Application US20210081804

TENSOR NETWORK MACHINE LEARNING SYSTEM
[+] Show Details

US20210081804

1. A machine learning based method of identifying candidate, small, drug-like molecules, comprising the step of providing molecular orbital representations of drug-like molecules and/or parts of proteins relevant to an interaction with the molecules, as an input to a machine learning system, to predict molecular properties and identify candidate drug-like molecules. (28) (0)

2. The method of claim 1, in which molecular orbital representations of drug-like molecules and/or parts of proteins relevant to an interaction with the molecules are represented as tensor networks. (9) (0)

8. The method of claim 2 in which a tensor network is any mathematical object in an exponentially large Hilbert Space. (0)

9. The method of claim 2 in which the tensor networks are one or more of the following: matrix product state (MPS), tensor train, multi-scale entanglement renormalization ansatz (MERA), projected entangled-pair states (PEPS), correlator product states, tree tensor networks, complete graph network states, tensor network states. (0)

11. The method of claim 2 in which the tensor network representations include networks describing states with volume law entanglement that allow for tensor networks describing density matrices, including both those that obey the area law, and those that do not, and arbitrary superpositions of tensor networks, containing elements in general from distinct types of architectures (e.g. a superposition of two MPS tensor networks with a MERA network). (0)

14. The method of claim 2 in which tensorial networks are used to decompose high rank structures appearing in graph models, hence enabling quantum featurisation and modelling in the context of graph models. (0)

15. The method of claim 2 in which a feature map is applied to input data to transform that input into a high dimensional tensor network structure. (0)

16. The method of claim 2 in which a feature map is applied to input data to transform that input into a high dimensional tensor network structure. (0)

19. The method of claim 2 in which the tensor network represents a set of simple descriptions of a molecule whole dataset of molecules. (0)

20. The method of claim 2 in which the tensor network representations include a training data set of small drug-like molecules and a target, in which the small drug-like molecules are known to bind to the target. (0)

21. The method of claim 2 in which the training dataset is a tensor network representation of the molecular quantum states of small drug-like molecules and also a tensor network representation of the molecular quantum states of one or more targets, such as a protein to which a drug-like molecule may potentially bind. (0)

3. The method of claim 1 in which a training dataset is used to train the machine learning system, and the training dataset is a molecular orbital representation of the drug-like molecules and the machine learning system uses generative models to capture statistically meaningful distributions or patterns in the training dataset sand the trained machine learning system is configured to predict novel candidate drug-like molecules. (0)

4. The method of claim 1 in which the machine learning system is configured to predict molecular properties and to optimize the selection of candidates across multiple parameters, such as binding affinity and favourable toxicity. (2) (0)

5. The method of claim 1 in which the machine learning system is configured as a tensor network or includes tensor decompositions in its components and further configured to predict molecular properties and to optimize the selection of candidates across multiple parameters, such as binding affinity and favorable toxicity. (0)

6. The method of claim 1 used for analyzing a chemical compound dataset so as to determine suitable chemical compounds, the method comprising the steps of: (18) (3)

(i) determining a tensorial space for said chemical compounds dataset, or other chemical compounds;

(ii) correlating the tensorial space with a large search or latent space, that search or latent space itself being tensorial; and

(iii) outputting a further dataset of candidate chemical compounds using a generative model.

53. The method of claim 6, in which the search or latent space represents physically relevant wavefunctions. (0)

54. The method of claim 6, in which the search or latent space represents physically relevant wave functions in conjunction with a decoder network. (0)

55. The method of claim 6, in which the further dataset of candidate chemical compounds is a filtered dataset of candidate chemical compounds. (0)

56. The method of claim 6, in which the further dataset of candidate chemical compounds is a novel dataset of candidate chemical compounds. (0)

57. The method of claim 6 in which wavefunctions describing elements in some datasets are each represented by an appropriate tensor network. (0)

58. The method of claim 6 in which a whole set of wavefunctions are described by an appropriate tensor network. (0)

59. The method of claim 6 in which the tensor network is one of: complete-graph tensor network state, correlator product state, projected entanglement pair states (PEPS), matrix product states (MPS)/tensor train, tree tensor network, and matrix product operator. (0)

60. The method of claim 6, in which the dataset is mapped into a higher dimensional tensor. (0)

61. The method of claim 6, in which the correlating process includes optimising over the search or latent space. (0)

62. The method of claim 6, in which the correlating process includes determining an optimal tensor network to describe the mapping of the input data set to a more general description. (0)

63. The method of claim 6, in which the search or latent space represents desired properties of a set of candidate chemical compounds. (0)

64. The method of claim 6, in which the correlating process includes maximising a cost function, said cost function representing desirable properties of the candidate chemical compounds. (0)

65. The method of claim 6, in which determining the tensorial space for the chemical compounds dataset comprises encoding said dataset using a generative model. (1) (0)

67. The method of claim 65, in which the generative model comprises a neural network. (4) (0)

68. The method of claim 67, in which the generative model comprises a graph variational auto encoder or graph based GAN, with enhanced featurisation based on separate quantum mechanical calculations (for example done using tensor network methods, or other quantum chemistry methods). (0)

69. The method of claim 67 in which the neural network is in the form of an autoencoder. (1) (0)

71. The method of claim 67, in which the neural network is any of the following: generative auto encoder, RNN, GAN, Monte-Carlo tree search model, an Ising model or a restricted Boltzmann machine trained in an unsupervised manner. (0)

72. The method of claim 67, in which the neural network comprises tensorial decompositions or tensor networks, or is constructed as a tensor network in its entirety. (1) (0)

66. The method of claim 6, in which the outputting process includes the process of decoding using a generative model. (0)

74. The method of claim 6 in which a visualisation of the tensorial space is generated and output. (0)

75. The method of claim 6, in which candidate chemical compounds are active elements of a pharmaceutical compound. (0)

76. The method of claim 6, in which candidate chemical compounds are suitable for use in an organic light emitting diode. (0)

77. The method of claim 6, in which the chemical compound dataset is a dataset of possible molecules containing up to a predetermined maximum number of atoms. (0)

7. The method of claim 1 in which the molecular orbital representations of drug-like molecules and/or parts of proteins relevant to an interaction with the molecules contain direct information about quantum correlations and entanglement properties. (0)

10. The method of claim 1 in which molecular orbital representations of drug-like molecules and/or parts of proteins relevant to an interaction with the molecules include representations describing states with volume law entanglement, which can, for example, provide descriptions of highly excited states present in transition states of small molecules in a reaction or a binding process. (0)

12. The method of claim 1 in which the molecular orbital representations of drug-like molecules and/or parts of protein relevant to an interaction with the molecules efficiently represent a set of molecular quantum states of small drug-like molecules. (1) (0)

17. The method of claim 1 in which featurization of graph models is by introducing entanglement features or other quantum mechanically features, derived from approximate molecular wave functions. (0)

18. The method of claim 1 in which parts of protein relevant to an interaction with the molecules are provided as an input to the machine learning system. (0)

22. The method of claim 1 in which a training dataset is used to train a generative model. (0)

23. The method of claim 1 in which a training dataset is used to train a predictive model. (0)

24. The method of claim 1 in which a training dataset is used to train a generative model, which in turn feeds the predictive model. (0)

25. The method of claim 1 in which a predictive model is configured to predict whether a candidate small drug-like molecule is appropriate to a specific requirement. (5) (0)

31. The method of claim 1 in which a training dataset is fed high-quality molecules by a predictive model. (0)

34. The method of claim 1 in which the machine learning system is configured to handle tensor network inputs by applying tensorial layers to the input, and using a mechanism to decrease the complexity of the resulting tensor network, so that it can be feasibly run on classical computers. (0)

35. The method of claim 1 in which the machine learning system has been trained on a training dataset. (0)

36. The method of claim 1 in which the machine learning system is supervised, semi-supervised or unsupervised. (0)

37. The method of claim 1 in which the machine learning system is a generative network, or an autoencoder, or RNN, or Monte-Carlo tree search model, or an Ising model or a restricted Boltzmann machine trained in an unsupervised manner, a graph model, a graph to graph autoencoder. (0)

38. The method of claim 1 in which the machine learning system is a generative adversarial network. (0)

40. The method of claim 1 in which the machine learning system is a neural network that comprises tensorial decompositions or tensor networks. (0)

41. The method of claim 1 in which the machine learning system is a quantum computer with quantum circuits configured as tensor networks. (0)

42. The method of claim 1 in which the machine learning system is a quantum machine learning circuit. (1) (0)

44. The method of claim 1 in which the machine learning system outputs molecular orbital representations of drug-like molecules to a predictive model. (3) (0)

45. The method of claim 1 in which a predictive model screens the output from the machine learning system. (0)

52. The method of claim 1 in which the machine learning system is one of the following: generative model, generative adversarial network, autoencoder or recursive neural network. (0)

78. A system configured to perform method of claim 1. (0)

79. A molecule or class of molecules identified using the method of claim 1. (0)

32. (canceled) (0)

33. (canceled) (0)

39. (canceled) (0)

47. (canceled) (0)

View Abstract and Specification Size

PDF with Images and Document Face >

Full Text Publication >

Application US20210081804

TENSOR NETWORK MACHINE LEARNING SYSTEM

Patent Matrix® Search