats. However, accurate reconstruction of chemical bonds, especially from distorted geometries produced by generative models, remains a critical challenge. Here, we present YuelBond, a multimodal graph neural network framework for robust bond reconstruction across three key scenarios: (i) recovery of bonds from accurate 3D atomic coordinates, (ii) reconstruction of chemically valid bonds in crude de novo generated compounds (CDGs) with perturbed geometries, and (iii) reassignment of bond orders in 2D topological graphs. YuelBond outperforms traditional rule-based methods such as RDKit, achieving 98.4% F1 score on standard 3D structures and maintaining strong performance (92.7% F1 score) on distorted CDGs, even when RDKit fails in most cases. Our results demonstrate that YuelBond enables accurate and reliable bond reconstruction from imperfect molecular data, bridging a critical gap in generative drug discovery pipelines.