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.

N-aryl-piperidine-4-carboxamides as a novel class of potent inhibitors of MALT1 proteolytic activity.

Starting from a weak screening hit, potent and selective inhibitors of the MALT1 protease function were elaborated. Advanced compounds displayed high potency in biochemical and cellular assays. Compounds showed activity in a mechanistic Jurkat T cell activation assay as well as in the B-cell lymphoma line OCI-Ly3, which suggests potential use of MALT1 inhibitors in the treatment of autoimmune diseases as well as B-cell lymphomas with a dysregulated NF-κB pathway. Initially, rat pharmacokinetic properties of this compound series were dominated by very high clearance which could be linked to amide cleavage. Using a rat hepatocyte assay a good in vitro-in vivo correlation could be established which led to the identification of compounds with improved PK properties.

Dark chemical matter as a promising starting point for drug lead discovery.

High-throughput screening (HTS) is an integral part of early drug discovery. Herein, we focused on those small molecules in a screening collection that have never shown biological activity despite having been exhaustively tested in HTS assays. These compounds are referred to as 'dark chemical matter' (DCM). We quantified DCM, validated it in quality control experiments, described its physicochemical properties and mapped it into chemical space. Through analysis of prospective reporter-gene assay, gene expression and yeast chemogenomics experiments, we evaluated the potential of DCM to show biological activity in future screens. We demonstrated that, despite the apparent lack of activity, occasionally these compounds can result in potent hits with unique activity and clean safety profiles, which makes them valuable starting points for lead optimization efforts. Among the identified DCM hits was a new antifungal chemotype with strong activity against the pathogen Cryptococcus neoformans but little activity at targets relevant to human safety.

Discovery of New Binders for DCAF1, an Emerging Ligase Target in the Targeted Protein Degradation Field.

In this study, we describe the rapid identification of potent binders for the WD40 repeat domain (WDR) of DCAF1. This was achieved by two rounds of iterative focused screening of a small set of compounds selected on the basis of internal WDR domain knowledge followed by hit expansion. Subsequent structure-based design led to nanomolar potency binders with a clear exit vector enabling DCAF1-based bifunctional degrader exploration.

Bioactivity-guided mapping and navigation of chemical space.

The structure- and chemistry-based hierarchical organization of library scaffolds in tree-like arrangements provides a valid, intuitive means to map and navigate chemical space. We demonstrate that scaffold trees built using bioactivity as the key selection criterion for structural simplification during tree construction allow efficient and intuitive mapping, visualization and navigation of the chemical space defined by a given library, which in turn allows correlation of this chemical space with the investigated bioactivity and further compound design. Brachiation along the branches of such trees from structurally complex to simple scaffolds with retained yet varying bioactivity is feasible at high frequency for the five major pharmaceutically relevant target classes and allows for the identification of new inhibitor types for a given target. We provide proof of principle by identifying new active scaffolds for 5-lipoxygenase and the estrogen receptor ERalpha.

Mining for bioactive scaffolds with scaffold networks: improved compound set enrichment from primary screening data.

Identification of meaningful chemical patterns in the increasing amounts of high-throughput-generated bioactivity data available today is an increasingly important challenge for successful drug discovery. Herein, we present the scaffold network as a novel approach for mapping and navigation of chemical and biological space. A scaffold network represents the chemical space of a library of molecules consisting of all molecular scaffolds and smaller "parent" scaffolds generated therefrom by the pruning of rings, effectively leading to a network of common scaffold substructure relationships. This algorithm provides an extension of the scaffold tree algorithm that, instead of a network, generates a tree relationship between a heuristically rule-based selected subset of parent scaffolds. The approach was evaluated for the identification of statistically significantly active scaffolds from primary screening data for which the scaffold tree approach has already been shown to be successful. Because of the exhaustive enumeration of smaller scaffolds and the full enumeration of relationships between them, about twice as many statistically significantly active scaffolds were identified compared to the scaffold-tree-based approach. We suggest visualizing scaffold networks as islands of active scaffolds.

MultiMCS: a fast algorithm for the maximum common substructure problem on multiple molecules.

Several efficient correspondence graph-based algorithms for determining the maximum common substructure (MCS) of a pair of molecules have been published in the literature. The extension of the problem to three or more molecules is however nontrivial; heuristics used to increase the efficiency in the two-molecule case are either inapplicable to the many-molecule case or do not provide significant speedups. Our specific algorithmic contribution is two-fold. First, we show how the correspondence graph approach for the two-molecule case can be generalized to obtain an algorithm that is guaranteed to find the optimum connected MCS of multiple molecules, and that runs fast on most families of molecules using a new divide-and-conquer strategy that has hitherto not been reported in this context. Second, we provide a characterization of those compound families for which the algorithm might run slowly, along with a heuristic for speeding up computations on these families. We also extend the above algorithm to a heuristic algorithm to find the disconnected MCS of multiple molecules and to an algorithm for clustering molecules into groups, with each group sharing a substantial MCS. Our methods are flexible in that they provide exquisite control on various matching criteria used to define a common substructure.

Splitting chemical structure data sets for federated privacy-preserving machine learning.

With the increase in applications of machine learning methods in drug design and related fields, the challenge of designing sound test sets becomes more and more prominent. The goal of this challenge is to have a realistic split of chemical structures (compounds) between training, validation and test set such that the performance on the test set is meaningful to infer the performance in a prospective application. This challenge is by its own very interesting and relevant, but is even more complex in a federated machine learning approach where multiple partners jointly train a model under privacy-preserving conditions where chemical structures must not be shared between the different participating parties. In this work we discuss three methods which provide a splitting of a data set and are applicable in a federated privacy-preserving setting, namely: a. locality-sensitive hashing (LSH), b. sphere exclusion clustering, c. scaffold-based binning (scaffold network). For evaluation of these splitting methods we consider the following quality criteria (compared to random splitting): bias in prediction performance, classification label and data imbalance, similarity distance between the test and training set compounds. The main findings of the paper are a. both sphere exclusion clustering and scaffold-based binning result in high quality splitting of the data sets, b. in terms of compute costs sphere exclusion clustering is very expensive in the case of federated privacy-preserving setting.

Compound set enrichment: a novel approach to analysis of primary HTS data.

The main goal of high-throughput screening (HTS) is to identify active chemical series rather than just individual active compounds. In light of this goal, a new method (called compound set enrichment) to identify active chemical series from primary screening data is proposed. The method employs the scaffold tree compound classification in conjunction with the Kolmogorov-Smirnov statistic to assess the overall activity of a compound scaffold. The application of this method to seven PubChem data sets (containing between 9389 and 263679 molecules) is presented, and the ability of this method to identify compound classes with only weakly active compounds (potentially latent hits) is demonstrated. The analysis presented here shows how methods based on an activity cutoff can distort activity information, leading to the incorrect activity assignment of compound series. These results suggest that this method might have utility in the rational selection of active classes of compounds (and not just individual active compounds) for followup and validation.

Gene-signature-derived IC50s/EC50s reflect the potency of causative upstream targets and downstream phenotypes.

Multiplexed gene-signature-based phenotypic assays are increasingly used for the identification and profiling of small molecule-tool compounds and drugs. Here we introduce a method (provided as R-package) for the quantification of the dose-response potency of a gene-signature as EC50 and IC50 values. Two signaling pathways were used as models to validate our methods: beta-adrenergic agonistic activity on cAMP generation (dedicated dataset generated for this study) and EGFR inhibitory effect on cancer cell viability. In both cases, potencies derived from multi-gene expression data were highly correlated with orthogonal potencies derived from cAMP and cell growth readouts, and superior to potencies derived from single individual genes. Based on our results we propose gene-signature potencies as a novel valid alternative for the quantitative prioritization, optimization and development of novel drugs.

Discovery of Potent, Highly Selective, and In Vivo Efficacious, Allosteric MALT1 Inhibitors by Iterative Scaffold Morphing.

MALT1 plays a central role in immune cell activation by transducing NF-κB signaling, and its proteolytic activity represents a key node for therapeutic intervention. Two cycles of scaffold morphing of a high-throughput biochemical screening hit resulted in the discovery of MLT-231, which enabled the successful pharmacological validation of MALT1 allosteric inhibition in preclinical models of humoral immune responses and B-cell lymphomas. Herein, we report the structural activity relationships (SARs) and analysis of the physicochemical properties of a pyrazolopyrimidine-derived compound series. In human T-cells and B-cell lymphoma lines, MLT-231 potently and selectively inhibits the proteolytic activity of MALT1 in NF-κB-dependent assays. Both in vitro and in vivo profiling of MLT-231 support further optimization of this in vivo tool compound toward preclinical characterization.

Evolution of Novartis' Small Molecule Screening Deck Design.

This article summarizes the evolution of the screening deck at the Novartis Institutes for BioMedical Research (NIBR). Historically, the screening deck was an assembly of all available compounds. In 2015, we designed a first deck to facilitate access to diverse subsets with optimized properties. We allocated the compounds as plated subsets on a 2D grid with property based ranking in one dimension and increasing structural redundancy in the other. The learnings from the 2015 screening deck were applied to the design of a next generation in 2019. We found that using traditional leadlikeness criteria (mainly MW, clogP) reduces the hit rates of attractive chemical starting points in subset screening. Consequently, the 2019 deck relies on solubility and permeability to select preferred compounds. The 2019 design also uses NIBR's experimental assay data and inferred biological activity profiles in addition to structural diversity to define redundancy across the compound sets.

Systematic Chemogenetic Library Assembly.

Chemogenetic libraries, collections of well-defined chemical probes, provide tremendous value to biomedical research but require substantial effort to ensure diversity as well as quality of the contents. We have assembled a chemogenetic library by data mining and crowdsourcing institutional expertise. We are sharing our approach, lessons learned, and disclosing our current collection of 4,185 compounds with their primary annotated gene targets (https://github.com/Novartis/MoaBox). This physical collection is regularly updated and used broadly both within Novartis and in collaboration with external partners.

Cheminformatics analysis of natural products: lessons from nature inspiring the design of new drugs.

Natural products (NPs) have evolved over a very long natural selection process to form optimal interactions with biological macromolecules. NPs are therefore an extremely useful source of inspiration for the design of new drugs. In the present study we report the results of a cheminformatics analysis of more than 130,000 NP structures. The physicochemical properties of NPs and their typical structural features are compared to those of bioactive molecules and average organic molecules. The relationship between the structure of NPs and the type of organism from which they have come has also been analyzed. The aim of this study was to identify those properties and structural features which are typical for NPs and discriminate this class of molecules from common synthetic molecules, with the ultimate goal being to provide a guide for the design of novel NP-like bioactive structures. Hopefully the results of this analysis help to eliminate the old myth about NPs as being 'too complex' or having 'bad properties', as well as help us to focus on these areas of NP structural space which are essential for biological activity, taking advantage of the process of natural selection over billions of years to guide us to new and as yet unexplored areas of the Chemical Structure Universe.

Identification of SPPL2a Inhibitors by Multiparametric Analysis of a High-Content Ultra-High-Throughput Screen.

The intramembrane protease signal peptide peptidase-like 2a (SPPL2a) is a potential drug target for the treatment of autoimmune diseases due to an essential role in B cells and dendritic cells. To screen a library of 1.4 million compounds for inhibitors of SPPL2a, we developed an imaging assay detecting nuclear translocation of the proteolytically released cytosolic substrate fragment. The state-of-the-art hit calling approach based on nuclear translocation resulted in numerous false-positive hits, mainly interrupting intracellular protein trafficking. To filter the false positives, we extracted 340 image-based readouts and developed a novel multiparametric analysis method that successfully triaged the primary hit list. The identified scaffolds were validated by demonstrating activity on endogenous SPPL2a and substrate CD74/p8 in B cells. The multiparametric analysis discovered diverse cellular phenotypes and provided profiles for the whole library. The principle of the presented imaging assay, the screening strategy, and multiparametric analysis are potentially applicable in future screening campaigns.

Complex molecules: do they add value?

The concept of complexity in chemistry has received much interest in the research community. Various measures to assess molecular complexity have been published, ranging from abstract complexity definitions to very specific application-oriented definitions. In this article we focus on molecular complexity in relation to biological activity. Connectivity and feature-based structural descriptors have been evaluated with reference to their potential as complexity measures. Our goal was to discuss the potential of the complexity concept to support the drug discovery process, helping to design suitable lead candidates. The studies have shown that highly active compounds, on average, are more complex than inactive compounds. However, complexity must be balanced with other molecular properties because more complex molecules have a higher probability to exhibit pharmacokinetic problems.

Quest for the rings. In silico exploration of ring universe to identify novel bioactive heteroaromatic scaffolds.

Bioactive molecules only contain a relatively limited number of unique ring types. To identify those ring properties and structural characteristics that are necessary for biological activity, a large virtual library of nearly 600 000 heteroaromatic scaffolds was created and characterized by calculated properties, including structural features, bioavailability descriptors, and quantum chemical parameters. A self-organizing neural network was used to cluster these scaffolds and to identify properties that best characterize bioactive ring systems. The analysis shows that bioactivity is very sparsely distributed within the scaffold property and structural space, forming only several relatively small, well-defined "bioactivity islands". Various possible applications of a large database of rings with calculated properties and bioactivity scores in the drug design and discovery process are discussed, including virtual screening, support for the design of combinatorial libraries, bioisosteric design, and scaffold hopping.

A chemoinformatics analysis of hit lists obtained from high-throughput affinity-selection screening.

The high-throughput affinity-selection screening platform SpeedScreen was recently reported by the Novartis Institutes for BioMedical Research as a homogeneous, label-free screening technology with mass-spectrometry readout. SpeedScreen relies on the screening of compound mixtures with various target proteins and uses fast size-exclusion chromatography to separate target-bound from unbound substances. After disintegration of the target-binder complex, the binder molecules are identified by their molecular masses using liquid chromatography/mass spectrometry. The authors report an analysis of the molecular properties of hits obtained with SpeedScreen on 26 targets screened within the past few years at Novartis using this technology. Affinity-based SpeedScreen is a robust high-throughput screening technology that does not accumulate frequent hitters or potential covalent binders. The hits are representative of the most commonly identified scaffold classes observed for known drugs. Validated SpeedScreen hits tend to be enriched on more lipophilic and larger-molecular-weight compounds compared to the whole library. The potential for a reduced SpeedScreen screening set to be used in case only limited protein quantities are available is evaluated. Such a reduced compound set should also maximize the coverage of the high-performing regions of the chemical property and class spaces; chemoinformatics methods including genetic algorithms and divisive K-means clustering are used for this aim.

Library design for fragment based screening.

According to Hann's model of molecular complexity an increased probability of detection binding to a target protein can be expected when small, low complex molecular fragments are screened with high sensitivity instead of full-sized ligands with lower sensitivity. Analysis of the HTS summary data of Novartis and comparison with NMR screening results obtained on generic fragment libraries indicate this expectation to be true with hitrates of 0.001% - 0.151% observed in the identification of ligands with an IC(50) threshold in the micromolar range in an HTS setup and hitrates above or equal to 3% observed in NMR screening of fragments with an affinity threshold in the millimolar range. It is however necessary to keep in mind that the sets of target studied were not identical for both method and the experience in NMR screening is too limited for a final conclusion. The term hitrate as used here reflects only the success rate in the observation of ligand binding event. It must not be confused with the overall success rate of fragment and high throughput screening in the lead finding process, which can be entirely different, since the steps required to follow-up a ligand binding event to a lead are different for both methods. A survey of fragment-based lead discovery case studies given in the literature shows that in approximately half of the cases the initial hit fragment was discovered by screening a generic library, whereas in the other cases some knowledge about an initial ligands or the protein binding site has been used, whereas systematic virtual screening of fragment databases has been only rarely reported. As comparatively high hitrates were obtained, further consideration to optimize the generic fragment screening library were directed to the chemical tractability of the fragment. As several functional groups preferred by chemists for modification and linking of the fragments are also preferentially involved in interactions between the fragments and the target protein, a set of screening fragments was derived from chemical building blocks by masking its linker group by a chemical transformation which can be later on used in the chemical follow-up of the fragment hit. For example primary amines can be masked as acetamides. If the screening fragment is active the related building block can then be used for synthesis of a follow-up library.

Key aspects of the Novartis compound collection enhancement project for the compilation of a comprehensive chemogenomics drug discovery screening collection.

The NIBR (Novartis Institutes for BioMedical Research) compound collection enrichment and enhancement project integrates corporate internal combinatorial compound synthesis and external compound acquisition activities in order to build up a comprehensive screening collection for a modern drug discovery organization. The main purpose of the screening collection is to supply the Novartis drug discovery pipeline with hit-to-lead compounds for today's and the future's portfolio of drug discovery programs, and to provide tool compounds for the chemogenomics investigation of novel biological pathways and circuits. As such, it integrates designed focused and diversity-based compound sets from the synthetic and natural paradigms able to cope with druggable and currently deemed undruggable targets and molecular interaction modes. Herein, we will summarize together with new trends published in the literature, scientific challenges faced and key approaches taken at NIBR to match the chemical and biological spaces.