Methods and systems for machine-learning based molecule generation and scoring

A system comprising:

• one or more computers;
• and
• one or more storage devices communicatively coupled to the one or more computers, wherein the one or more storage devices store instructions that, when executed by the one or more computers, cause the one or more computers to perform operations comprising:
• identifying one or more ligands for binding to a target molecule by performing operations comprising:
• obtaining data identifying an input ligand;
• generating, using a generative machine learning model, a noised latent vector by combining random noise with data representing the input ligand,
• wherein the generative machine learning model has been trained, by a machine learning training technique on a training dataset of ligand structures, to generate a predicted structure of a ligand by denoising an initial latent vector over a sequence of denoising iterations;
• iteratively denoising the noised latent vector using the generative machine learning model over the sequence of denoising iterations to generate a final denoised latent vector that defines an output ligand;
• processing a model input characterizing a 3D structure of the output ligand using a scoring model to generate a quality score for the output ligand;
• determining gradients of the quality score for the output ligand by backpropagating the gradients of the quality score through the scoring model and the generative machine learning model and into an internal latent vector of the generative machine learning model corresponding to a target denoising iteration, wherein the target denoising iteration is any denoising iteration before a final denoising iteration in the sequence of denoising iterations;
• updating the internal latent vector at the target denoising iteration using the gradients of the quality score;
• and
• iteratively denoising the updated internal latent vector, using the generative machine learning model and starting from the target denoising iteration in the sequence of denoising iterations, to generate an optimized denoised latent vector that defines an optimized ligand;
• and
• outputting data defining the optimized ligand.

One or more non-transitory computer storage media storing instructions that when executed by one or more computers cause the one or more computers to perform operations comprising:

• identifying one or more ligands for binding to a target molecule by performing operations comprising:
• obtaining data identifying an input ligand;
• generating, using a generative machine learning model, a noised latent vector by combining random noise with data representing the input ligand,
• wherein the generative machine learning model has been trained, by a machine learning training technique on a training dataset of ligand structures, to generate a predicted structure of a ligand by denoising an initial latent vector over a sequence of denoising iterations;
• iteratively denoising the noised latent vector using the generative machine learning model over the sequence of denoising iterations to generate a final denoised latent vector that defines an output ligand;
• processing a model input characterizing a 3D structure of the output ligand using a scoring model to generate a quality score for the output ligand;
• determining gradients of the quality score for the output ligand by backpropagating the gradients of the quality score through the scoring model and the generative machine learning model and into an internal latent vector of the generative machine learning model corresponding to a target denoising iteration, wherein the target denoising iteration is any denoising iteration before a final denoising iteration in the sequence of denoising iterations;
• updating the internal latent vector at the target denoising iteration using the gradients of the quality score;
• and
• iteratively denoising the updated internal latent vector, using the generative machine learning model and starting from the target denoising iteration in the sequence of denoising iterations, to generate an optimized denoised latent vector that defines an optimized ligand;
• and
• outputting data defining the optimized ligand.