Actionable Predictions of Human Pharmacokinetics at the Drug Design Stage.

We present a novel computational approach for predicting human pharmacokinetics (PK) that addresses the challenges of early stage drug design. Our study introduces and describes a large-scale data set of 11 clinical PK end points, encompassing over 2700 unique chemical structures to train machine learning models. To that end multiple advanced training strategies are compared, including the integration of in vitro data and a novel self-supervised pretraining task. In addition to the predictions, our final model provides meaningful epistemic uncertainties for every data point. This allows us to successfully identify regions of exceptional predictive performance, with an absolute average fold error (AAFE/geometric mean fold error) of less than 2.5 across multiple end points. Together, these advancements represent a significant leap toward actionable PK predictions, which can be utilized early on in the drug design process to expedite development and reduce reliance on nonclinical studies.

Discovery of a first-in-class orally available HBV cccDNA inhibitor.

BACKGROUND & AIMS: The persistence of covalently closed circular DNA (cccDNA) in infected hepatocytes is the major barrier preventing viral eradication with existing therapies in patients with chronic hepatitis B. Therapeutic agents that can eliminate cccDNA are urgently needed to achieve viral eradication and thus HBV cure. METHODS: A phenotypic assay with HBV-infected primary human hepatocytes (PHHs) was employed to screen for novel cccDNA inhibitors. A HBVcircle mouse model and a uPA-SCID (urokinase-type plasminogen activator-severe combined immunodeficiency) humanized liver mouse model were used to evaluate the anti-HBV efficacy of the discovered cccDNA inhibitors. RESULTS: Potent and dose-dependent reductions in extracellular HBV DNA, HBsAg, and HBeAg levels were achieved upon the initiation of ccc_R08 treatment two days after the HBV infection of PHHs. More importantly, the level of cccDNA was specifically reduced by ccc_R08, while it did not obviously affect mitochondrial DNA. Additionally, ccc_R08 showed no significant cytotoxicity in PHHs or in multiple proliferating cell lines. The twice daily oral administration of ccc_R08 to HBVcircle model mice, which contained surrogate cccDNA molecules, significantly decreased the serum levels of HBV DNA and antigens, and these effects were sustained during the off-treatment follow-up period. Moreover, at the end of follow-up, the levels of surrogate cccDNA molecules in the livers of ccc_R08-treated HBVcircle mice were reduced to below the lower limit of quantification. CONCLUSIONS: We have discovered a small-molecule cccDNA inhibitor that reduces HBV cccDNA levels. cccDNA inhibitors potentially represent a new approach to completely cure patients chronically infected with HBV. IMPACT AND IMPLICATIONS: Covalently closed circular DNA (cccDNA) persistence in HBV-infected hepatocytes is the root cause of chronic hepatitis B. We discovered a novel small-molecule cccDNA inhibitor that can specifically reduce cccDNA levels in HBV-infected hepatocytes. This type of molecule could offer a new approach to completely cure patients chronically infected with HBV.

Computational Predictions of Nonclinical Pharmacokinetics at the Drug Design Stage.

Although computational predictions of pharmacokinetics (PK) are desirable at the drug design stage, existing approaches are often limited by prediction accuracy and human interpretability. Using a discovery data set of mouse and rat PK studies at Roche (9,685 unique compounds), we performed a proof-of-concept study to predict key PK properties from chemical structure alone, including plasma clearance (CLp), volume of distribution at steady-state (Vss), and oral bioavailability (F). Ten machine learning (ML) models were evaluated, including Single-Task, Multitask, and transfer learning approaches (i.e., pretraining with in vitro data). In addition to prediction accuracy, we emphasized human interpretability of outcomes, especially the quantification of uncertainty, applicability domains, and explanations of predictions in terms of molecular features. Results show that intravenous (IV) PK properties (CLp and Vss) can be predicted with good precision (average absolute fold error, AAFE of 1.96-2.84 depending on data split) and low bias (average fold error, AFE of 0.98-1.36), with AutoGluon, Gaussian Process Regressor (GP), and ChemProp displaying the best performance. Driven by higher complexity of oral PK studies, predictions of F were more challenging, with the best AAFE values of 2.35-2.60 and higher overprediction bias (AFE of 1.45-1.62). Multi-Task approaches and pretraining of ChemProp neural networks with in vitro data showed similar precision to Single-Task models but helped reduce the bias and increase correlations between observations and predictions. A combination of GP-computed prediction variance, molecular clustering, and dimensionality-reduction provided valuable quantitative insights into prediction uncertainty and applicability domains. SHAPley Additive exPlanations (SHAPs) highlighted molecular features contributing to prediction outcomes of Vss, providing explanations that could aid drug design. Combined results show that computational predictions of PK are feasible at the drug design stage, with several ML technologies converging to successfully leverage historical PK data sets. Further studies are needed to unlock the full potential of this approach, especially with respect to data set sizes and quality, transfer learning between in vitro and in vivo data sets, model-independent quantification of uncertainty, and explainability of predictions.

Inducers of the endothelial cell barrier identified through chemogenomic screening in genome-edited hPSC-endothelial cells.

The blood-retina barrier and blood-brain barrier (BRB/BBB) are selective and semipermeable and are critical for supporting and protecting central nervous system (CNS)-resident cells. Endothelial cells (ECs) within the BRB/BBB are tightly coupled, express high levels of Claudin-5 (CLDN5), a junctional protein that stabilizes ECs, and are important for proper neuronal function. To identify novel CLDN5 regulators (and ultimately EC stabilizers), we generated a CLDN5-P2A-GFP stable cell line from human pluripotent stem cells (hPSCs), directed their differentiation to ECs (CLDN5-GFP hPSC-ECs), and performed flow cytometry-based chemogenomic library screening to measure GFP expression as a surrogate reporter of barrier integrity. Using this approach, we identified 62 unique compounds that activated CLDN5-GFP. Among them were TGF-ß pathway inhibitors, including RepSox. When applied to hPSC-ECs, primary brain ECs, and retinal ECs, RepSox strongly elevated barrier resistance (transendothelial electrical resistance), reduced paracellular permeability (fluorescein isothiocyanate-dextran), and prevented vascular endothelial growth factor A (VEGFA)-induced barrier breakdown in vitro. RepSox also altered vascular patterning in the mouse retina during development when delivered exogenously. To determine the mechanism of action of RepSox, we performed kinome-, transcriptome-, and proteome-profiling and discovered that RepSox inhibited TGF-ß, VEGFA, and inflammatory gene networks. In addition, RepSox not only activated vascular-stabilizing and barrier-establishing Notch and Wnt pathways, but also induced expression of important tight junctions and transporters. Taken together, our data suggest that inhibiting multiple pathways by selected individual small molecules, such as RepSox, may be an effective strategy for the development of better BRB/BBB models and novel EC barrier-inducing therapeutics.

Blocked inverted indices for exact clustering of large chemical spaces.

The calculation of pairwise compound similarities based on fingerprints is one of the fundamental tasks in chemoinformatics. Methods for efficient calculation of compound similarities are of the utmost importance for various applications like similarity searching or library clustering. With the increasing size of public compound databases, exact clustering of these databases is desirable, but often computationally prohibitively expensive. We present an optimized inverted index algorithm for the calculation of all pairwise similarities on 2D fingerprints of a given data set. In contrast to other algorithms, it neither requires GPU computing nor yields a stochastic approximation of the clustering. The algorithm has been designed to work well with multicore architectures and shows excellent parallel speedup. As an application example of this algorithm, we implemented a deterministic clustering application, which has been designed to decompose virtual libraries comprising tens of millions of compounds in a short time on current hardware. Our results show that our implementation achieves more than 400 million Tanimoto similarity calculations per second on a common desktop CPU. Deterministic clustering of the available chemical space thus can be done on modern multicore machines within a few days.

G-PLIP: Knowledge graph neural network for structure-free protein-ligand bioactivity prediction.

Protein-ligand interactions (PLIs) determine the efficacy and safety profiles of small molecule drugs. Existing methods rely on either structural information or resource-intensive computations to predict PLI, casting doubt on whether it is possible to perform structure-free PLI predictions at low computational cost. Here we show that a light-weight graph neural network (GNN), trained with quantitative PLIs of a small number of proteins and ligands, is able to predict the strength of unseen PLIs. The model has no direct access to structural information about the protein-ligand complexes. Instead, the predictive power is provided by encoding the entire chemical and proteomic space in a single heterogeneous graph, encapsulating primary protein sequence, gene expression, the protein-protein interaction network, and structural similarities between ligands. This novel approach performs competitively with, or better than, structure-aware models. Our results suggest that existing PLI prediction methods may be improved by incorporating representation learning techniques that embed biological and chemical knowledge.

Optimization of the TeraTox Assay for Preclinical Teratogenicity Assessment.

Current animal-free methods to assess teratogenicity of drugs under development still deliver high numbers of false negatives. To improve the sensitivity of human teratogenicity prediction, we characterized the TeraTox test, a newly developed multilineage differentiation assay using 3D human-induced pluripotent stem cells. TeraTox produces primary output concentration-dependent cytotoxicity and altered gene expression induced by each test compound. These data are fed into an interpretable machine-learning model to perform prediction, which relates to the concentration-dependent human teratogenicity potential of drug candidates. We applied TeraTox to profile 33 approved pharmaceuticals and 12 proprietary drug candidates with known in vivo data. Comparing TeraTox predictions with known human or animal toxicity, we report an accuracy of 69% (specificity: 53%, sensitivity: 79%). TeraTox performed better than 2 quantitative structure-activity relationship models and had a higher sensitivity than the murine embryonic stem cell test (accuracy: 58%, specificity: 76%, and sensitivity: 46%) run in the same laboratory. The overall prediction accuracy could be further improved by combining TeraTox and mouse embryonic stem cell test results. Furthermore, patterns of altered gene expression revealed by TeraTox may help grouping toxicologically similar compounds and possibly deducing common modes of action. The TeraTox assay and the dataset described here therefore represent a new tool and a valuable resource for drug teratogenicity assessment.

Clinical Characterization of Epilepsy in Children With Angelman Syndrome.

BACKGROUND: Epilepsy is highly prevalent in children with Angelman syndrome (AS), and its detailed characterization and relationship to the genotype (deletion vs nondeletion) is important both for medical practice and for clinical trial design. METHODS AND MATERIALS: We retrospectively analyzed the main clinical features of epilepsy in 265 children with AS who were enrolled in the AS Natural History Study, a multicenter, observational study conducted at six centers in the United States. Participants were prospectively followed up and classified by genotype. RESULTS: Epilepsy was reported in a greater proportion of individuals with a deletion than a nondeletion genotype (171 of 187 [91%] vs. 48 of 78 [61%], P < 0.001). Compared with participants with a nondeletion genotype, those with deletions were younger at the time of the first seizure (age: median [95% confidence interval]: 24 [21-24] months vs. 57 [36-85] months, P < 0.001) and had a higher prevalence of generalized motor seizures. Hospitalization following a seizure was reported in more children with a deletion than a nondeletion genotype (92 of 171 [54%] vs. 17 of 48 [36%], P = 0.04). The overall prevalence of absence seizures was not significantly different between genotype groups. Forty-six percent (102/219) of the individuals reporting epilepsy were diagnosed with AS concurrently or after their first seizure. CONCLUSIONS: Significant differences exist in the clinical expression of epilepsy in AS according to the underlying genotype, with earlier age of onset and more severe epilepsy in individuals with AS due to a chromosome 15 deletion.

Multiscale modelling of drug mechanism and safety.

Here, we introduce models at three levels-molecular level, cellular and omics level, and organ and system level-that study drug mechanism and safety in preclinical drug discovery. The models differ in both their scope of study and technical details, but are all rooted in mathematical descriptions of complex biological systems, and all require informatics tools that handle large-volume, heterogeneous, and noisy data. We present principles and recent developments with examples at each level and highlight the synergy by a case study. We proffer a multiscale modelling view of drug discovery, call for a seamless flow of information in the form of models, and examine potential impacts.

Refinement of a model of repeated cerebrospinal fluid collection in conscious rats.

The cannulation of the cisterna magna in rats for in vivo sampling of cerebrospinal fluid serves as a valuable model for studying the delivery of new drugs into the central nervous system or disease models. It offers the advantages of repeated sampling without anesthesia-induced bias and using animals as their own controls. An established model was retrospectively reviewed for the outcomes and it was hypothesized that by refining the method, i.e. by (1) implementing pathophysiological-based anesthesia and analgesia, (2) using state-of-the-art peri-operative monitoring and supportive care, (3) increasing stability of the cement-cannula assembly, and (4) selecting a more adaptable animal strain, the outcome in using the model - quantified by peri-operative mortality, survival time and stability of the implant - could be improved and could enhance animal welfare. After refinement of the technique, peri-operative mortality decreased significantly (7 animals out of 73 compared with 4 out of 322; P = 0.001), survival time increased significantly (36 ± 14 days compared with 28 ± 18 days; P < 0.001), as well as the stability of the cement-cannula assembly (47 ± 8 days of adhesion compared with 33 ± 15 days and 34 ± 13 days using two other cement types; P < 0.001). Overall, the 3R concept of Russell and Burch was successfully addressed and animal welfare was improved by (1) the reduction in the total number of animals needed as a result of lower mortality or fewer euthanizations due to technical failure, and frequent use of individual rats over a time frame; and (2) improving the scientific quality of the model.