# US 11,030,275 B2 · Modelling ordinary differential equations using a variational auto encoder

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 11,030,275 B2
- Document ID: US11030275B2
- Status: Granted
- Kind: B2
- Application number: 16/255,778
- Filed: 2019-01-23
- Published: 2020-07-23
- Granted: 2021-06-08
- Earliest claimed: 2019-01-18
- Assignees: Microsoft Technology Licensing, LLC
- Applicants: Microsoft Technology Licensing, LLC
- Inventors: Edward Meeds; Geoffrey Roeder; Neil Dalchau
- Web page: https://patentgrep.com/patent/US11030275B2

## CPC

- G06N 3/084: Backpropagation, e.g. using gradient descent
- G06F 17/13: Differential equations (using digital differential analysers G06F7/64)
- G06F 18/214: Generating training patterns; Bootstrap methods, e.g. bagging or boosting
- G06F 18/2411: based on the proximity to a decision surface, e.g. support vector machines
- G06F 18/2414: {Smoothing the distance, e.g. radial basis function networks [RBFN]}
- G06N 3/045: Combinations of networks
- G06N 3/0455: Auto-encoder networks; Encoder-decoder networks
- G06N 3/0464: Convolutional networks [CNN, ConvNet]
- G06N 3/047: Probabilistic or stochastic networks
- G06N 3/0475: Generative networks
- G06N 3/08: Learning methods
- G06V 10/82: using neural networks
- G06V 20/698: {Matching; Classification}
- G06N 3/044: Recurrent networks, e.g. Hopfield networks
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)
- G06V: IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING

## Abstract

A computer-implemented method comprising: from each of multiple trials, obtaining a respective series of observations y(t) of a subject over time t; using a variational auto encoder to model an ordinary differential equation, ODE, wherein the variational auto encoder comprises an encoder for encoding the observations into a latent vector z and a decoder for decoding the latent vector, the encoder comprising a first neural network and the decoder comprising one or more second neural networks, wherein the ODE as modelled by the decoder has a state x(t) representing one or more physical properties of the subject which result in the observations y, and the decoder models a rate of change of x with respect to time t as a function f of at least x and z: dx/dt=f(x, z); and operating the variational auto encoder to learn the function f based on the obtained observations y.

## Claims

### Claim 1

A computer-implemented method comprising:
from each of multiple trials, obtaining a respective series of observations y(t) of a subject over time t;
using a variational auto encoder to model an ordinary differential equation, ODE, wherein the variational auto encoder comprises an encoder for encoding the observations into a latent vector z and a decoder for decoding the latent vector, the encoder comprising at least a first neural network and the decoder comprising one or more second neural networks, wherein the ODE as modelled by the decoder has a state x(t) representing one or more physical properties of the subject which result in the observations y, and the decoder models a rate of change of x with respect to time t as a function f of at least x and z: dx/dt=f(x, z); and
operating the variational auto encoder to learn the function f based on the obtained observations y by tuning weights of the first and second neural networks, the learning comprising performing variational inference over the latent vector z, whereby z has a plurality of dimensions and in the encoder each dimension is constrained to comprising one or more samples of a predetermined type of probabilistic distribution parameterized by a respective one or more parameters.

### Claim 2 (depends on 1)

The method of claim 1, wherein said learning comprises minimizing a measure of difference between the observations y and corresponding output values ŷ, the output values ŷ being predicted versions of the corresponding observations y related to the decoder state x by ŷ=g(x), where g is a predetermined observation function estimating a relation between the observations and said one or more properties.

### Claim 3 (depends on 1)

The method of claim 1, wherein the multiple trials comprise trials under conditions of different variants d of the subject and/or different external signals u applied to the subject, where f is further modelled as a function of one or both of d and u:
dx/dt=f(x,z,d),dx/dt=f(x,z,u) or dx/dt=f(x,z,d,u).

### Claim 4 (depends on 3)

The method of claim 3, comprising, following said learning, either:
a) for a given candidate variant d, external signal u or combination thereof, predicting what the time profile of x or y would be caused by the given u and/or d; or
b) for a desired time profile of x or y, inferring what variant d, external signal u or combination thereof would cause the desired time profile.

### Claim 5 (depends on 4)

The method of claim 4, comprising implementing a copy of the subject with the predicted or inferred u and/or d.

### Claim 6 (depends on 1)

The method of claim 1, wherein the state vector x is a state vector of at least one of the one or more second neural networks in the decoder.

### Claim 7 (depends on 1)

The method of claim 1, wherein the distributions from which at least some of the dimensions of z are sampled are dependent on x.

### Claim 8 (depends on 1)

The method of claim 1, wherein:
the latent vector z generated by the encoder comprises at least a local vector encoded from the observations y by the first neural network, each dimension of the local vector comprising one or more samples of a first predetermined type of a probabilistic distribution which represents a variation per observation, the local probabilistic distribution being dependent on x and d.

### Claim 9 (depends on 8)

The method of claim 8, wherein:
the latent vector z further comprises a device-dependent vector, each dimension of the device-dependent vector comprising one or more samples of a second predetermined type of a probabilistic distribution which represents a variation in the observations per device, the second predetermined type of the probabilistic distribution being a device-dependent distribution and dependent on d but not x; and/or
the latent vector z further comprises a global vector, each dimension of the global vector comprising one or more samples of a third predetermined type of a probabilistic distribution which represents a global variation in the observations, the third predetermined type of the probabilistic distribution being a global distribution being dependent on neither d nor x.

### Claim 10 (depends on 1)

The method of claim 1, wherein in the decoder, the function f is modelled in the form of a known analytical equation, the equation comprising a plurality of parts; and wherein the one or more second neural networks comprise a plurality of second neural networks, each of said parts being modelled by at least a respective one of the second neural networks.

### Claim 11 (depends on 10)

The method of claim 10, wherein said form comprises dx/dt=f1−x·f2, where each of f1 and f2 are modeled as at least a respective one of the second neural networks, and each is a function of x and z and optionally one or both of d and u.

### Claim 12 (depends on 1)

The method of claim 1, wherein the subject comprises a device of a genetic circuit inserted into a cell for producing one or more proteins, the device comprising one or more cassettes each comprising at least one respective gene and at least one respective promoter; and wherein the one or more properties comprises one or more of:
a density of copies of the cell in a population from which the observations are observed in the respective trial, and/or
an intercellular concentration of any one or more proteins or other species between cells in the population from which the observations are observed in the respective trial.

### Claim 13 (depends on 12)

The method of claim 12, wherein the one or more proteins comprises a florescent reporter protein and the observations comprises a measure of an amount of light received from the fluorescent protein.

### Claim 14 (depends on 12)

The method of claim 12, wherein:
the multiple trials comprise trials under conditions of different variants d of the device and/or different external signals u applied to the device, where f is further modelled as a function of one or both of d and u: dx/dt=f(x,z,d), dx/dt=f(x,z,u) or dx/dt=f(x,z,d,u); and
the variants d of the device comprise different cassettes or different combinations of cassettes.

### Claim 15 (depends on 12)

The method of claim 12, wherein:
the multiple trials comprise trials under conditions of different variants d of the device and/or different external signals u applied to the devices, where f is further modelled as a function of one or both of d and u: dx/dt=f(x,z,d), dx/dt=f(x,z,u) or dx/dt=f(x,z,d,u); and
the different external signals u comprise one or more different treatments or combinations of treatments applied to the device.

### Claim 16 (depends on 12)

The method of claim 12, wherein:
the multiple trials comprise trials under conditions of different variants d of the device and/or different external signals u applied to the device, where f is further modelled as a function of one or both of d and u: dx/dt=f(x,z,d), dx/dt=f(x,z,u) or dx/dt=f(x,z,d,u);
the variants d of the device comprise different cassettes or different combinations of cassette, and/or the different external signals u comprise one or more different treatments or combinations of treatments applied to the device; and
the method comprises, following said learning, either:
a) for a given candidate of the variants d of the device, treatment or treatments u, or combination of the variants d of the device and treatment or treatments u, predicting what time profile of x or y would be caused by the candidate of the variants d of the device, treatment or treatments, or combination; or
b) for a desired time profile of x or y, inferring what the variants d of the device, treatment or treatments u, or combination of the variants d of the device and treatment or treatments u, would cause the desired time profile.

### Claim 17 (depends on 16)

The method of claim 16, comprising, in vitro:
engineering a genetic circuit including the candidate or the inferred variants of the device in a cell, and thereby causing the cell to produce said one or more proteins with the predicted or desired time profile, respectively;
engineering a genetic circuit including the variants d of the device in a cell, applying the candidate or inferred treatment or treatments to the variants d of the device, and thereby causing the cell to produce said one or more proteins with the predicted or desired time profile, respectively;
engineering a genetic circuit including the candidate or the inferred variants d of the device in a cell, applying the candidate or inferred treatment or treatments to the variants d of the device, and thereby causing the cell to produce said one or more proteins with the predicted or desired time profile, respectively;
engineering a genetic circuit including in a cell a first additional device approximated by the candidate or inferred variants d of the device, and thereby causing the cell to produce at least one protein resulting from the first additional device;
engineering a genetic circuit including in a cell a second additional device approximated by the variants d of the device, and applying the candidate or inferred treatment or treatments to the variants d of the device, thereby causing the cell to produce at least one protein resulting from the second additional device; or
engineering a genetic circuit including in a cell a third additional approximated by the candidate or the inferred variants d of the device, applying the candidate or inferred treatment or treatments to the variants d of the device, and thereby causing the cell to produce at least one protein resulting from the third additional device.

### Claim 18 (depends on 1)

A computer program on a memory of a computer and configured so as when run on the computer to perform the method of claim 1.

### Claim 19

A computer system comprising storage and one or more processors, the storage storing code arranged to run on at least one of the processors, wherein the code is configured so as when run on the at least one processor to perform operations of:
from each of multiple trials, obtaining a respective series of observations y(t) of a subject over time t;
implementing a variational auto encoder to model an ordinary differential equation, ODE, wherein the variational auto encoder comprises an encoder for encoding the observations into a latent vector z and a decoder for decoding the latent vector, the encoder comprising at least a first neural network and the decoder comprising one or more second neural networks, wherein the ODE as modelled by the decoder has a state x(t) representing one or more physical properties of the subject which result in the observations y, and the decoder models a rate of change of x with respect to time t as a function f of at least x and z: dx/dt=f(x, z); and
operating the variational auto encoder to learn the function f based on the obtained observations y by tuning weights of the first and second neural networks, the learning comprising performing variational inference over the latent vector z, whereby z has a plurality of dimensions and in the encoder each dimension is constrained to comprising one or more samples of a predetermined type of probabilistic distribution parameterized by a respective one or more parameters.

### Claim 20

A genetic engineering facility comprising:
a computer system comprising storage and one or more processors, the storage storing code arranged to run on at least one of the processors, wherein the code is configured so as when run on the at least one processor to perform operations of:
from each of multiple trials, obtaining a respective series of observations y(t) of a subject over time t;
implementing a variational auto encoder to model an ordinary differential equation, ODE, wherein the variational auto encoder comprises an encoder for encoding the observations into a latent vector z and a decoder for decoding the latent vector, the encoder comprising at least a first neural network and the decoder comprising one or more second neural networks, wherein the ODE as modelled by the decoder has a state x(t) representing one or more physical properties of the subject which result in the observations y, and the decoder models a rate of change of x with respect to time t as a function f of at least x and z: dx/dt=f(x, z); and
operating the variational auto encoder to learn the function f based on the obtained observations y by tuning weights of the first and second neural networks, the learning comprising performing variational inference over the latent vector z, whereby z has a plurality of dimensions and in the encoder each dimension is constrained to comprising one or more samples of a predetermined type of probabilistic distribution parameterized by a respective one or more parameters;
wherein the subject comprises a device of a genetic circuit inserted into a cell for producing one or more proteins, the device comprising one or more cassettes each comprising at least one respective gene and at least one respective promoter; and wherein the one or more properties comprises one or more of: a density of copies of the cell in a population from which the observations are observed in the respective trial, and/or an intercellular concentration of any one or more proteins or other species between cells in the population from which the observations are observed in the respective trial;
wherein the multiple trials comprise trials under conditions of different variants d of the device and/or different external signals u applied to the device, where f is further modelled as a function of one or both of d and u: dx/dt=f(x,z,d), dx/dt=f(x,z,u) or dx/dt=f(x,z,d,u), the variants d of the device comprising different cassettes or different combinations of cassette, and/or the different external signals u comprising one or more different treatments or combinations of treatments applied to the device; and
wherein the code is further configured so as, following said learning, to either:
a) for a given candidate variant of a device d, treatment or treatments u applied to the device, or combination of the variant of the device d and treatment or treatments u, predicting what time profile of x or y would be caused by the candidate device, treatment or treatments, or combination; or
b) for a desired time profile of x or y, inferring what the variant of the device d, treatment or treatments u, or combination of the variant of the device d and treatment or treatments u, would cause the desired time profile;
wherein the genetic engineering facility further comprises in vitro genetic engineering equipment arranged to produce a genetic circuit based on the running of the program, the production comprising:
engineering a genetic circuit including the candidate or the inferred variant of the device d in a cell, and thereby causing the cell to produce said one or more proteins with the predicted or desired time profile, respectively;
engineering a genetic circuit including the variant of the device d in a cell, applying the candidate or inferred treatment or treatments to the device, and thereby causing the cell to produce said one or more proteins with the predicted or desired time profile, respectively;
engineering a genetic circuit including the candidate or the inferred variant of the device d in a cell, applying the candidate or inferred treatment or treatments to the variant of the device d, and thereby causing the cell to produce said one or more proteins with the predicted or desired time profile, respectively;
engineering a genetic circuit including in a cell a first additional device approximated by the candidate or inferred variant of the device d, and thereby causing the cell to produce at least one protein resulting from the first additional device;
engineering a genetic circuit including in a cell a second additional device approximated by the variant of the device d, and applying the candidate or inferred treatment or treatments to the variant of the device d, thereby causing the cell to produce at least one protein resulting from the second additional device; or
engineering a genetic circuit including in a cell a third additional device approximated by the candidate or the inferred variant of the device d, applying the candidate or inferred treatment or treatments to the variant of the device d, and thereby causing the cell to produce at least one protein resulting from the third additional device.