MELLODDY: Cross-pharma Federated Learning at Unprecedented Scale Unlocks Benefits in QSAR without Compromising Proprietary Information.

Federated multipartner machine learning has been touted as an appealing and efficient method to increase the effective training data volume and thereby the predictivity of models, particularly when the generation of training data is resource-intensive. In the landmark MELLODDY project, indeed, each of ten pharmaceutical companies realized aggregated improvements on its own classification or regression models through federated learning. To this end, they leveraged a novel implementation extending multitask learning across partners, on a platform audited for privacy and security. The experiments involved an unprecedented cross-pharma data set of 2.6+ billion confidential experimental activity data points, documenting 21+ million physical small molecules and 40+ thousand assays in on-target and secondary pharmacodynamics and pharmacokinetics. Appropriate complementary metrics were developed to evaluate the predictive performance in the federated setting. In addition to predictive performance increases in labeled space, the results point toward an extended applicability domain in federated learning. Increases in collective training data volume, including by means of auxiliary data resulting from single concentration high-throughput and imaging assays, continued to boost predictive performance, albeit with a saturating return. Markedly higher improvements were observed for the pharmacokinetics and safety panel assay-based task subsets.

KiSSim: Predicting Off-Targets from Structural Similarities in the Kinome.

Protein kinases are among the most important drug targets because their dysregulation can cause cancer, inflammatory and degenerative diseases, and many more. Developing selective inhibitors is challenging due to the highly conserved binding sites across the roughly 500 human kinases. Thus, detecting subtle similarities on a structural level can help explain and predict off-targets among the kinase family. Here, we present the kinase-focused, subpocket-enhanced KiSSim fingerprint (Kinase Structural Similarity). The fingerprint builds on the KLIFS pocket definition, composed of 85 residues aligned across all available protein kinase structures, which enables residue-by-residue comparison without a computationally expensive alignment. The residues' physicochemical and spatial properties are encoded within their structural context including key subpockets at the hinge region, the DFG motif, and the front pocket. Since structure was found to contain information complementary to sequence, we used the fingerprint to calculate all-against-all similarities within the structurally covered kinome. We could identify off-targets that are unexpected if solely considering the sequence-based kinome tree grouping; for example, Erlobinib's known kinase off-targets SLK and LOK show high similarities to the key target EGFR (TK group), although belonging to the STE group. KiSSim reflects profiling data better or at least as well as other approaches such as KLIFS pocket sequence identity, KLIFS interaction fingerprints (IFPs), or SiteAlign. To rationalize observed (dis)similarities, the fingerprint values can be visualized in 3D by coloring structures with residue and feature resolution. We believe that the KiSSim fingerprint is a valuable addition to the kinase research toolbox to guide off-target and polypharmacology prediction. The method is distributed as an open-source Python package on GitHub and as a conda package: https://github.com/volkamerlab/kissim.

Large-Scale Assessment of Binding Free Energy Calculations in Active Drug Discovery Projects.

Accurate ranking of compounds with regards to their binding affinity to a protein using computational methods is of great interest to pharmaceutical research. Physics-based free energy calculations are regarded as the most rigorous way to estimate binding affinity. In recent years, many retrospective studies carried out both in academia and industry have demonstrated its potential. Here, we present the results of large-scale prospective application of the FEP+ method in active drug discovery projects in an industry setting at Merck KGaA, Darmstadt, Germany. We compare these prospective data to results obtained on a new diverse, public benchmark of eight pharmaceutically relevant targets. Our results offer insights into the challenges faced when using free energy calculations in real-life drug discovery projects and identify limitations that could be tackled by future method development. The new public data set we provide to the community can support further method development and comparative benchmarking of free energy calculations.

Relative binding affinity prediction of farnesoid X receptor in the D3R Grand Challenge 2 using FEP.

Physics-based free energy simulations have increasingly become an important tool for predicting binding affinity and the recent introduction of automated protocols has also paved the way towards a more widespread use in the pharmaceutical industry. The D3R 2016 Grand Challenge 2 provided an opportunity to blindly test the commercial free energy calculation protocol FEP+ and assess its performance relative to other affinity prediction methods. The present D3R free energy prediction challenge was built around two experimental data sets involving inhibitors of farnesoid X receptor (FXR) which is a promising anticancer drug target. The FXR binding site is predominantly hydrophobic with few conserved interaction motifs and strong induced fit effects making it a challenging target for molecular modeling and drug design. For both data sets, we achieved reasonable prediction accuracy (RMSD ≈ 1.4 kcal/mol, rank 3-4 according to RMSD out of 20 submissions) comparable to that of state-of-the-art methods in the field. Our D3R results boosted our confidence in the method and strengthen our desire to expand its applications in future in-house drug design projects.

KinMap: a web-based tool for interactive navigation through human kinome data.

BACKGROUND: Annotations of the phylogenetic tree of the human kinome is an intuitive way to visualize compound profiling data, structural features of kinases or functional relationships within this important class of proteins. The increasing volume and complexity of kinase-related data underlines the need for a tool that enables complex queries pertaining to kinase disease involvement and potential therapeutic uses of kinase inhibitors. RESULTS: Here, we present KinMap, a user-friendly online tool that facilitates the interactive navigation through kinase knowledge by linking biochemical, structural, and disease association data to the human kinome tree. To this end, preprocessed data from freely-available sources, such as ChEMBL, the Protein Data Bank, and the Center for Therapeutic Target Validation platform are integrated into KinMap and can easily be complemented by proprietary data. The value of KinMap will be exemplarily demonstrated for uncovering new therapeutic indications of known kinase inhibitors and for prioritizing kinases for drug development efforts. CONCLUSION: KinMap represents a new generation of kinome tree viewers which facilitates interactive exploration of the human kinome. KinMap enables generation of high-quality annotated images of the human kinome tree as well as exchange of kinome-related data in scientific communications. Furthermore, KinMap supports multiple input and output formats and recognizes alternative kinase names and links them to a unified naming scheme, which makes it a useful tool across different disciplines and applications. A web-service of KinMap is freely available at http://www.kinhub.org/kinmap/ .

Coupling Matched Molecular Pairs with Machine Learning for Virtual Compound Optimization.

Matched molecular pair (MMP) analyses are widely used in compound optimization projects to gain insights into structure-activity relationships (SAR). The analysis is traditionally done via statistical methods but can also be employed together with machine learning (ML) approaches to extrapolate to novel compounds. The here introduced MMP/ML method combines a fragment-based MMP implementation with different machine learning methods to obtain automated SAR decomposition and prediction. To test the prediction capabilities and model transferability, two different compound optimization scenarios were designed: (1) "new fragments" which occurs when exploring new fragments for a defined compound series and (2) "new static core and transformations" which resembles for instance the identification of a new compound series. Very good results were achieved by all employed machine learning methods especially for the new fragments case, but overall deep neural network models performed best, allowing reliable predictions also for the new static core and transformations scenario, where comprehensive SAR knowledge of the compound series is missing. Furthermore, we show that models trained on all available data have a higher generalizability compared to models trained on focused series and can extend beyond chemical space covered in the training data. Thus, coupling MMP with deep neural networks provides a promising approach to make high quality predictions on various data sets and in different compound optimization scenarios.

Identification and Visualization of Kinase-Specific Subpockets.

The identification and design of selective compounds is important for the reduction of unwanted side effects as well as for the development of tool compounds for target validation studies. This is, in particular, true for therapeutically important protein families that possess conserved folds and have numerous members such as kinases. To support the design of selective kinase inhibitors, we developed a novel approach that allows identification of specificity determining subpockets between closely related kinases solely based on their three-dimensional structures. To account for the intrinsic flexibility of the proteins, multiple X-ray structures of the target protein of interest as well as of unwanted off-target(s) are taken into account. The binding pockets of these protein structures are calculated and fused to a combined target and off-target pocket, respectively. Subsequently, shape differences between these two combined pockets are identified via fusion rules. The approach provides a user-friendly visualization of target-specific areas in a binding pocket which should be explored when designing selective compounds. Furthermore, the approach can be easily combined with in silico alanine mutation studies to identify selectivity determining residues. The potential impact of the approach is demonstrated in four retrospective experiments on closely related kinases, i.e., p38α vs Erk2, PAK1 vs PAK4, ITK vs AurA, and BRAF vs VEGFR2. Overall, the presented approach does not require any profiling data for training purposes, provides an intuitive visualization of a large number of protein structures at once, and could also be applied to other target classes.

Pocketome of human kinases: prioritizing the ATP binding sites of (yet) untapped protein kinases for drug discovery.

Protein kinases are involved in a variety of diseases including cancer, inflammation, and autoimmune disorders. Although the development of new kinase inhibitors is a major focus in pharmaceutical research, a large number of kinases remained so far unexplored in drug discovery projects. The selection and assessment of targets is an essential but challenging area. Today, a few thousands of experimentally determined kinase structures are available, covering about half of the human kinome. This large structural source allows guiding the target selection via structure-based druggability prediction approaches such as DoGSiteScorer. Here, a thorough analysis of the ATP pockets of the entire human kinome in the DFG-in state is presented in order to prioritize novel kinase structures for drug discovery projects. For this, all human kinase X-ray structures available in the PDB were collected, and homology models were generated for the missing part of the kinome. DoGSiteScorer was used to calculate geometrical and physicochemical properties of the ATP pockets and to predict the potential of each kinase to be druggable. The results indicate that about 75% of the kinome are in principle druggable. Top ranking structures comprise kinases that are primary targets of known approved drugs but additionally point to so far less explored kinases. The presented analysis provides new insights into the druggability of ATP binding pockets of the entire kinome. We anticipate this comprehensive druggability assessment of protein kinases to be helpful for the community to prioritize so far untapped kinases for drug discovery efforts.

Predicting enzymatic function from global binding site descriptors.

Due to the rising number of solved protein structures, computer-based techniques for automatic protein functional annotation and classification into families are of high scientific interest. DoGSiteScorer automatically calculates global descriptors for self-predicted pockets based on the 3D structure of a protein. Protein function predictors on three levels with increasing granularity are built by use of a support vector machine (SVM), based on descriptors of 26632 pockets from enzymes with known structure and enzyme classification. The SVM models represent a generalization of the available descriptor space for each enzyme class, subclass, and substrate-specific sub-subclass. Cross-validation studies show accuracies of 68.2% for predicting the correct main class and accuracies between 62.8% and 80.9% for the six subclasses. Substrate-specific recall rates for a kinase subset are 53.8%. Furthermore, application studies show the ability of the method for predicting the function of unknown proteins and gaining valuable information for the function prediction field.

DoGSiteScorer: a web server for automatic binding site prediction, analysis and druggability assessment.

MOTIVATION: Many drug discovery projects fail because the underlying target is finally found to be undruggable. Progress in structure elucidation of proteins now opens up a route to automatic structure-based target assessment. DoGSiteScorer is a newly developed automatic tool combining pocket prediction, characterization and druggability estimation and is now available through a web server. AVAILABILITY: The DoGSiteScorer web server is freely available for academic use at http://dogsite.zbh.uni-hamburg.de CONTACT: rarey@zbh.uni-hamburg.de.

TRAPP: a tool for analysis of transient binding pockets in proteins.

We present TRAPP (TRAnsient Pockets in Proteins), a new automated software platform for tracking, analysis, and visualization of binding pocket variations along a protein motion trajectory or within an ensemble of protein structures that may encompass conformational changes ranging from local side chain fluctuations to global backbone motions. TRAPP performs accurate grid-based calculations of the shape and physicochemical characteristics of a binding pocket for each structure and detects the conserved and transient regions of the pocket in an ensemble of protein conformations. It also provides tools for tracing the opening of a particular subpocket and residues that contribute to the binding site. TRAPP thus enables an assessment of the druggability of a disease-related target protein taking its flexibility into account.

Protein pocket and ligand shape comparison and its application in virtual screening.

Understanding molecular recognition is one major requirement for drug discovery and design. Physicochemical and shape complementarity between two binding partners is the driving force during complex formation. In this study, the impact of shape within this process is analyzed. Protein binding pockets and co-crystallized ligands are represented by normalized principal moments of inertia ratios (NPRs). The corresponding descriptor space is triangular, with its corners occupied by spherical, discoid, and elongated shapes. An analysis of a selected set of sc-PDB complexes suggests that pockets and bound ligands avoid spherical shapes, which are, however, prevalent in small unoccupied pockets. Furthermore, a direct shape comparison confirms previous studies that on average only one third of a pocket is filled by its bound ligand, supplemented by a 50 % subpocket coverage. In this study, we found that shape complementary is expressed by low pairwise shape distances in NPR space, short distances between the centers-of-mass, and small deviations in the angle between the first principal ellipsoid axes. Furthermore, it is assessed how different binding pocket parameters are related to bioactivity and binding efficiency of the co-crystallized ligand. In addition, the performance of different shape and size parameters of pockets and ligands is evaluated in a virtual screening scenario performed on four representative targets.

SUMO: In Silico Sequence Assessment Using Multiple Optimization Parameters.

To select the most promising screening hits from antibody and VHH display campaigns for subsequent in-depth profiling and optimization, it is highly desirable to assess and select sequences on properties beyond only their binding signals from the sorting process. In addition, developability risk criteria, sequence diversity, and the anticipated complexity for sequence optimization are relevant attributes for hit selection and optimization. Here, we describe an approach for the in silico developability assessment of antibody and VHH sequences. This method not only allows for ranking and filtering multiple sequences with regard to their predicted developability properties and diversity, but also visualizes relevant sequence and structural features of potentially problematic regions and thereby provides rationales and starting points for multi-parameter sequence optimization.

Combining global and local measures for structure-based druggability predictions.

Predicting druggability and prioritizing certain disease modifying targets for the drug development process is of high practical relevance in pharmaceutical research. DoGSiteScorer is a fully automatic algorithm for pocket and druggability prediction. Besides consideration of global properties of the pocket, also local similarities shared between pockets are reflected. Druggability scores are predicted by means of a support vector machine (SVM), trained, and tested on the druggability data set (DD) and its nonredundant version (NRDD). The DD consists of 1069 targets with assigned druggable, difficult, and undruggable classes. In 90% of the NRDD, the SVM model based on global descriptors correctly classifies a target as either druggable or undruggable. Nevertheless, global properties suffer from binding site changes due to ligand binding and from the pocket boundary definition. Therefore, local pocket properties are additionally investigated in terms of a nearest neighbor search. Local similarities are described by distance dependent histograms between atom pairs. In 88% of the DD pocket set, the nearest neighbor and the structure itself conform with their druggability type. A discriminant feature between druggable and undruggable pockets is having less short-range hydrophilic-hydrophilic pairs and more short-range lipophilic-lipophilic pairs. Our findings for global pocket descriptors coincide with previously published methods affirming that size, shape, and hydrophobicity are important global pocket descriptors for automatic druggability prediction. Nevertheless, the variety of pocket shapes and their flexibility upon ligand binding limit the automatic projection of druggable features onto descriptors. Incorporating local pocket properties is another step toward a reliable descriptor-based druggability prediction.

Pharmacophore alignment search tool: influence of the third dimension on text-based similarity searching.

Previously (Hähnke et al., J Comput Chem 2010, 31, 2810) we introduced the concept of nonlinear dimensionality reduction for canonization of two-dimensional layouts of molecular graphs as foundation for text-based similarity searching using our Pharmacophore Alignment Search Tool (PhAST), a ligand-based virtual screening method. Here we apply these methods to three-dimensional molecular conformations and investigate the impact of these additional degrees of freedom on virtual screening performance and assess differences in ranking behavior. Best-performing variants of PhAST are compared with 16 state-of-the-art screening methods with respect to significance estimates for differences in screening performance. We show that PhAST sorts new chemotypes on early ranks without sacrificing overall screening performance. We succeeded in combining PhAST with other virtual screening techniques by rank-based data fusion, significantly improving screening capabilities. We also present a parameterization of double dynamic programming for the problem of small molecule comparison, which allows for the calculation of structural similarity between compounds based on one-dimensional representations, opening the door to a holistic approach to molecule comparison based on textual representations.

A normal mode-based geometric simulation approach for exploring biologically relevant conformational transitions in proteins.

A three-step approach for multiscale modeling of protein conformational changes is presented that incorporates information about preferred directions of protein motions into a geometric simulation algorithm. The first two steps are based on a rigid cluster normal-mode analysis (RCNMA). Low-frequency normal modes are used in the third step (NMSim) to extend the recently introduced idea of constrained geometric simulations of diffusive motions in proteins by biasing backbone motions of the protein, whereas side-chain motions are biased toward favorable rotamer states. The generated structures are iteratively corrected regarding steric clashes and stereochemical constraint violations. The approach allows performing three simulation types: unbiased exploration of conformational space; pathway generation by a targeted simulation; and radius of gyration-guided simulation. When applied to a data set of proteins with experimentally observed conformational changes, conformational variabilities are reproduced very well for 4 out of 5 proteins that show domain motions, with correlation coefficients r > 0.70 and as high as r = 0.92 in the case of adenylate kinase. In 7 out of 8 cases, NMSim simulations starting from unbound structures are able to sample conformations that are similar (root-mean-square deviation = 1.0-3.1 Å) to ligand bound conformations. An NMSim generated pathway of conformational change of adenylate kinase correctly describes the sequence of domain closing. The NMSim approach is a computationally efficient alternative to molecular dynamics simulations for conformational sampling of proteins. The generated conformations and pathways of conformational transitions can serve as input to docking approaches or as starting points for more sophisticated sampling techniques.

Identifying biophysical assays and in silico properties that enrich for slow clearance in clinical-stage therapeutic antibodies.

Understanding the pharmacokinetic (PK) properties of a drug, such as clearance, is a crucial step for evaluating efficacy. The PK of therapeutic antibodies can be complex and is influenced by interactions with the target, Fc-receptors, anti-drug antibodies, and antibody intrinsic factors. A growing body of literature has linked biophysical properties of antibodies, particularly nonspecific-binding propensity, hydrophobicity and charged regions to rapid clearance in preclinical species and selected human PK studies. A clear understanding of the connection between biophysical properties and their impact on PK would allow for early selection and optimization of antibodies and reduce costly attrition during clinical trials due to sub-optimal human clearance. Due to the difficulty in obtaining large and unbiased human PK data, previous studies have focused mostly on preclinical PK. For this study, we obtained and curated the most comprehensive clinical PK dataset to date and calculated accurate estimates of linear clearance for 64 monoclonal antibodies ranging from investigational candidates in Phase 2 trials to marketed products. This allows for the first time a deep analysis of the influence of biophysical and sequence-based in silico properties directly on human clearance. We use statistical analysis and a Random Forest classifier to identify properties that have the greatest influence in our dataset. Our findings indicate that in vitro poly-specificity assay and in silico estimated isoelectric point can discriminate fast and slow clearing antibodies, extending previous observations on preclinical clearance. This provides a simple yet powerful approach to select antibodies with desirable PK during early-stage screening.

Short linear motif candidates in the cell entry system used by SARS-CoV-2 and their potential therapeutic implications.

The first reported receptor for SARS-CoV-2 on host cells was the angiotensin-converting enzyme 2 (ACE2). However, the viral spike protein also has an RGD motif, suggesting that cell surface integrins may be co-receptors. We examined the sequences of ACE2 and integrins with the Eukaryotic Linear Motif (ELM) resource and identified candidate short linear motifs (SLiMs) in their short, unstructured, cytosolic tails with potential roles in endocytosis, membrane dynamics, autophagy, cytoskeleton, and cell signaling. These SLiM candidates are highly conserved in vertebrates and may interact with the µ2 subunit of the endocytosis-associated AP2 adaptor complex, as well as with various protein domains (namely, I-BAR, LC3, PDZ, PTB, and SH2) found in human signaling and regulatory proteins. Several motifs overlap in the tail sequences, suggesting that they may act as molecular switches, such as in response to tyrosine phosphorylation status. Candidate LC3-interacting region (LIR) motifs are present in the tails of integrin ß3 and ACE2, suggesting that these proteins could directly recruit autophagy components. Our findings identify several molecular links and testable hypotheses that could uncover mechanisms of SARS-CoV-2 attachment, entry, and replication against which it may be possible to develop host-directed therapies that dampen viral infection and disease progression. Several of these SLiMs have now been validated to mediate the predicted peptide interactions.

Mechanism of praziquantel action at a parasitic flatworm ion channel.

Praziquantel (PZQ) is an essential medicine for treating parasitic flatworm infections such as schistosomiasis, which afflicts over 250 million people. However, PZQ is not universally effective, lacking activity against liver flukes of the Fasciola genus. The reason for this insensitivity is unclear, as the mechanism of PZQ action is unknown. Here, we use ligand- and target-based methods to demonstrate that PZQ activates a transient receptor potential melastatin ion channel (TRPMPZQ) in schistosomes by engaging a hydrophobic ligand binding pocket within the voltage sensor­like domain of the channel to cause calcium entry and worm paralysis. PZQ activates TRPMPZQ homologs in other PZQ-sensitive flukes, but not Fasciola hepatica. However, a single amino acid change in the F. hepatica TRPMPZQ binding pocket, to mimic schistosome TRPMPZQ, confers PZQ sensitivity. After decades of clinical use, the molecular basis of PZQ action at a druggable TRP channel is resolved.

Pharmacophore alignment search tool: Influence of canonical atom labeling on similarity searching.

Previously, (Hähnke et al., J Comput Chem 2009, 30, 761) we presented the Pharmacophore Alignment Search Tool (PhAST), a ligand-based virtual screening technique representing molecules as strings coding pharmacophoric features and comparing them by global pairwise sequence alignment. To guarantee unambiguity during the reduction of two-dimensional molecular graphs to one-dimensional strings, PhAST employs a graph canonization step. Here, we present the results of the comparison of 11 different algorithms for graph canonization with respect to their impact on virtual screening. Retrospective screenings of a drug-like data set were evaluated using the BEDROC metric, which yielded averaged values between 0.4 and 0.14 for the best-performing and worst-performing canonization technique. We compared five scoring schemes for the alignments and found preferred combinations of canonization algorithms and scoring functions. Finally, we introduce a performance index that helps prioritize canonization approaches without the need for extensive retrospective evaluation.