3D-Scaffold: Deep Learning Framework to Generate 3D Coordinates of Drug-like Molecules with Desired Scaffolds

Abstract

The prerequisite of therapeutic drug design is to identify novel molecules with desired biophysical and biochemical properties. Deep generative models have demonstrated their ability to find such molecules by exploring a huge chemical space efficiently. An effective way to obtain molecules with desired target properties is the preservation of critical scaffolds in the generation process. To this end, we propose a domain aware generative framework called 3D-Scaffold that takes 3D coordinates of the desired scaffold as an input and generates 3D coordinates of novel therapeutic candidates as an output while always preserving the desired scaffolds in generated structures. We show that our framework generates predominantly valid, unique, novel, and experimentally synthesizable molecules that have drug-like properties similar to the molecules in the training set. Using domain specific datasets, we generate covalent and non-covalent antiviral inhibitors. To measure the success of our framework in generating therapeutic candidates, generated structures were subjected to high throughput virtual screening via docking simulations, which shows favorable interaction against SARS-CoV-2 main protease and non-structural protein endoribonuclease (NSP15) targets. Most importantly, our model performs well with relatively small volumes of training data and generalizes to new scaffolds, making it applicable to other domains.

Inverse design of 3d molecular structures with conditional generative neural networks

“…Further work is required to apply the cG-SchNet architecture to the exploration of significantly larger systems and a more diverse set of atom types. Although an unconditional G-SchNet has been trained on druglike molecules with 50+ atoms in the 3D-Scaffold framework [54], adjustments will be necessary to ensure scalability to materials. In the current implementation, we employ all preceding atoms to predict the type and reconstruct the positional distribution of the next atom.…”

“…This includes specifically designed approaches to translate given molecular graphs to 3d conformations [32][33][34][35][36][37][38], map from coarse-grained to fine-grained structures [39], sample unbiased equilibrium configurations of a given system [40,41], or focus on protein folding [42][43][44][45][46]. In contrast, other models aim at sampling directly from distributions of 3d molecules with arbitrary composition [47][48][49][50][51][52][53][54][55][56], making them suitable for general inverse design settings. These models need to be biased towards structures with properties of interest, e.g.…”

“…These models need to be biased towards structures with properties of interest, e.g. using reinforcement learning [51,52,56], fine-tuning on a biased data set [48], or other heuristics [54]. Some of us have previously proposed G-SchNet [48], an auto-regressive deep neural network that generates diverse, small organic molecules by placing atom after atom in Euclidean space.…”

“…Some of us have previously proposed G-SchNet [48], an auto-regressive deep neural network that generates diverse, small organic molecules by placing atom after atom in Euclidean space. It has been applied in the 3D-Scaffold framework to build molecules around a functional group associated with properties of interest in order to discover novel drug candidates [54]. Such an approach requires prior knowledge about the relationship between functional groups and target properties and might prevent the model from unfolding its potential by limiting sampling to very specific molecules.…”

Artificial Intelligence based Autonomous Molecular Design for Medical Therapeutic: A Perspective

“…In their models, information is propagated back and forth in the molecules in the form of waves making it possible to pass the information locally while simultaneously travelling the entire molecule in a single pass. With the unprecedented success of learned molecular representations for predictive modelling, they are also adopted with success for generative models [54,66]…”