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.

Development of an Automatic Pipeline for Participation in the CELPP Challenge.

The prediction of how a ligand binds to its target is an essential step for Structure-Based Drug Design (SBDD) methods. Molecular docking is a standard tool to predict the binding mode of a ligand to its macromolecular receptor and to quantify their mutual complementarity, with multiple applications in drug design. However, docking programs do not always find correct solutions, either because they are not sampled or due to inaccuracies in the scoring functions. Quantifying the docking performance in real scenarios is essential to understanding their limitations, managing expectations and guiding future developments. Here, we present a fully automated pipeline for pose prediction validated by participating in the Continuous Evaluation of Ligand Pose Prediction (CELPP) Challenge. Acknowledging the intrinsic limitations of the docking method, we devised a strategy to automatically mine and exploit pre-existing data, defining-whenever possible-empirical restraints to guide the docking process. We prove that the pipeline is able to generate predictions for most of the proposed targets as well as obtain poses with low RMSD values when compared to the crystal structure. All things considered, our pipeline highlights some major challenges in the automatic prediction of protein-ligand complexes, which will be addressed in future versions of the pipeline.

Accurate Representation of Protein-Ligand Structural Diversity in the Protein Data Bank (PDB).

The number of available protein structures in the Protein Data Bank (PDB) has considerably increased in recent years. Thanks to the growth of structures and complexes, numerous large-scale studies have been done in various research areas, e.g., protein-protein, protein-DNA, or in drug discovery. While protein redundancy was only simply managed using simple protein sequence identity threshold, the similarity of protein-ligand complexes should also be considered from a structural perspective. Hence, the protein-ligand duplicates in the PDB are widely known, but were never quantitatively assessed, as they are quite complex to analyze and compare. Here, we present a specific clustering of protein-ligand structures to avoid bias found in different studies. The methodology is based on binding site superposition, and a combination of weighted Root Mean Square Deviation (RMSD) assessment and hierarchical clustering. Repeated structures of proteins of interest are highlighted and only representative conformations were conserved for a non-biased view of protein distribution. Three types of cases are described based on the number of distinct conformations identified for each complex. Defining these categories decreases by 3.84-fold the number of complexes, and offers more refined results compared to a protein sequence-based method. Widely distinct conformations were analyzed using normalized B-factors. Furthermore, a non-redundant dataset was generated for future molecular interactions analysis or virtual screening studies.

rDock: a fast, versatile and open source program for docking ligands to proteins and nucleic acids.

Identification of chemical compounds with specific biological activities is an important step in both chemical biology and drug discovery. When the structure of the intended target is available, one approach is to use molecular docking programs to assess the chemical complementarity of small molecules with the target; such calculations provide a qualitative measure of affinity that can be used in virtual screening (VS) to rank order a list of compounds according to their potential to be active. rDock is a molecular docking program developed at Vernalis for high-throughput VS (HTVS) applications. Evolved from RiboDock, the program can be used against proteins and nucleic acids, is designed to be computationally very efficient and allows the user to incorporate additional constraints and information as a bias to guide docking. This article provides an overview of the program structure and features and compares rDock to two reference programs, AutoDock Vina (open source) and Schrödinger's Glide (commercial). In terms of computational speed for VS, rDock is faster than Vina and comparable to Glide. For binding mode prediction, rDock and Vina are superior to Glide. The VS performance of rDock is significantly better than Vina, but inferior to Glide for most systems unless pharmacophore constraints are used; in that case rDock and Glide are of equal performance. The program is released under the Lesser General Public License and is freely available for download, together with the manuals, example files and the complete test sets, at http://rdock.sourceforge.net/

Dynamics of hERG closure allow novel insights into hERG blocking by small molecules.

Today, drug discovery routinely uses experimental assays to determine very early if a lead compound can yield certain types of off-target activity. Among such off targets is hERG. The ion channel plays a primordial role in membrane repolarization and altering its activity can cause severe heart arrhythmia and sudden death. Despite routine tests for hERG activity, rather little information is available for helping medicinal chemists and molecular modelers to rationally circumvent hERG activity. In this article novel insights into the dynamics of hERG channel closure are described. Notably, helical pairwise closure movements have been observed. Implications and relations to hERG inactivation are presented. Based on these dynamics novel insights on hERG blocker placement are presented, compared to literature, and discussed. Last, new evidence for horizontal ligand positioning is shown in light of former studies on hERG blockers.

MDpocket: open-source cavity detection and characterization on molecular dynamics trajectories.

MOTIVATION: A variety of pocket detection algorithms are now freely or commercially available to the scientific community for the analysis of static protein structures. However, since proteins are dynamic entities, enhancing the capabilities of these programs for the straightforward detection and characterization of cavities taking into account protein conformational ensembles should be valuable for capturing the plasticity of pockets, and therefore allow gaining insight into structure-function relationships. RESULTS: This article describes a new method, called MDpocket, providing a fast, free and open-source tool for tracking small molecule binding sites and gas migration pathways on molecular dynamics (MDs) trajectories or other conformational ensembles. MDpocket is based on the fpocket cavity detection algorithm and a valuable contribution to existing analysis tools. The capabilities of MDpocket are illustrated for three relevant cases: (i) the detection of transient subpockets using an ensemble of crystal structures of HSP90; (ii) the detection of known xenon binding sites and migration pathways in myoglobin; and (iii) the identification of suitable pockets for molecular docking in P38 Map kinase. AVAILABILITY: MDpocket is free and open-source software and can be downloaded at http://fpocket.sourceforge.net. CONTACT: pschmidtke@ub.edu SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

fpocket: online tools for protein ensemble pocket detection and tracking.

Computational small-molecule binding site detection has several important applications in the biomedical field. Notable interests are the identification of cavities for structure-based drug discovery or functional annotation of structures. fpocket is a small-molecule pocket detection program, relying on the geometric alpha-sphere theory. The fpocket web server allows: (i) candidate pocket detection--fpocket; (ii) pocket tracking during molecular dynamics, in order to provide insights into pocket dynamics--mdpocket; and (iii) a transposition of mdpocket to the combined analysis of homologous structures--hpocket. These complementary online tools allow to tackle various questions related to the identification and annotation of functional and allosteric sites, transient pockets and pocket preservation within evolution of structural families. The server and documentation are freely available at http://bioserv.rpbs.univ-paris-diderot.fr/fpocket.

Extracting Atomic Contributions to Binding Free Energy Using Molecular Dynamics Simulations with Mixed Solvents (MDmix).

BACKGROUND: Mixed solvents MD (MDmix) simulations have proved to be a useful and increasingly accepted technique with several applications in structure-based drug discovery. One of the assumptions behind the methodology is the transferability of free energy values from the simulated cosolvent molecules to larger drug-like molecules. However, the binding free energy maps (ΔGbind) calculated for the different moieties of the cosolvent molecules (e.g. a hydroxyl map for the ethanol) are largely influenced by the rest of the solvent molecule and do not reflect the intrinsic affinity of the moiety in question. As such, they are hardly transferable to different molecules. METHOD: To achieve transferable energies, we present here a method for decomposing the molecular binding free energy into accurate atomic contributions. RESULT: We demonstrate with two qualitative visual examples how the corrected energy maps better match known binding hotspots and how they can reveal hidden hotspots with actual drug design potential. CONCLUSION: Atomic decomposition of binding free energies derived from MDmix simulations provides transferable and quantitative binding free energy maps.

Shielded hydrogen bonds as structural determinants of binding kinetics: application in drug design.

Time scale control of molecular interactions is an essential part of biochemical systems, but very little is known about the structural factors governing the kinetics of molecular recognition. In drug design, the lifetime of drug-target complexes is a major determinant of pharmacological effects but the absence of structure-kinetic relationships precludes rational optimization of this property. Here we show that almost buried polar atoms--a common feature on protein binding sites--tend to form hydrogen bonds that are shielded from water. Formation and rupture of this type of hydrogen bonds involves an energetically penalized transition state because it occurs asynchronously with dehydration/rehydration. In consequence, water-shielded hydrogen bonds are exchanged at slower rates. Occurrence of this phenomenon can be anticipated from simple structural analysis, affording a novel tool to interpret and predict structure-kinetics relationships. The validity of this principle has been investigated on two pairs of Hsp90 inhibitors for which detailed thermodynamic and kinetic data has been experimentally determined. The agreement between macroscopic observables and molecular simulations confirms the role of water-shielded hydrogen bonds as kinetic traps and illustrates how our finding could be used as an aid in structure-based drug discovery.

Large-scale comparison of four binding site detection algorithms.

A large-scale evaluation and comparison of four cavity detection algorithms was carried out. The algorithms SiteFinder, fpocket, PocketFinder, and SiteMap were evaluated on a protein test set containing 5416 protein-ligand complexes and 9900 apo forms, corresponding to a subset of the set used earlier for benchmarking the PocketFinder algorithm. For the holo structures, all four algorithms correctly identified a similar amount of pockets (around 95%). SiteFinder, using optimized parameters, SiteMap, and fpocket showed similar pocket ranking performance, which was defined by ranking the correct binding site on rank 1 of the predictions or within the first 5 ranks of the predictions. On the apo structures, PocketFinder especially and also SiteFinder (optimized parameters) performed best, identifying 96% and 84% of all binding sites, respectively. The fpocket program predicts binding sites most accurately among the algorithms evaluated here. SiteFinder needed an average calculation time of 1.6 s compared with 2 min for SiteMap and around 2 s for fpocket.