Target-to-catalyst translation networks

1 . A computer system for generating the molecular structure of a catalytic activator for a reaction in which input reactants are converted into an output product, the computer system comprising:

one or more processors; and

a non-transitory computer readable medium storing a plurality of instructions, which when executed, cause the one or more processors to:

receive, by a trained machine learning model, an operating feature set defining chemical features of the input reactants and chemical features of the output product of a reaction, the trained machine learning model being a Variational AutoEncoder comprising: i) a chemical encoder configured to encode the chemical features into a latent vector that defines a point in a latent manifold space, and ii) a catalyst decoder configured to decode the defined point in the latent space into a latent vector defining the set of catalyst features for catalysing the reaction that converts the input reactants to the output product;

generate, by the trained machine learning model, a set of catalyst features defining one or more catalytic activators for catalysing a reaction to convert the input reactants to the output product; and

cause the reaction to convert the input reactants to the output product using the generated set of catalyst features defining one or more catalytic activators.

2 . The computer system according to claim 1 , wherein the machine learning model is a Variational AutoEncoder comprising a catalyst encoder operable to encode the catalyst features into a latent vector which defines a point in a latent space and a chemical decoder operable to decode the defined point in the latent space into a latent vector which defines the chemical features of the input reactants and output product of a reaction.

3 . The computer system according to claim 1 , wherein the encoders and decoders each comprise a neural network having an encoder comprising an input layer of nodes, at least one hidden layer of nodes and an output layer of nodes, the function of each node being parameterised by weights.

4 . The computer system according to claim 1 , wherein the trained machine learning model comprises kernel memory in which the weights are stored.

5 . A method of training a Variational AutoEncoder machine learning model, the method comprising:

providing a first set of training data defining chemical features of input reactants and output products for each of a plurality of reactions;

encoding the chemical features into latent vectors defining points in a latent space and defining the distribution of the latent vectors in the latent space, wherein defining the distribution of the latent vectors in the latent space comprises:

decoding the chemical latent vectors using a catalyst decoder to generate catalyst features in a catalyst domain,

comparing the decoded catalyst features with catalyst features of the second set of training data,

decoding the catalyst latent vectors using a chemical decoder into chemical features in a chemical domain,

comparing the decoded chemical features with the chemical feature of the first set of training data, and

minimising, jointly, a loss function in the comparison steps to match the distributions in the common latent space of the latent vectors in the chemical domain and the latent vector in the catalyst domain;

providing a second set of training data defining catalyst features of catalytic activators for each of the plurality of reactions; and

encoding the catalyst features into catalyst latent vectors in the latent space and matching the distribution of the catalyst latent vectors to the chemical latent vectors such that decoding a chemical latent vector at a point in the latent space generates catalyst features of a catalytic activator for catalysing a reaction to convert the input reactants of that reaction to an output product, the input reactant and the output product defined by the chemical features encoded in the latent vector.

6 . The method according to claim 5 , comprising the step of decoding the chemical latent vectors using a chemical decoder and comparing the decoded chemical features with the chemical features of the first set of training data and minimising the loss of an objective training function by adjusting weights in a first domain of the variational autoencoder; and

decoding the catalyst latent vector into decoded catalyst features and comparing the decoded catalyst features with catalyst features of the second set of training data to minimise the loss of an objective training function by adjusting weights of a second domain of the variational autoencoder, and sharing the weights of the first and second domains.

7 . The computer system according to claim 6 , wherein the encoders and decoders each comprise a neural network having an encoder comprising an input layer of nodes, at least one hidden layer of nodes and an output layer of nodes, the function of each node being parameterised by weights.

8 . The computer system according to claim 5 , wherein the catalytic activator is a catalyst.

9 . The computer system according to claim 8 , wherein the catalyst is an enzyme.

10 . A computer program product, comprising a non-transitory computer-readable medium having computer-readable program code embodied therein to be executed by one or more processors, the program code including instructions to:

provide a first set of training data defining chemical features of input reactants and output products for each of a plurality of reactions;

encode the chemical features into latent vectors defining points in a latent space and defining the distribution of the latent vectors in the latent space, wherein defining the distribution of the latent vectors in the latent space comprises:

decode the chemical latent vectors using a catalyst decoder to generate catalyst features in a catalyst domain,

compare the decoded catalyst features with catalyst features of the second set of training data,

decode the catalyst latent vectors using a chemical decoder into chemical features in a chemical domain,

compare the decoded chemical features with the chemical feature of the first set of training data, and

minimize, jointly, a loss function in the comparison steps to match the distributions in the common latent space of the latent vectors in the chemical domain and the latent vector in the catalyst domain;

provide a second set of training data defining catalyst features of catalytic activators for each of the plurality of reactions; and

encode the catalyst features into catalyst latent vectors in the latent space and matching the distribution of the catalyst latent vectors to the chemical latent vectors such that decoding a chemical latent vector at a point in the latent space generates catalyst features of a catalytic activator for catalysing a reaction to convert the input reactants of that reaction to an output product, the input reactant and the output product defined by the chemical features encoded in the latent vector.

11 . The computer program product according to claim 10 , wherein the program code further includes instructions to:

decode the chemical latent vectors using a chemical decoder and compare the decoded chemical features with the chemical features of the first set of training data and minimize the loss of an objective training function by adjusting weights in a first domain of the variational autoencoder; and

decode the catalyst latent vector into decoded catalyst features and compare the decoded catalyst features with catalyst features of the second set of training data to minimize the loss of an objective training function by adjusting weights of a second domain of the variational autoencoder, and sharing the weights of the first and second domains.

12 . The computer program product according to claim 10 , wherein the program code further includes instructions to:

decode the chemical latent vectors using a chemical decoder and compare the decoded chemical features with the chemical features of the first set of training data and minimize the loss of an objective training function by adjusting weights in a first domain of the variational autoencoder; and

decode the catalyst latent vector into decoded catalyst features and compare the decoded catalyst features with catalyst features of the second set of training data to minimize the loss of an objective training function by adjusting weights of a second domain of the variational autoencoder, and share the weights of the first and second domains.

13 . A computer program product for generating the molecular structure of a catalytic activator for a reaction in which input reactants are converted into an output product, computer program product comprising a non-transitory computer-readable medium having computer-readable program code embodied therein to be executed by one or more processors, the program code including instructions to:

receive, by a trained machine learning model, an operating feature set defining chemical features of the input reactants and chemical features of the output product of a reaction, the trained machine learning model being a Variational AutoEncoder comprising: i) a chemical encoder configured to encode the chemical features into a latent vector that defines a point in a latent manifold space, and ii) a catalyst decoder configured to decode the defined point in the latent space into a latent vector defining the set of catalyst features for catalysing the reaction that converts the input reactants to the output product;

generate, by the trained machine learning model, a set of catalyst features defining one or more catalytic activators for catalysing a reaction to convert the input reactants to the output product; and

cause the reaction to convert the input reactants to the output product using the generated set of catalyst features defining one or more catalytic activators.

14 . The computer program product according to claim 13 , wherein the machine learning model is a Variational AutoEncoder comprising a catalyst encoder operable to encode the catalyst features into a latent vector which defines a point in a latent space and a chemical decoder operable to decode the defined point in the latent space into a latent vector which defines the chemical features of the input reactants and output product of a reaction.

15 . The computer program product according to claim 13 , wherein the encoders and decoders each comprise a neural network having an encoder comprising an input layer of nodes, at least one hidden layer of nodes and an output layer of nodes, the function of each node being parameterised by weights.

16 . The computer program product according to claim 13 , wherein the trained machine learning model comprises kernel memory in which the weights are stored.