Finton Sirockin

Scaffold hopping-the design of novel scaffolds for existing lead candidates-is a multifaceted and nontrivial task, for medicinal chemists and computational approaches alike. Generative reinforcement learning can iteratively optimize desirable properties of de novo designs, thereby offering opportunities to accelerate scaffold hopping. Current approaches confine the generation to a predefined molecular substructure (e.g., a linker or scaffold) for scaffold hopping. This confined generation may limit the exploration of the chemical space and require intricate molecule (dis)assembly rules. In this work, we aim to advance reinforcement learning for scaffold hopping, by allowing "unconstrained", full-molecule generation. This is achieved via the RuSH (Reinforcement Learning for Unconstrained Scaffold Hopping) approach. RuSH steers the generation toward the design of full molecules having a high three-dimensional and pharmacophore similarity to a reference molecule, but low scaffold similarity. In this first study, we show the flexibility and effectiveness of RuSH in exploring analogs of known scaffold-hops and in designing scaffold-hopping candidates that match known binding mechanisms. Finally, the comparison between RuSH and two established methods highlights the benefit of its unconstrained molecule generation to systematically achieve scaffold diversity while preserving optimal three-dimensional properties.

Scaffold hopping – the design of novel scaffolds for existing lead candidates – is a multi-faceted and non-trivial task, for medicinal chemists and computational approaches alike. Generative reinforcement learning can iteratively optimize desirable properties of de novo designs, thereby offering opportunities to accelerate scaffold hopping. Current approaches confine the generation to a pre-defined molecular substructure (e.g., a linker or scaffold) for scaffold hopping. This confined generation may limit the exploration of the chemical space, and require intricate molecule (dis)assembly rules. In this work, we aim to advance reinforcement learning for scaffold hopping, by allowing ‘unconstrained’, full-molecule generation. This is achieved via the RuSH (Reinforcement Learning for Unconstrained Scaffold Hopping) approach. RuSH steers the generation towards the design of full molecules having a high three-dimensional and pharmacophore similarity to a reference molecule, but low scaffold similarity. In this first study, we show the flexibility and effectiveness of RuSH in exploring analogs of known scaffold-hops and in designing scaffold-hopping candidates that match known binding mechanisms. Finally, the comparison between RuSH and two established methods highlights the benefit of its unconstrained molecule generation to systematically achieve scaffold diversity while preserving optimal three-dimensional properties.

Chemical structure optimization is a vital part of early drug discovery projects. Starting with compounds that show activity on the target of interest, the chemical structures are subsequently optimized toward a development candidate (DC) molecule with the best chances of clinical success. However, the DCs in the context of such optimization programs, as well as detailed characterization of major limiting factors, have not been investigated in detail so far. Here, we report an analysis of the historical DC molecules at Novartis since 2005 in the context of their optimization projects. Mapping the DCs into their respective chemical optimization series, we find that these tend to be synthesized rather early in a substantial number of cases. Further analysis of structural properties, ADMET, and potency-related readouts revealed that DC compounds tend to be generally significantly smaller, more permeable, and have higher ligand efficiency than other compounds sent to in vivo PK studies, which we also show for compounds from the same chemical series. Although this might seem obvious to most practitioners in medicinal chemistry, for all of these properties, we could show that they tend to evolve in an undesired direction during structure optimization. This highlights the difficulty of successfully translating our knowledge to medicinal chemistry optimizations.

Early in silico assessment of the potential of a series of compounds to deliver a drug is one of the major challenges in computer-assisted drug design. The goal is to identify the right chemical series of compounds out of a large chemical space to then subsequently prioritize the molecules with the highest potential to become a drug. Although multiple approaches to assess compounds have been developed over decades, the quality of these predictors is often not good enough and compounds that agree with the respective estimates are not necessarily druglike. Here, we report a novel deep learning approach that leverages large-scale predictions of ∼100 ADMET assays to assess the potential of a compound to become a relevant drug candidate. The resulting score, which we termed bPK score, substantially outperforms previous approaches and showed strong discriminative performance on data sets where previous approaches did not.

Machine-learning and deep-learning models have been extensively used in cheminformatics to predict molecular properties, to reduce the need for direct measurements, and to accelerate compound prioritization. However, different setups and frameworks and the large number of molecular representations make it difficult to properly evaluate, reproduce, and compare them. Here we present a new PREdictive modeling FramEwoRk for molecular discovery (PREFER), written in Python (version 3.7.7) and based on AutoSklearn (version 0.14.7), that allows comparison between different molecular representations and common machine-learning models. We provide an overview of the design of our framework and show exemplary use cases and results of several representation-model combinations on diverse data sets, both public and in-house. Finally, we discuss the use of PREFER on small data sets. The code of the framework is freely available on GitHub.

The accurate description of protein binding sites is essential to the determination of similarity and the application of machine learning methods to relate the binding sites to observed functions. This work describes CAVIAR, a new open source tool for generating descriptors for binding sites, using protein structures in PDB and mmCIF format as well as trajectory frames from molecular dynamics simulations as input. The applicability of CAVIAR descriptors is showcased by computing machine learning predictions of binding site ligandability. The method can also automatically assign subcavities, even in the absence of a bound ligand. The defined subpockets mimic the empirical definitions used in medicinal chemistry projects. It is shown that the experimental binding affinity scales relatively well with the number of subcavities filled by the ligand, with compounds binding to more than three subcavities having nanomolar or better affinities to the target. The CAVIAR descriptors and methods can be used in any machine learning-based investigations of problems involving binding sites, from protein engineering to hit identification. The full software code is available on GitHub and a conda package is hosted on Anaconda cloud.

Motivation. The detection of small molecules binding sites in proteins is central to structure-based drug design and chemical biology. Many tools were developed in the last 40 years, but few of them are still available in 2020, open-source, and suitable for the analysis of large databases or for the integration in automatic workflows. No software can characterize subpockets solely with the information of the protein structure, a pivotal concept in fragment-based drug design. Results. CAVIAR is a new open source tool for protein cav ity i dentification a nd r ationalization, supporting PDB and mmCIF files as well as DCD trajectories from molecular dynamics simulations. The protein structure serves as input for automatic cavity detection and computation of properties, including ligandability. A subcavity segmentation algorithm decomposes binding sites into subpockets without requiring the presence of a ligand. The defined subpockets mimick the empirical definitions of subpockets in medicinal chemistry projects. A tool like CAVIAR may be valuable to support chemical biology, medicinal chemistry and ligand identification efforts. Our analysis of the PDB shows that liganded cavities tend to be bigger, more hydrophobic and more complex than apo cavities. Moreover, in line with the paradigm of fragment-based drug design, the binding affinity scales relatively well with the number of subcavities filled by the ligand. Compounds binding to more than three subcavities are mostly in the nanomolar or better range of affinities to their target. Availability and implementation. Installation notes, user manual and support for CAVIAR are available at https://jr-marchand.github.io/caviar/ . The CAVIAR GUI and CAVIAR command line tool are available on GitHub at https://github.com/jr-marchand/caviar and a conda package is hosted on Anaconda cloud at https://anaconda.org/jr-marchand/caviar . The software suite is free and all of the source code is available under a permissive MIT license. The lists of PDB files used for validation, as well as the results of subpocket decomposition with CAVIAR and DoGSite are hosted on GitHub at https://github.com/jr-marchand/caviar/tree/master/validation_sets . Contact: jean-remy.marchand@novartis.com ; finton.sirockin@novartis.com

The alternative pathway (AP) of the complement system is a key contributor to the pathogenesis of several human diseases including age-related macular degeneration, paroxysmal nocturnal hemoglobinuria (PNH), atypical hemolytic uremic syndrome (aHUS), and various glomerular diseases. The serine protease factor B (FB) is a key node in the AP and is integral to the formation of C3 and C5 convertase. Despite the prominent role of FB in the AP, selective orally bioavailable inhibitors, beyond our own efforts, have not been reported previously. Herein we describe in more detail our efforts to identify FB inhibitors by high-throughput screening (HTS) and leveraging insights from several X-ray cocrystal structures during optimization efforts. This work culminated in the discovery of LNP023 (41), which is currently being evaluated clinically in several diverse AP mediated indications.

Several recent reports have shown that long short-term memory generative neural networks (LSTM) of the type used for grammar learning efficiently learn to write SMILES of drug-like compounds when trained with SMILES from a database of bioactive compounds such as ChEMBL and can later produce focused sets upon transfer learning with compounds of specific bioactivity profiles. Here we trained an LSTM using molecules taken either from ChEMBL, DrugBank, commercially available fragments, or from FDB-17 (a database of fragments up to 17 atoms) and performed transfer learning to a single known drug to obtain new analogs of this drug. We found that this approach readily generates hundreds of relevant and diverse new drug analogs and works best with training sets of around 40,000 compounds as simple as commercial fragments. These data suggest that fragment-based LSTM offer a promising method for new molecule generation.