Marc Bianciotto

We introduce a framework based on the Design–Make–Test–Analyze (DMTA) paradigm for simulating the outcome of prioritization strategies during lead optimization and assessing their performance in exploring and exploiting the chemical space.

Drug discovery requires traversing vast chemical spaces to identify compounds exhibiting favorable potency, selectivity and absorption–distribution–metabolism–excretion–toxicity (ADMET) profiles. During this process, the synthetic and assay throughput is generally markedly lower than the ideation of new propositions by the project team members, so that the prioritization of new syntheses is indispensable. We introduce a framework for simulating the outcome of multi-objective prioritization strategies during lead optimization. Based on the Design–Make–Test–Analyze (DMTA) paradigm, historical discovery programs are replayed round by round using user-defined compound selection strategies. We developed qualitative and quantitative tools to assess their performance in retrieving the best compounds and exploring the project’s chemical space. We demonstrate our pipeline using four industrial datasets, each containing chemical structures, assays values and time stamps. Multiple selection strategies are implemented, including approaches inspired by active learning (AL), multicriteria decision analysis (MCDA), and medicinal chemistry heuristics, that display distinct behavior in the selection of compounds. Retrospective analysis provides a rigorous, low-cost test bed for investigating selection strategies in lead optimization and could help reducing cost, duration and risk of lead optimization projects.

Over the past decade, collective intelligence, i.e., the intelligence that emerges from collective efforts, has transformed complex problem-solving and decision-making. In drug discovery, decision-making often relies on medicinal chemistry intuition. The present study explores the application of collective intelligence in drug discovery, focusing on lead optimization. Ninety-two Sanofi researchers with diverse expertise participated anonymously in an exercise centered on ADMET-related questions. Their feedback was used to build a collective intelligence agent, which was compared to an artificial intelligence model. The study led to three major conclusions: first, collective intelligence improves decision-making in optimizing ADMET endpoints, compared to individual decisions. Second, collective intelligence outperforms artificial intelligence for all other endpoints but hERG inhibition. Finally, we observe complementarity between collective human and artificial intelligence. Overall, this research highlights the potential of collective intelligence in drug discovery and the importance of a synergistic approach combining human and artificial intelligence in project decision making.

Using a mean-variance framework, we present a simple model to bridge the optimization objective of goal-directed generation with the need for diversity. We motivate theoretically that it minimizes risk measures when selecting multiple molecules.

Over the last decade, the combination of collective intelligence with computational methods has transformed complex problem-solving. Here, we investigate if and how collective intelligence can be applied to drug discovery, focusing on the lead optimization stage of the discovery process. For this study, 92 Sanofi researchers with diverse scientific expertise participated anonymously in a lead optimization exercise. Their feedback was used to build a collective intelligence agent that was compared to an artificial intelligence model developed in parallel. This work has led to three major conclusions. First, a significant improvement of collective versus individual decisions in optimizing ADMET endpoints is observed. Second, for all endpoints apart from hERG inhibition, the collective intelligence performance exceeds the artificial intelligence model. Third, we observe a complementarity between collective intelligence and AI for complex tasks, demonstrating the potential of hybrid predictions. Overall, this research highlights the potential of collective intelligence in drug discovery. The entire dataset, including questionnaire responses, and developed models are available for access on GitHub.

The heat shock protein 90 (Hsp90) is a molecular chaperone that controls the folding and activation of client proteins using the free energy of ATP hydrolysis. The Hsp90 active site is in its N-terminal domain (NTD). Our goal is to characterize the dynamics of NTD using an autoencoder-learned collective variable (CV) in conjunction with adaptive biasing force Langevin dynamics. Using dihedral analysis, we cluster all available experimental Hsp90 NTD structures into distinct native states. We then perform unbiased molecular dynamics (MD) simulations to construct a dataset that represents each state and use this dataset to train an autoencoder. Two autoencoder architectures are considered, with one and two hidden layers, respectively, and bottlenecks of dimension k ranging from 1 to 10. We demonstrate that the addition of an extra hidden layer does not significantly improve the performance, while it leads to complicated CVs that increase the computational cost of biased MD calculations. In addition, a two-dimensional (2D) bottleneck can provide enough information of the different states, while the optimal bottleneck dimension is five. For the 2D bottleneck, the 2D CV is directly used in biased MD simulations. For the five-dimensional (5D) bottleneck, we perform an analysis of the latent CV space and identify the pair of CV coordinates that best separates the states of Hsp90. Interestingly, selecting a 2D CV out of the 5D CV space leads to better results than directly learning a 2D CV and allows observation of transitions between native states when running free energy biased dynamics.

Molecular generative artificial intelligence is drawing significant attention in the drug design community, with several experimentally validated proof of concepts already published. Nevertheless, generative models are known for sometimes generating unrealistic, unstable, unsynthesizable, or uninteresting structures. This calls for methods to constrain those algorithms to generate structures in drug-like portions of the chemical space. While the concept of applicability domains for predictive models is well studied, its counterpart for generative models is not yet well-defined. In this work, we empirically examine various possibilities and propose applicability domains suited for generative models. Using both public and internal data sets, we use generative methods to generate novel structures that are predicted to be actives by a corresponding quantitative structure–activity relationships model while constraining the generative model to stay within a given applicability domain. Our work looks at several applicability domain definitions, combining various criteria, such as structural similarity to the training set, similarity of physicochemical properties, unwanted substructures, and quantitative estimate of drug-likeness. We assess the structures generated from both qualitative and quantitative points of view and find that the applicability domain definitions have a strong influence on the drug-likeness of generated molecules. An extensive analysis of our results allows us to identify applicability domain definitions that are best suited for generating drug-like molecules with generative models. We anticipate that this work will help foster the adoption of generative models in an industrial context.

Molecular generative artificial intelligence is drawing significant attention in the drug design community, with several experimentally validated proofs of concepts already published. Nevertheless, generative models are known for sometimes generating unrealistic, unsynthesizable or unstable structures. This calls for methods to constrain those algorithms to generate structures in reasonable portions of the chemical space. While the concept of applicability domains (AD) for predictive models is well studied, its counterpart for generative models is not yet defined. In this work, we examine empirically various possibilities and propose applicability domains suited for generative models. Using both public and internal datasets, we use state-of-the-art generative methods to generate novel structures that are predicted actives by a corresponding QSAR model, while constraining the generative model to stay within a given applicability domain. Our work looks at several applicability domain definitions, combining various criteria, such as structural similarity to the training set, similarity of physico-chemical properties, unwanted substructures, and Quantitative Estimate of Drug- Likeness (QED). We assess both from a qualitative and quantitative point of view the structures generated, and find that the applicability domain definitions have a strong influence on the chemical beauty of generated molecules. An extensive analysis of our results allows us to identify applicability domain definitions that are best suited for generating drug-like molecules with generative models. We anticipate that this work will help foster the adoption of generative models in an industrial context.

Despite growing interest and success in automated in-silico molecular design, questions remain regarding the ability of goal-directed generation algorithms to perform unbiased exploration of novel chemical spaces. A specific phenomenon has recently been highlighted: goal-directed generation guided with machine learning models produce molecules with high scores according to the optimization model, but low scores according to control models, even when trained on the same data distribution and the same target. In this work, we show that this worrisome behavior is actually due to issues with the predictive models and not the goal-directed generation algorithms. We show that with appropriate predictive models, this issue can be resolved, and molecules generated have high scores according to both the optimization and the control models.