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.

Integrating synthetic accessibility with AI-based generative drug design.

Generative models are frequently used for de novo design in drug discovery projects to propose new molecules. However, the question of whether or not the generated molecules can be synthesized is not systematically taken into account during generation, even though being able to synthesize the generated molecules is a fundamental requirement for such methods to be useful in practice. Methods have been developed to estimate molecule "synthesizability", but, so far, there is no consensus on whether or not a molecule is synthesizable. In this paper we introduce the Retro-Score (RScore), which computes a synthetic accessibility score of molecules by performing a full retrosynthetic analysis through our data-driven synthetic planning software Spaya, and its dedicated API: Spaya-API (https://spaya.ai). We start by comparing several synthetic accessibility scores to a binary "chemist score" as estimated by chemists on a bench of generated molecules, as a first experimental validation that the RScore is a reliable synthetic accessibility score. We then describe a pipeline to generate molecules that validate a list of targets while still being easy to synthesize. We further this idea by performing experiments comparing molecular generator outputs across a range of constraints and conditions. We show that the RScore can be learned by a Neural Network, which leads to a new score: RSPred. We demonstrate that using the RScore or RSPred as a constraint during molecular generation enables our molecular generators to produce more synthesizable solutions, with higher diversity. The open-source Python code containing all the scores and the experiments can be found on ( https://github.com/iktos/generation-under-synthetic-constraint ).

Deep generative models for ligand-based de novo design applied to multi-parametric optimization.

Multi-parameter optimization (MPO) is a major challenge in new chemical entity (NCE) drug discovery. Recently, promising results were reported for deep learning generative models applied to de novo molecular design, but, to our knowledge, until now no report was made of the value of this new technology for addressing MPO in an actual drug discovery project. In this study, we demonstrate the benefit of applying AI technology in a real drug discovery project. We evaluate the potential of a ligand-based de novo design technology using deep learning generative models to accelerate the obtention of lead compounds meeting 11 different biological activity objectives simultaneously. Using the initial dataset of the project, we built QSAR models for all the 11 objectives, with moderate to high performance (precision between 0.67 and 1.0 on an independent test set). Our DL-based AI de novo design algorithm, combined with the QSAR models, generated 150 virtual compounds predicted as active on all objectives. Eleven were synthetized and tested. The AI-designed compounds met 9.5 objectives on average (i.e., 86% success rate) versus 6.4 (i.e., 58% success rate) for the initial molecules measured on all objectives. One of the AI-designed molecules was active on all 11 measured objectives, and two were active on 10 objectives while being in the error margin of the assay for the last one. The AI algorithm designed compounds with functional groups, which, although being rare or absent in the initial dataset, turned out to be highly beneficial for the MPO.