Integrating synthetic accessibility with AI-based generative drug design

Abstract

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 feasibility 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 feasibility scores to a binary ground truth of synthetic feasibility as estimated by chemists on a bench of generated molecules, as a first experimental validation that the RScore is a reliable synthetic feasibility 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.

Predicting the price of molecules using their predicted synthetic pathways

“…The loss of the model is defined in Equation (1) as a mean squared error loss between the predicted price (p θ ) of the root molecule and its true price (y) weighted by the score (s) representing the ranking score (Rscore [4]) of the route. The Rscore was used as a weighting factor in the training loss in order to prioritize the information coming from well scored routes against the one coming from badly scored routes during the training.…”

“…During the research and development (R&D) process of a new drug, chemists use different metrics to refine and filter libraries of virtual molecules in order to prioritize their synthesis, such as QSAR models' predictions of biological or physchem properties [1], docking scores [2], druggability metrics [3], and synthetic feasibility scores [4] [5], to name only a few. On top of those well known metrics, a function which estimates the price of a novel virtual molecule taking into account its synthetic pathway and the associated starting materials (price, provider, and availability) has never been considered before.…”

“…RetroPriceNet takes the predicted routes found by Spaya as input, with the possibility of using search options such as a selection of providers, a maximal number of steps, a maximal price of starting materials, a set of name reactions to pass by or to avoid, the presence of key intermediates, and so on. For each molecule, out of the top routes found by Spaya, the best route is selected to feed RetroPriceNet, based on the Rscore [4], Iktos proprietary synthetic feasibility score, which is used to compare synthetic routes based on their probability of success. RetroPriceNet was trained on a large database of 6.2 million commercial molecules along with their prices and associated predicted synthetic routes.…”

Unlocking the Future of Drug Development: Generative AI, Digital Twins, and Beyond

“…In recent years, a subset of AI -generative AI has undergone significant advancements, transforming various domains by generating realistic content. Generative AI refers to a category of models designed to generate new content similar to, but not the same as, the input data it was trained on [7][8][9]. Unlike traditional AI systems that are often task-specific and deterministic, generative AI systems can produce novel outputs by learning the underlying patterns and structures of the training data.…”