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.

Developing a Kinase-Specific Target Selection Method Using a Structure-Based Machine Learning Approach.

INTRODUCTION: Despite recent advances in the drug discovery field, developing selective kinase inhibitors remains a complicated issue for a number of reasons, one of which is that there are striking structural similarities in the ATP-binding pockets of kinases. OBJECTIVE: To address this problem, we have designed a machine learning model utilizing various structure-based and energy-based descriptors to better characterize protein-ligand interactions. METHODS: In this work, we use a dataset of 104 human kinases with available PDB structures and experimental activity data against 1202 small-molecule compounds from the PubChem BioAssay dataset "Navigating the Kinome". We propose structure-based interaction descriptors to build activity predicting machine learning model. RESULTS AND DISCUSSION: We report a ligand-oriented computational method for accurate kinase target prioritizing. Our method shows high accuracy compared to similar structure-based activity prediction methods, and more importantly shows the same prediction accuracy when tested on the special set of structurally remote compounds, showing that it is unbiased to ligand structural similarity in the training set data. We hope that our approach will be useful for the development of novel highly selective kinase inhibitors.

Systematic search for structural motifs of peptide binding to double-stranded DNA.

A large variety of short biologically active peptides possesses antioxidant, antibacterial, antitumour, anti-ageing and anti-inflammatory activity, involved in the regulation of neuro-immuno-endocrine system functions, cell apoptosis, proliferation and differentiation. Therefore, the mechanisms of their biological activity are attracting increasing attention not only in modern molecular biology, biochemistry and biophysics, but also in pharmacology and medicine. In this work, we systematically analysed the ability of dipeptides (all possible combinations of the 20 standard amino acids) to bind all possible combinations of tetra-nucleotides in the central part of dsDNA in the classic B-form using molecular docking and molecular dynamics. The vast majority of the dipeptides were found to be unable to bind dsDNA. However, we were able to identify 57 low-energy dipeptide complexes with peptide-dsDNA possessing high selectivity for DNA binding. The analysis of the dsDNA complexes with dipeptides with free and blocked N- and C-terminus showed that selective peptide binding to dsDNA can increase dramatically with the peptide length.

Prediction of the secondary structure of short DNA aptamers.

Aptamers have a spectrum of applications in biotechnology and drug design, because of the relative simplicity of experimental protocols and advantages of stability and specificity associated with their structural properties. However, to understand the structure-function relationships of aptamers, robust structure modeling tools are necessary. Several such tools have been developed and extensively tested, although most of them target various forms of biological RNA. In this study, we tested the performance of three tools in application to DNA aptamers, since DNA aptamers are the focus of many studies, particularly in drug discovery. We demonstrated that in most cases, the secondary structure of DNA can be reconstructed with acceptable accuracy by at least one of the three tools tested (Mfold, RNAfold, and CentroidFold), although the G-quadruplex motif found in many of the DNA aptamer structures complicates the prediction, as well as the pseudoknot interaction. This problem should be addressed more carefully to improve prediction accuracy.

Human long intrinsically disordered protein regions are frequent targets of positive selection.

Intrinsically disordered regions occur frequently in proteins and are characterized by a lack of a well-defined three-dimensional structure. Although these regions do not show a higher order of structural organization, they are known to be functionally important. Disordered regions are rapidly evolving, largely attributed to relaxed purifying selection and an increased role of genetic drift. It has also been suggested that positive selection might contribute to their rapid diversification. However, for our own species, it is currently unknown whether positive selection has played a role during the evolution of these protein regions. Here, we address this question by investigating the evolutionary pattern of more than 6600 human proteins with intrinsically disordered regions and their ordered counterparts. Our comparative approach with data from more than 90 mammalian genomes uses a priori knowledge of disordered protein regions, and we show that this increases the power to detect positive selection by an order of magnitude. We can confirm that human intrinsically disordered regions evolve more rapidly, not only within humans but also across the entire mammalian phylogeny. They have, however, experienced substantial evolutionary constraint, hinting at their fundamental functional importance. We find compelling evidence that disordered protein regions are frequent targets of positive selection and estimate that the relative rate of adaptive substitutions differs fourfold between disordered and ordered protein regions in humans. Our results suggest that disordered protein regions are important targets of genetic innovation and that the contribution of positive selection in these regions is more pronounced than in other protein parts.

AquaBridge: A novel method for systematic search of structural water molecules within the protein active sites.

We have developed a novel method for calculation of the water bridges that can be formed in the active sites of proteins in the absence or in the presence of small-molecule ligands. We tested its efficiency on a representative set of human ATP-binding proteins, and show that the docking accuracy of ligands can be substantially improved when water bridges are included in the modeling of protein-ligand interactions. Our analysis of binding pocket hydration can be a useful addition to the in silico approaches of Drug Design.

Lytic water dynamics reveal evolutionarily conserved mechanisms of ATP hydrolysis by TIP49 AAA+ ATPases.

Eukaryotic TIP49a (Pontin) and TIP49b (Reptin) AAA+ ATPases play essential roles in key cellular processes. How their weak ATPase activity contributes to their important functions remains largely unknown and difficult to analyze because of the divergent properties of TIP49a and TIP49b proteins and of their homo- and hetero-oligomeric assemblies. To circumvent these complexities, we have analyzed the single ancient TIP49 ortholog found in the archaeon Methanopyrus kandleri (mkTIP49). All-atom homology modeling and molecular dynamics simulations validated by biochemical assays reveal highly conserved organizational principles and identify key residues for ATP hydrolysis. An unanticipated crosstalk between Walker B and Sensor I motifs impacts the dynamics of water molecules and highlights a critical role of trans-acting aspartates in the lytic water activation step that is essential for the associative mechanism of ATP hydrolysis.