The rapid proliferation of per- and polyfluoroalkyl substances (PFASs) necessitates high-throughput tools to assess safety and prevent regrettable substitution. However, conventional toxicological testing is resource-intensive, while standard computational methods often lack the mechanistic transparency required for rational molecular design. This study presents a deep learning framework coupling graph neural networks (GNNs) with transfer learning to predict acute oral rat LD50 toxicity in data-scarce environments. Among the architectures evaluated, GIN with stratified transfer learning (GIN-TL-Strat) achieved the highest predictive correlation (test R[2] = 0.7426). By leveraging knowledge transferred from a broad chemical space, the model demonstrates promising generalization potential to structurally diverse fluorinated compounds. GNNExplainer was integrated as an interpretability mechanism to map prediction logic back to specific atomic substructures, enabling quantitative disentanglement of the contributions of chain length, functional group composition, and structural isomerism to acute toxicity. The results demonstrate that the fluorinated carbon chain length is the primary structural determinant of toxicity in continuous-chain alkyl PFAS. Crucially, attribution magnitude does not linearly dictate toxicity impact: while fluorinated backbones and specific nitrogenous/polar functionalities drive positive toxicity shifts, heavy halogens (I, Br, and Cl) exhibit the lowest mean attribution scores and correlate with reduced predicted toxicity, highlighting a decoupling between model attention and net toxicological outcomes. Structural isomerism between straight-chain and branched PFAS was found to have no statistically meaningful effect on predicted toxicity. This work transitions computational toxicology from a passive screening tool to an active platform for the safety-by-design of next-generation fluorinated alternatives.

Accurate prediction of variants within intrinsically disordered regions (IDRs) is crucial for advancing disease diagnosis and biomedical interpretation. However, the intrinsic lack of stable structural conformations and the high sequence variability of IDRs make it challenging for existing predictors to achieve robust performance in these regions. Here, we introduce DisoPatho, a deep learning framework specifically tailored for predicting disease-associated variants in IDRs. DisoPatho features a novel mutation-centric architecture that utilizes the variant site as an anchor for feature construction and interaction. The core innovation lies in a cross-view adaptive-feature interaction mechanism, which synergistically integrates IDR-specific energy representations with embeddings from protein language models, including xTrimoPGLM and Evolutionary Scale Modeling. This strategy enables the comprehensive capture of evolutionary constraints and physicochemical patterns without requiring explicit structural descriptors, multiple-sequence alignments, or hand-crafted conservation scores. Consequently, DisoPatho exhibits enhanced discriminative power better adapted to the highly flexible nature of IDRs. Comprehensive evaluations across multiple IDR data sets demonstrate that DisoPatho substantially outperforms existing methods. In 5-fold cross-validation, it achieves average AUCs of 0.899 and 0.840, with ACCs of 0.862 and 0.860 on two data sets. Notably, on a highly confounded independent test set where phylogenetic constraints offer limited discriminative signals, DisoPatho yields a 50.2% relative improvement in MCC over AlphaMissense on their respective predictable variants, while achieving broader prediction coverage. In-depth analyses of the prediction results further confirm the effectiveness and stability of the framework in IDR-specific scenarios. The code, data sets, and predictions for DisoPatho are available for academic use at https://github.com/IBHFLab/DisoPatho.

Cocrystal formation is a widely used strategy in solid-state chemistry and pharmaceutical development to improve the solubility, stability, and bioavailability of molecules with otherwise poor physicochemical properties. Identifying viable coformer combinations remains laborious and uncertain. A key but underappreciated challenge is that experimental databases overwhelmingly report successful cocrystals, while unsuccessful attempts are rarely documented, creating biased data sets that cause many machine-learning models to make overly optimistic and unreliable predictions when applied to new chemical systems. Here, we address this limitation by reframing cocrystal prediction as a learning problem with missing negative information and by adopting a conservative strategy that focuses on identifying molecular pairs that are very unlikely to form cocrystals. We leverage multiple, independent molecular descriptions─including structural, electronic, and physicochemical characteristics─that provide complementary views for identifying reliable negatives, and use their agreement to exclude implausible combinations from large sets of untested pairs. These highly confident pseudonegative examples are then used to mitigate data imbalance and to fine-tune a pretrained graph attention network for cocrystal prediction. Across large and chemically diverse data sets, this data-centric strategy significantly improves the reliability and generalization of cocrystal prediction models compared with existing deep-learning approaches, demonstrating that carefully correcting for missing negative information is critical for making computational screening more realistic and more useful for guiding future experimental discovery.

Identifying cryptic binding sites in proteins remains a challenge in structure-based drug discovery because these sites are often not apparent in apo structures. Here, we developed and validated a novel "induce-and-identify" workflow that integrates mixed solvent molecular dynamics (MxMD) simulations with SiteMap. This approach leverages MxMD to sample protein conformations to expose hidden pockets, which are then effectively identified and ranked by SiteMap. Using a challenging data set of 65 cryptic binding sites, the developed workflow identified the cryptic binding site within the top 5 predictions in 78.5% of cases. These results suggest that the proposed MxMD + SiteMap workflow provides a robust and valuable tool for early phase drug discovery, enabling the exploration of a broader range of druggable targets by effectively inducing and identifying cryptic binding sites.

The recent discovery that inverse agonists bind to a secondary orthosteric site in PPARγ reveals unexpected structural complexity in this nuclear receptor. To differentiate between full and partial agonists, we hypothesize that partial agonists exhibit dynamic positional behavior, moving among alternative binding sites at equilibrium when saturated, while full agonists bind directly and consistently to the primary orthosteric site. Using a two-state model derived from the Boltzmann distribution and molecular docking (MODO) sampling, we developed a computational metric called the Ligand Drift Rate (LDR) Descriptor to differentiate full agonists (Emax ≥ 90%) from partial agonists (Emax ≤ 45%) of PPARγ by simulating ligand positional instability within the primary orthosteric binding pocket. We combined molecular docking and hydrogen bond fluctuation analysis across 28 PPARγ-ligand complexes (18 full agonists and 10 partial agonists). MODO sampling was performed with several algorithms, each tested on five carefully selected receptor-binding pockets from 356 crystal complexes. Hydrogen bonds between ligands and key residues (His323, His449, Tyr473) were analyzed at various hydrogen-bond displacement (HBD) thresholds (6-12 Å). The LDR was calculated as the percentage of the top 20-scoring poses that lacked hydrogen bonds. We found that adding constraints on pose diversity and allowing receptor flexibility can improve LDR's ability to distinguish full from partial agonists. Finally, MODO-sampling-based LDR calculations were conducted on 1079 ChEMBL ligands (393 full agonists and 686 partial agonists). In this validation set, LDR successfully differentiated partial from full agonists (t-test, p = 0.00043) using the genetic algorithm with the 3B1M-KRC binding pocket and a 10 Å hydrogen-bond threshold. In QSAR applications for drug discovery, we have developed a dynamics-aware molecular descriptor, LDR, that can be a key factor in quantifying partial agonism arising from small-molecule ligand drift away from the primary orthosteric site.

Early on in the emergence of virtual high-throughput screening (VHTS), it was recognized that, for validation to be robust and reliable, decoys should match actives as closely as possible in as many aspects as possible. This has given rise to several generations of validation sets that address previously reported shortcomings of earlier collections. This is an iterative and expensive method of curating validation sets that leaves ample scope for discrepancies between actives and decoys to creep in. It has previously been conjectured that in silico isomerization offers an attractive alternative for generating decoys for drug-like compounds that naturally mitigates many of these discrepancies. Here, we explore this proposition and prove the conjecture. We show that isomerization can produce molecules that have hydrogen bond acceptor, donor, rotatable bonds counts, charge, and surface area distributions that match experimental actives more closely than experimental decoys. While these are properties that receive a lot of attention in drug design, we also show that isomerization can effectively produce decoys that are positioned more closely to actives in property hyperspace than current experimental decoys, which tend to be highly dissimilar from the actives. The latter is a significant shortcoming that has thus far remained unreported and unaddressed. Additionally, we introduce the concept of pseudoisomers - decoys that have nearly identical atomic compositions to the actives - and show how they too can be used in VHTS validation at practically no cost. Herein, we build upon the methods, tools, and work of others to facilitate the generation of new and better validation sets more cheaply and efficiently, in the hope of moving the field of VHTS forward toward maturity. To that end, we make our code fully and freely available on GitHub (https://github.com/sivanovMU-Sofia/isomerization).

Equilibrist is a free, open-source platform for the quantitative analysis of chemical equilibria and kinetics. It runs in a local web browser, requires no installation, and uses a plain-text scripting language to specify reversible and irreversible reactions, polyprotic acid-base networks, and thermodynamic-cycle constraints. Five fitting modes cover species concentration, NMR shifts (fast exchange), NMR integrations (slow exchange), a mixed mode, and UV-vis/emission spectrophotometry, optimized by the L-BFGS-B and Nelder-Mead methods. A comprehensive statistical layer adds information criteria (AIC, AICc, BIC) with Akaike weights and a Fisher F-test for model selection, residual diagnostics (Durbin-Watson and Shapiro-Wilk statistics, Q-Q plot), and three uncertainty quantification families (bootstrap, jackknife, and Monte Carlo input propagation). Parameter identifiability is mapped by 1D and 2D RMSE profiles, sloppy-spectrum eigen decomposition, correlation heatmaps, t-tests, and a novel local sensitivity coefficient ξ. The combined diagnostics let the user move past R[2] coefficients and Hessian standard errors to assess parameter estimates and the chemical model itself.

Balancing target-specific biological affinity with drug-likeness remains a central challenge in de novo molecular design, where existing generative models often exhibit limited controllability or reduced structural diversity. Here, we present DF-S4, a conditional molecular generation framework based on Structured State Space Models (S4), which addresses this limitation through a disentangled latent representation and hierarchical feature-wise linear modulation (FiLM). By decoupling structural and property variables and injecting conditional signals across multiple representation levels, DF-S4 enables fine-grained and stable multi-objective control beyond conventional input-level conditioning. Evaluated via a rigorous progressive multi-denominator auditing framework across three kinase targets (EGFR, BRAF, and FGFR1), DF-S4 exhibits robust target-steering performance, yielding favorable intradomain active ratios (74.2-83.9%) while maintaining high novelty (>95%) and competitive internal diversity (∼0.85). Furthermore, DF-S4 shifts the multi-objective Pareto Frontier toward regions of simultaneous high apparent affinity and drug-likeness, mitigating the distributional-trapping trade-offs commonly observed in prior approaches. Molecular docking analyses confirm the physical plausibility of generated candidates, yielding stronger binding affinities than reference inhibitors. Finally, comprehensive ablation studies and latent space dependence analyses mathematically validate that both explicit latent disentanglement and hierarchical FiLM modulation are critical for robust feature isolation, tighter property alignment, and generative stability.

Reliable prediction of hydrogen adsorption free energy (ΔGH) is essential for accelerating electrocatalyst discovery for the alkaline hydrogen evolution reaction (HER), yet practical machine-learning workflows remain limited by inconsistent energetic definitions, heterogeneous density functional theory (DFT) protocols, and poor out-of-distribution (OOD) generalization. Here, we present an energetically anchored machine-learning framework that integrates machine-learning interatomic potential (MLIP)-derived energetic descriptors with pretrained Crystal Hamiltonian Graph Neural Network (CHGNet) latent embeddings to predict DFT-defined hydrogen adsorption energetics across chemically diverse catalyst surfaces. The framework employs protocol-consistent single-point MLIP energy evaluation on reference geometries to construct thermodynamically aligned energetic descriptors while combining them with structural representations through gradient-boosting regression. Using curated data sets from Catalysis Hub and AQCat25 containing single-, bi-, and trimetallic adsorption systems, the embedding-augmented energetic model achieved high predictive accuracy (R[2] = 0.976, MAE = 0.054 eV, RMSE = 0.105 eV) under site-level data splitting, substantially outperforming embedding-only and physicochemical-descriptor-only models. Shapley additive explanations analysis revealed a hierarchical learning mechanism in which the protocol-consistent energetic descriptor serves as the dominant thermodynamic anchor, whereas structural descriptors provide secondary refinements that improve adsorption-energy discrimination, particularly near the thermoneutral regime relevant to catalyst screening. Explicit evaluation of MLIP-only geometry workflows further demonstrated that the framework retains meaningful adsorption-energy ranking capability despite degradation introduced by MLIP-relaxed geometries. Additional OOD and transfer-learning analyses showed that predictive robustness depends strongly on the balance between compositional diversity and reference-protocol consistency. These results establish energetically anchored MLIP embeddings as an effective strategy for scalable post-DFT adsorption-energy refinement and data-efficient electrocatalyst screening while clarifying the practical limitations of MLIP-driven workflows for heterogeneous catalysis.

Predicting Drug-Target Interactions (DTIs) is a crucial task in drug discovery. Recent advances in deep learning, particularly the application of Large Language Models (LLMs), have shown promise in encoding sequential information from SMILES strings and protein sequences. However, integrating these diverse modalities remains a challenge. In this paper, we propose TextDTI, a multimodal framework that simultaneously exploits sequential and structural representations. First, TextDTI utilizes Pretrained Language Models (PLMs) to generate functional description texts for proteins represented by their amino acid sequences, which includes biological functions, molecular mechanisms, and pathway involvement. Next, the generated texts and the SMILES sequences of drugs are encoded into corresponding feature representations by other separate LLMs. Third, drug and target characteristics are fused through convolutional and graph-based modules. Finally, unidentified drug-target interactions are classified using a multilayer perceptron neural network. We further enhance feature alignment through adversarial learning and contrastive loss, resulting in robust and high-performance DTI prediction. Experiments conducted on multiple data sets in both single-domain and cross-domain settings demonstrate that our model outperforms other baseline methods. The source code and data sets are available at https://github.com/xiaoyiliu-usc/TextDTI.

Triple-negative breast cancer (TNBC) is an aggressive breast cancer subtype with a poor prognosis. The absence of effective targeted therapies and endocrine treatment options leads to limited therapeutic options, which remains one of the major clinical challenges in TNBC management. Drug discovery is typically a lengthy and costly process that could be significantly improved through drug repurposing. However, the biological complexity and insufficient repurposing strategies hinder the reuse. This study aims to develop a deep learning-based framework to accelerate drug discovery for TNBC, identify novel therapeutic candidates, and uncover potential drug targets. We developed a deep neural network framework to predict the anticancer efficacy, toxicity profiles, and structural similarities of compounds. By applying this platform to screen over 6,000 compounds from the Drug Repurposing Hub, we identified promising candidates with potential therapeutic efficacy and safety profiles against TNBC. The top-predicted compounds were subsequently validated through comprehensive in vitro and in vivo functional assays. Furthermore, we employed transcriptomic sequencing and mass spectrometry-based proteomics to elucidate the molecular mechanisms underlying the anti-TNBC activity. We identified emodepside, a structurally unique molecule diverging from conventional anticancer agents that exhibited potent antitumor efficacy across multiple TNBC cell lines. Significantly, emodepside administration (5 mg/kg) inhibited tumor growth in xenograft models. Integrated multiomics analyses (RNA-seq/CETSA-MS) identified NAMPT as the primary target. This study demonstrates the viability of our deep learning models to discover structurally novel anticancer agents that are distinct from conventional drugs, thereby expanding the therapeutic arsenal for TNBC patients. Emodepside emerges as a promising TNBC therapeutic candidate, with a possible mechanism of promoting TNBC cell apoptosis via NAMPT inhibition.

Electroporation triggers topological transitions in lipid membranes, yet the nanoscale spatiotemporal reorganization of cholesterol (Chol), acyl chains, and local electrostatic fields remains unexplored. Employing molecular dynamics (MD) simulations coupled with enhanced sampling, we map the structural thermodynamics of pore nucleation across multiple lipid species and Chol concentrations, analyzing 99 membrane systems (33 zero- and 66 applied-field) totaling >0.40 ms of MD sampling. We reveal that pore-induced acyl-chain interdigitation acts as a thermodynamic barrier, forcibly expelling rigid Chol. Whole-membrane z-profiles of membrane potential and water-dipole orientation largely mirror bulk behavior, showing that global averaging masks the pore-core electrostatic response and necessitating characteristic-radius-resolved local sampling. Crucially, high applied potentials induce an electro-topological disparity: voltage-driven leaflet interpenetration structurally outpaces aqueous pore dilation. Radius-resolved electrostatic matrices, phenotype clustering, and state-progression analyses show that interdigitation-defined local domains concentrate core water ordering and potential amplification. Rather than a simple nanoscale hole, this extensive electro-deformation generates an expanded "lipid scar" characterized by substantial tail entanglement, sharp lateral sterol gradients, and localized electrostatic enrichment. This amplified structural prestrain creates a thermodynamically vulnerable domain, potentially lowering the activation energy for consecutive poration, clarifying the molecular origins of electrosensitization in pulsed electric field therapies.

In this work, we explore the uses of biased sampling simulations including constant-velocity SMD (cv-SMD) and umbrella sampling to study molecular permeation of three highly distinct solutes spanning over 3 orders of magnitude in the permeability space, across a model of the blood-brain barrier (BBB). These molecules range from among the slowest permeants known (temozolomide; Papp ∼ 10[-6] cm s[-1]) to among the fastest compound known (ethanol; Papp ∼ 10[-3] cm s[-1]). We arrive at the transmembrane solute permeability, Psim, by the Inhomogeneous Solubility-Diffusion equation combined with an unbiasing procedure involving the Green-Kubo relations, to find a ∼1 order of magnitude agreement with experiments. We find that the converged thermodynamics of crossing span distinct topologies for each of the compounds considered, while noting that there is an apparent relationship between the thermodynamic barrier height (ΔG[‡]) and experimental permeability, Papp for these molecules.

The rapid expansion of compound libraries has significantly advanced drug discovery, especially through ultralarge library screenings that provide access to vast chemical spaces. However, the sheer scale of these libraries introduces substantial challenges in the early stages of drug discovery. While it is true that searching larger libraries can improve the hit rate of a virtual screening campaign, as the available chemical space has increased to over 64 billion molecules, identifying relevant information becomes a complex and time-consuming task. Machine learning (ML)-based docking methods and scoring functions offer a potential solution by providing increased speed and scalability. However, much of their reported success is based on benchmark data sets that rely on constructed decoys and ligand sets, which often have hidden biases that can artificially inflate performance and fail to capture the challenges of real-world screening. In this work, we systematically evaluate the performance of a popular ML-based docking method, DiffDock-Pocket, on high-throughput screening (HTS) data sets derived from the PubChem BioAssay database, a premier source of bioactivity data. By using HTS data sets as a more realistic benchmark, we aim to provide a clearer picture of how ML models perform in practical virtual screening scenarios compared to traditional physics-based docking approaches. Our work highlights the strengths and limitations of current ML methods and offers insights into their reliability and applicability to prospective applications in drug discovery.