Bayer's in silico ADMET platform: a journey of machine learning over the past two decades.

Over the past two decades, an in silico absorption, distribution, metabolism, and excretion (ADMET) platform has been created at Bayer Pharma with the goal to generate models for a variety of pharmacokinetic and physicochemical endpoints in early drug discovery. These tools are accessible to all scientists within the company and can be a useful in assisting with the selection and design of novel leads, as well as the process of lead optimization. Here. we discuss the development of machine-learning (ML) approaches with special emphasis on data, descriptors, and algorithms. We show that high company internal data quality and tailored descriptors, as well as a thorough understanding of the experimental endpoints, are essential to the utility of our models. We discuss the recent impact of deep neural networks and show selected application examples.

Discovery of Rogaratinib (BAY 1163877): a pan-FGFR Inhibitor.

Rogaratinib (BAY 1163877) is a highly potent and selective small-molecule pan-fibroblast growth factor receptor (FGFR) inhibitor (FGFR1-4) for oral application currently being investigated in phase 1 clinical trials for the treatment of cancer. In this publication, we report its discovery by de novo structure-based design and medicinal chemistry optimization together with its pharmacokinetic profile.

Novel TIE-2 inhibitor BAY-826 displays in vivo efficacy in experimental syngeneic murine glioma models.

Targeting the vascular endothelial growth factor signaling axis in glioblastoma inevitably leads to tumor recurrence and a more aggressive phenotype. Therefore, other angiogenic pathways, like the angiopoietin/tunica interna endothelial cell kinase (TIE) signaling axis, have become additional targets for therapeutic intervention. Here, we explored whether targeting the receptor tyrosine kinase TIE-2 using a novel, highly potent, orally available small molecule TIE-2 inhibitor (BAY-826) improves tumor control in syngeneic mouse glioma models. BAY-826 inhibits TIE-2 phosphorylation in vitro and in vivo as demonstrated by suppression of Angiopoietin-1- or Na3 VO4 -induced TIE-2 phosphorylation in glioma cells or extracts of lungs from BAY-826-treated mice. There was a trend toward prolonged survival upon single-agent treatment in two of four models (SMA-497 and SMA-540) and there was a significant survival benefit in one model (SMA-560). Co-treatment with BAY-826 and irradiation was ineffective in one model (SMA-497), but provided synergistic prolongation of survival in another (SMA-560). Decreased vessel densities and increased leukocyte infiltration were observed, but might be independent processes as the effect was also observed in single treatment modalities. These data demonstrate that TIE-2 inhibition may improve tumor response to treatment in highly vascularized tumors such as glioblastoma.

Best of both worlds: combining pharma data and state of the art modeling technology to improve in Silico pKa prediction.

In a unique collaboration between a software company and a pharmaceutical company, we were able to develop a new in silico pKa prediction tool with outstanding prediction quality. An existing pKa prediction method from Simulations Plus based on artificial neural network ensembles (ANNE), microstates analysis, and literature data was retrained with a large homogeneous data set of drug-like molecules from Bayer. The new model was thus built with curated sets of ∼14,000 literature pKa values (∼11,000 compounds, representing literature chemical space) and ∼19,500 pKa values experimentally determined at Bayer Pharma (∼16,000 compounds, representing industry chemical space). Model validation was performed with several test sets consisting of a total of ∼31,000 new pKa values measured at Bayer. For the largest and most difficult test set with >16,000 pKa values that were not used for training, the original model achieved a mean absolute error (MAE) of 0.72, root-mean-square error (RMSE) of 0.94, and squared correlation coefficient (R(2)) of 0.87. The new model achieves significantly improved prediction statistics, with MAE = 0.50, RMSE = 0.67, and R(2) = 0.93. It is commercially available as part of the Simulations Plus ADMET Predictor release 7.0. Good predictions are only of value when delivered effectively to those who can use them. The new pKa prediction model has been integrated into Pipeline Pilot and the PharmacophorInformatics (PIx) platform used by scientists at Bayer Pharma. Different output formats allow customized application by medicinal chemists, physical chemists, and computational chemists.

CypScore: Quantitative prediction of reactivity toward cytochromes P450 based on semiempirical molecular orbital theory.

CypScore is an in silico approach for predicting the likely sites of cytochrome P450-mediated metabolism of druglike organic molecules. It consists of multiple models for the most important P450 oxidation reactions such as aliphatic hydroxylation, N-dealkylation, O-dealkylation, aromatic hydroxylation, double-bond oxidation, N-oxidation, and S-oxidation. Each of these models is based on atomic reactivity descriptors derived from surface-based properties calculated with ParaSurf and based on AM1 semiempirical molecular orbital theory. The models were trained with data derived from Bayer Schering Pharma's in-house MajorMetabolite Database with more than 2300 transformations and more than 800 molecules collected from the primary literature. The models have been balanced to allow the treatment of relative intramolecular, intra-chemotype, and inter-chemotype reactivities of the labile sites toward oxidation. The models were evaluated with promising hit rates on three public datasets of varying quality in the annotation of the experimental positions. For 39 well-characterized compounds from 14 in-house lead optimization programs, we could detect at least one major metabolite for the three highest-ranked positions in 87 % of the compounds and overall more than 62 % of all major metabolites, with promising true- to false-positive ratios of 0.9.

Design and synthesis of potent and selective azaindole-based Rho kinase (ROCK) inhibitors.

Rho kinase plays a pivotal role in several cellular processes such as vasoregulation, making it a suitable target for the treatment of hypertension and related disorders. We discovered a new compound class of Rho kinase (ROCK) inhibitors containing a 7-azaindole hinge-binding scaffold tethered to an aminopyrimidine core. Herein we describe the structure-activity relationships elucidated through biochemical and functional assays. The introduction of suitable substituents at the 3-position of the bicyclic moiety led to an increase in activity, which was required to design compounds with favorable pharmacokinetic profile. Azaindole 32 was identified as a highly selective and orally available ROCK inhibitor able to cause a sustained blood pressure reduction in vivo.

In silico ADMET traffic lights as a tool for the prioritization of HTS hits.

The need for in silico characterization of HTS hit structures as part of a data-driven hit-selection process is demonstrated. A solution is described in the form of an in silico ADMET traffic light and PhysChem scoring system. This has been extensively validated with in-house data at Bayer, published data, and a collection of launched small-molecule oral drugs.

Improving the hit-to-lead process: data-driven assessment of drug-like and lead-like screening hits.

Drug-like and lead-like hits derived from HTS campaigns provide good starting points for lead optimization. However, too strong emphasis on potency as hit-selection parameter might hamper the success of such projects. A detailed absorption, distribution, metabolism, excretion and toxicology (ADME-Tox) profiling is needed to help identify hits with a minimum number of (known) liabilities. This is particularly true for drug-like hits. Herein, we describe how to break down large numbers of screening hits and we provide a comprehensive overview of the strengths and weaknesses for each structural class. The overall profile (e.g. ligand efficiency, selectivity and ADME-Tox) is the distinctive feature that will define the priority for follow-up.

Molecular characterization of the principal substrate binding site of the ubiquitous folding catalyst protein disulfide isomerase.

Disulfide bond formation in the endoplasmic reticulum of eukaryotes is catalyzed by the ubiquitously expressed enzyme protein disulfide isomerase (PDI). The effectiveness of PDI as a catalyst of native disulfide bond formation in folding polypeptides depends on the ability to catalyze disulfide-dithiol exchange, to bind non-native proteins, and to trigger conformational changes in the bound substrate, allowing access to buried cysteine residues. It is known that the b' domain of PDI provides the principal peptide binding site of PDI and that this domain is critical for catalysis of isomerization but not oxidation reactions in protein substrates. Here we use homology modeling to define more precisely the boundaries of the b' domain and show the existence of an intradomain linker between the b' and a' domains. We have expressed the recombinant b' domain thus defined; the stability and conformational properties of the recombinant product confirm the validity of the domain boundaries. We have modeled the tertiary structure of the b' domain and identified the primary substrate binding site within it. Mutations within this site, expressed both in the isolated domain and in full-length PDI, greatly reduce the binding affinity for small peptide substrates, with the greatest effect being I272W, a mutation that appears to have no structural effect.

Recent advances in the prediction of blood-brain partitioning from molecular structure.

The widely distributed software tools Cerius2 and ACD/log D Suite have been used to develop a new method for the prediction of the ratio of concentrations of a drug in the brain and blood (BB,quantified as log BB) from structure. The performances of all known blood-brain partitioning prediction methods are compared to give an up-to-date account on their accuracy, limitations, and usefulness. It is demonstrated that the new log BB prediction method is superior to other methods with regard to low-to-medium throughput log BB prediction, whereas the C2-ADME log BB two-dimensional (2D) method seems to offer the best compromise between speed and accuracy for ultra-high throughput processing of large compound databases for log BB prediction.

In silico prediction of aqueous solubility, human plasma protein binding and volume of distribution of compounds from calculated pKa and AlogP98 values.

We have investigated whether three important ADME (absorption, distribution, metabolism, excretion) related properties (aqueous solubility, human plasma protein binding, and human volume of distribution at steady-state) can be predicted from chemical structure alone if only the predicted predominant ionisation state and lipophilicity (calculated logP [P = octanol-water partition coefficient]) are considered. A simple, fast method for the in silico prediction of aqueous solubility of predominantly uncharged compounds has been developed, while some potential is shown for the prediction of predominantly charged or zwitterionic compounds. Ten other known in silico prediction methods for aqueous solubility have also been evaluated. It has furthermore been demonstrated that the molecular weight (MW) profile of training sets for the development of aqueous solubility prediction methods can influence their predictive performance with regard to test sets of either matching or diverging profiles. The same property descriptors which have been found most relevant for the prediction of aqueous solubility have also proved useful for the prediction of human plasma protein binding and human volume of distribution at steady-state.