Chemical sensing system

• US 10,386,379 B1
• Filed: 01/29/2019
• Issued: 08/20/2019
• Est. Priority Date: 01/29/2018
• Status: Active Grant

• Current Assignee: Stratuscent Inc.
• Original Assignee: Stratuscent Inc.
• Inventors: Khomami Abadi, Mojtaba, Gahroosi, Amir Bahador, Gould, Matthew V., Masilamani, Ashok Prabhu
• Primary Examiner(s): Ishizuka, Yoshihisa
• Application Number:
  US16/261,195
• Publication Number:
  US 20190234973A1
• Time in Patent Office: 203 Days
• Field of Search: 702 23
• CPC Class Codes: B82Y 15/00 Nanotechnology for interacting, sensing or actuating, e.g. quantum dots as markers in protein assays or molecular motors
• CPC Class Codes: G01N 2035/00891 Displaying information to the operator
• CPC Class Codes: G01N 33/0031 comprising two or more sensors, e.g. a sensor array
• CPC Class Codes: G01N 33/0034 comprising neural networks or related mathematical techniques
• CPC Class Codes: G01N 35/00722 Communications; Identification
• CPC Class Codes: G06N 3/044 Recurrent networks, e.g. Hopfield networks
• CPC Class Codes: G06N 3/045 Combinations of networks
• CPC Class Codes: G06N 3/088 Non-supervised learning, e.g. competitive learning

Abstract

A chemical sensing system is described. The chemical sensing system can include a plurality of sensors arranged on at least one substrate. The sensors may have differing sensitivities to sense different analytes, and may each be configured to output a signal in response to sensing one or more of the different analytes. The chemical sensing system can further include a computer processor programmed to receive the signals output from the plurality of sensors. The computer processor may be further programmed to determine a concentration of analytes in the sample based, at least in part, on the received signals and a model relating the signals or information derived from the signals to an output representation having bases corresponding to analytes.

Claims

1. A chemical sensing system, comprising:

a plurality of sensors arranged on at least one substrate, wherein a first sensor and a second sensor of the plurality of sensors have different sensitivities to sense at least one analyte in a sample, each of the plurality of sensors being configured to output a signal in response to sensing the at least one analyte; and

a computer processor programmed to;

receive the signals output from the plurality of sensors;

determine a concentration of the at least one analyte in the sample by;

providing the received signals as input to a first model trained to relate the signals output from the plurality of sensors to a feature representation to generate feature representation values;

providing the feature representation values as input to a second model trained to relate the feature representation to a latent representation to generate latent representation values, the latent representation having a lower dimensionality than the feature representation;

providing the latent representation values as input to a third model trained to relate the latent representation to a straightened orthogonal representation to generate straightened orthogonal representation values, wherein;

one-dimensional manifolds in the straightened orthogonal representation corresponding to varying concentrations of a same analyte have zero angle between them; and

one-dimensional manifolds in the straightened orthogonal representation corresponding to different analytes are orthogonal;

providing the straightened orthogonal representation values as input to a fourth model trained to relate the straightened orthogonal representation to an output representation having bases corresponding to analytes, the analytes including the at least one analyte; and

determining the concentration of the at least one analyte in the sample based on an output of the fourth model; and

provide an indication of the determined concentration of the at least one analyte in the sample to a user.

2. The chemical sensing system of claim 1, wherein the first model comprises:

a mapping from the signals to response values; and

a weighting relating the response values to the feature representation.

3. The chemical sensing system of claim 2, wherein the weighting comprises weights of a recurrent neural network or convolutional neural network that relate the response values to the feature representation.

4. The chemical sensing system of claim 2, wherein the mapping assigns to each response value an amplitude characteristic and/or a temporal characteristic of a corresponding signal.

5. The chemical sensing system of claim 4, wherein the amplitude characteristic comprises a normalized change in an amplitude of the corresponding signal.

6. The chemical sensing system of claim 4, wherein the temporal characteristic comprises a rise time or a fall time of an amplitude of the corresponding signal.

7. The chemical sensing system of claim 1, wherein a change in a location of a point in the feature representation corresponding to an analyte concentration in the sample is a non-linear function of a change in the analyte concentration in the sample.

8. The chemical sensing system of claim 1, wherein each signal comprises a time series of measurements.

9. The chemical sensing system of claim 1, wherein

the first model and the second model are device-specific models;

and

the third model and the second model are device-independent models.

10. A chemical sensing system, comprising:

at least one computer processor; and

at least one computer readable medium including instructions that, when executed by the at least one computer processor, cause the chemical sensing system to perform a training process, the training process comprising;

receiving a training dataset including values associated with signals output from a plurality of sensors having different sensitivities to sense at least one analyte in a sample, each of the plurality of sensors configured to output the signal in response to sensing the at least one analyte;

training a set of models using the training dataset by;

training a first model to relate the signals output from the plurality of sensors to a feature representation;

training a second model to relate the feature representation to a latent representation, the latent representation having a lower dimensionality than the feature representation;

training a third model to relate the latent representation to a straightened orthogonal representation, wherein;

one-dimensional manifolds in the straightened orthogonal representation corresponding to varying concentrations of a same analyte have zero angle between them; and

one-dimensional manifolds in the straightened orthogonal representation corresponding to different analytes are orthogonal; and

training a fourth model to relate the straightened orthogonal representation to the output representation having bases corresponding to analytes; and

configuring the chemical sensing system to detect a concentration of the least one analyte in the sample using the trained set of models.

11. The system of claim 10, wherein training the first model comprises updating weights in a recurrent neural network or a convolutional neural network.

12. The system of claim 10, wherein the training dataset does not include labels for at least some of the values in the training set and wherein training the first model comprises training the first model without using the labels.

13. The system of claim 10, wherein training the second model comprises training at least a portion of an autoencoder or generative adversarial network.

14. The system of claim 10, wherein:

a continuous manifold embedded in the feature representation contains the signals output from the plurality of sensors, the continuous manifold having a dimensionality equal to the dimensionality of the latent representation; and

the second model relates the continuous manifold to the latent representation.

15. The system of claim 10, wherein training the set of models using the training dataset further comprises automatically identifying manifolds corresponding to samples of varying concentrations of a same analyte using local gradient changes in the feature representation.

16. The system of claim 10, wherein training the fourth model comprises associating bases in the straightened orthogonal representation with the analytes.

17. The system of claim 10, wherein training the fourth model comprises:

associating bases in the straightened orthogonal representation with the analytes using labels in the training data;

generating an introduction model that introduces the associated bases onto standard bases of a non-linearized output representation; and

generating a linearizing model that linearizes the non-linearized output representation using concentrations in the training dataset corresponding to the signals output from the plurality of sensors.

18. The system of claim 17, wherein the linearizing model comprises at least one feedforward neural network trained to linearize the non-linearized output representation using the signals output from the plurality of sensors and corresponding concentration values in the training dataset.

19. The system of claim 10, wherein training the first model and/or the second model comprises:

receiving a device-specific model trained using information acquired by a device other than the chemical sensing device; and

retraining the device-specific model using at least a portion of the training dataset.

20. A chemical sensing method, comprising:

receiving, by a chemical sensing system, signals output from a plurality of sensors arranged on at least one substrate, wherein a first sensor and a second sensor of the plurality of sensors have different sensitivities to sense at least one analyte in a sample, each of the plurality of sensors being configured to output a signal in response to sensing the at least one analyte;

determining, by the chemical sensing system, a concentration of the at least one analyte in the sample by;

providing the received signals as input to a first model trained to relate the signals output from the plurality of sensors to a feature representation to generate feature representation values;

providing the feature representation values as input to a second model trained to relate the feature representation to a latent representation to generate latent representation values, the latent representation having a lower dimensionality than the feature representation;

providing the latent representation values as input to a third model trained to relate the latent representation to a straightened orthogonal representation to generate straightened orthogonal representation values, wherein;

one-dimensional manifolds in the straightened orthogonal representation corresponding to varying concentrations of a same analyte have zero angle between them; and

one-dimensional manifolds in the straightened orthogonal representation corresponding to different analytes are orthogonal;

providing the straightened orthogonal representation values as input to a fourth model trained to relate the straightened orthogonal representation to an output representation having bases corresponding to analytes, the analytes including the at least one analyte; and

determining the concentration of the at least one analyte in the sample based on an output of the fourth model; and

provide, by the chemical sensing system, an indication of the determined concentration of the at least one analyte in the sample to a user.

21. The method of claim 20, wherein the first model comprises:

a mapping from the signals to response values; and

a weighting that relates the response values to the feature representation.

22. The method of claim 21, wherein the mapping assigns, to each response value, a normalized change in an amplitude of a corresponding signal and/or a rise time or a fall time of an amplitude of the corresponding signal.

23. The method of claim 20, wherein a change in a location of a point in the feature representation corresponding to an analyte concentration in the sample is a non-linear function of a change in the analyte concentration in the sample.

24. The method of claim 20, wherein

the first model and the second model are device-specific models;

and

the third model and the second model are device-independent models.

25. A computer-implemented training method, comprising:

receiving a training dataset including values associated with signals output from a plurality of sensors having different sensitivities to sense at least one analyte in a sample, each of the plurality of sensors configured to output the signal in response to sensing the at least one analyte;

training, using at least one computer processor, a set of models using the training dataset by;

training a first model to relate the signals output from the plurality of sensors to a feature representation;

training a second model to relate the feature representation to a latent representation, the latent representation having a lower dimensionality than the feature representation;

training a third model to relate the latent representation to a straightened orthogonal representation, wherein;

one-dimensional manifolds in the straightened orthogonal representation corresponding to varying concentrations of a same analyte have zero angle between them; and

one-dimensional manifolds in the straightened orthogonal representation corresponding to different analytes are orthogonal; and

training a fourth model to relate the straightened orthogonal representation to the output representation having bases corresponding to analytes; and

configuring a chemical sensing system to detect a concentration of the least one analyte in the sample using the trained set of models.

26. The method of claim 25, wherein:

training the first model comprises updating weights in a recurrent neural network or convolutional neural network;

training the second model comprises updating weights in at least a portion of an autoencoder or generative adversarial network; and

training the fourth model comprises updating weights in at least one feedforward neural network.

27. The method of claim 25, wherein:

a continuous manifold embedded in the feature representation contains the signals output from the plurality of sensors, the continuous manifold having a dimensionality equal to the dimensionality of the latent representation; and

the second model relates the continuous manifold to the latent representation.

28. The method of claim 25, wherein training the set of models using the training dataset further comprises automatically identifying manifolds corresponding to samples of varying concentrations of a same analyte using local gradient changes in the feature representation.

29. The method of claim 25, wherein training the fourth model comprises:

associating bases in the straightened orthogonal representation with the analytes using labels in the training data;

generating an introduction model that introduces the associated bases onto standard bases of a non-linearized output representation; and

generating a linearizing model that linearizes the non-linearized output representation using concentrations in the training dataset corresponding to the signals output from the plurality of sensors.

30. The method of claim 25, wherein training the first model and/or the second model comprises:

receiving a device-specific model trained using information acquired by a device other than the chemical sensing device; and

retraining the device-specific model using at least a portion of the training dataset.