Method and apparatus for identification and optimization of bioactive compounds using a neural network

A computerized double neural network system for predicting the chemical activity of at least one molecule of interest comprising:

• a) an outer neural network further comprising:
• i) an outer network input layer consisting of at least one neuron where input data is sent as a vector value;
• ii) an outer network weight matrix where every entry in the form of an input vector is multiplied by a set weight and then sent to at least one hidden layer neuron;
• iii) an outer network hidden layer consisting of at least one neuron such that when said input vector Is multiplied by a set weight said hidden layer contains said weight matrix, said weight matrix having the dimensions n by m where n is the length of an input vector and m is the number of hidden layer neurons available;
• iv) an outer network output layer consisting of at least one neuron where weight matrix data is sent before it is input into a transfer function;
• v) an outer network transfer function that is non-linear in form and is capable of taking any value generated by said output layer and returning a number between −1 and 1 or another predetermined range;
• vi) a training process for said outer neural network such that said neural network can accurately approximate a free energy of binding of at least one known training molecule with an output from said output layer;
• b) an inner neural network capable of receiving data from said outer neural network further comprising:
• i) an inner network weight matrix where every entry in the form of an input vector is multiplied by a set weight and then sent to at least one hidden layer neuron;
• ii) an inner network hidden layer consisting of at least one neuron such that when said input vector is multiplied by a set weight said hidden layer contains said weight matrix, said weight matrix having the dimensions n by m where n is the length of an input vector and m is the number of hidden layer neurons available;
• iii) an inner network output layer consisting of at least one neuron where weight matrix data is sent before it is input into a transfer function said inner network output layer having an output value;
• iv) an inner network transfer function that is non-linear in form and is capable of taking said output value generated by said output layer and returning a number between −1 and 1 or another predetermined range;
• v) a training process for said inner neural network such that said neural network can accurately approximate a free energy of binding of at least one known training molecule with an output from said output layer
• vi) a test process in which a trained neural network is used to predict a free energy of binding for said at least one molecule of interest;
• wherein said inner neural network is integrated to function with the data generated from said outer neural network such that the rules for said free energy of binding learned by said outer neural network are utilized by said inner neural network to model a quantum object such that said double neural network is used to predict the chemical characteristics of said quantum object, said quantum object describing a molecule with improved chemical properties of binding relative to said at least one molecule of interest;
• wherein said outer network output layer is the input layer of said inner neural network;
• wherein said outer network hidden layer includes an error term, said error term being used to calculate the correction terms for said outer network input layer such that the weights and biases of said double neural network are optimized;
• and
• wherein the physicochemical descriptor of said at least one molecule of interest is the quantum mechanical electrostatic potential of said at least one molecule of interest at the van der Waals surface of said at least one molecule of interest.

A computer implemented method for predicting the chemical activity of at least one molecule of interest by using a neural network comprising:

• a) inputting data into an input layer consisting of at least one neuron where input data is sent as a vector value;
• b) developing a weight matrix wherein every entry in the form of an input vector is multiplied by a set weight and then sent to at least one hidden layer neuron;
• c) providing a hidden layer consisting of at least one neuron such that when said input vector is multiplied by a set weight said hidden layer contains said weight matrix, said weight matrix having the dimensions n by m where n is the length of an input vector and m is the number of hidden layer neurons available;
• d) constructing an output layer consisting of at least one neuron where weight matrix data is sent before it is input into a transfer function;
• e) utilizing a transfer function that is non-linear in form and is capable of taking any value generated by said output layer and returning a number between −1 and 1;
• f) employing a training process for said neural network such that said neural network can accurately approximate a free energy of binding of at least one known training molecule with an output from said output layer;
• and
• employing a test process in which a trained neural network is used to predict a free energy of binding for said at least one molecule of interest wherein the physicochemical descriptor of said at least one molecule of interest is the quantum mechanical electrostatic potential of said at least one molecule of interest at the van der Waals surface of said at least one molecule of interest, wherein said test process includes the use of at least one adjuster molecule such that after said training process said neural network is used to predict a free energy of binding for said at least one adjuster molecule, said at least one adjuster molecule having a known free energy of binding and having been excluded from the set of molecules comprising the set of said of at least one known training molecule.

A computer implemented method for predicting the chemical activity of at least one molecule of interest by using a double neural network comprising:

• a) utilizing an outer neural network further comprising:
• i) an outer network input layer consisting of at least one neuron where input data is sent as a vector value;
• ii) an outer network weight matrix where every entry in the form of an input vector is multiplied by a set weight and then sent to at least one hidden layer neuron;
• iii) an outer network hidden layer consisting of at least one neuron such that when said input vector is multiplied by a set weight said hidden layer contains said weight matrix, said weight matrix having the dimensions n by m where n is the length of an input vector and m is the number of hidden layer neurons available;
• iv) an outer network output layer consisting of at least one neuron where weight matrix data is sent before it is input into a transfer function;
• v) an outer network transfer function that is non-linear in form and is capable of
• b) taking any value generated by said output layer and returning a number between −1 and 1;
• i) an outer network training process for said neural network such that said neural network can accurately approximate a free energy of binding of at least one known training molecule with an output from said output layer;
• c) providing an inner neural network capable of receiving data from said outer neural network further comprising:
• i) an inner network weight matrix where every entry in the form of an input vector is multiplied by a set weight and then sent to at least one hidden layer neuron;
• ii) an inner network hidden layer consisting of at least one neuron such that when said input vector is multiplied by a set weight said hidden layer contains said weight matrix, said weight matrix having the dimensions n by m where n is the length of an input vector and m is the number of hidden layer neurons available;
• iii) an inner network output layer consisting of at least one neuron where weight matrix data is sent before it is input into a transfer function said inner network output layer having an output value;
• iv) an inner network transfer function that is non-linear in form and is capable of taking any value generated by said output layer and returning a number between −1 and 1;
• v) an inner network training process for said neural network such that said neural network can accurately approximate a free energy of binding of at least one known training molecule with an output from said output layer;
• vi) a test process in which a trained neural network is used to predict a free energy of binding for said at least one molecule of interest:
• d) integrating said inner neural network to function with the data generated from said outer neural network such that the rules for said free energy of binding learned by said outer neural network are utilized by said inner neural network to model a quantum object such that said double neural network is used to predict the chemical characteristics of said quantum object, said quantum object describing a molecule with improved chemical properties of binding relative to said at least one molecule of interest;
• e) constructing said outer network input layer such that said output layer of said outer neural network is the input layer of said inner neural network;
• and
• f) providing said outer network hidden layer with an error term, said error term being used to calculate the correction terms for said outer network input layer such that the weights and biases of said double neural network are optimized;
• wherein the physicochemical descriptor of said at least one molecule of interest is the quantum mechanical electrostatic potential of said at least one molecule of interest at the van der Waals surface of said at least one molecule of interest.

A computer implemented method of determining a set of molecular descriptors:

• providing a neural network comprising an inner network trained to predict binding energy of a molecule of interest with a target molecule using a set of molecular descriptors based on geometric and/or electrostatic information for the molecule of interest the inner network having an input layer coupled to the output layer of an outer neural network for inputting molecular descriptors,
• in the inner neural network, setting the binding energy for an unknown molecule of interest to a desired level;
• and
• determining a set of molecular descriptors for an unknown molecule of interest by computing through the network a set of molecular descriptors that if output from the output layer of the outer neural network would yield a binding energy within a desired range of a predetermined binding energy set for the inner neural network.