# US 11,308,101 B2 · Multi-resolution modeling of discrete stochastic processes for computationally-efficient information search and retrieval

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

- Number: US 11,308,101 B2
- Document ID: US11308101B2
- Status: Granted
- Kind: B2
- Application number: 17/459,346
- Filed: 2021-08-27
- Published: 2022-03-24
- Granted: 2022-04-19
- Earliest claimed: 2020-09-19
- Applicants: Bonnie Berger Leighton; Maxwell Aaron Sherman; Adam Uri Yaari
- Inventors: Bonnie Berger Leighton; Maxwell Aaron Sherman; Adam Uri Yaari
- Web page: https://patentgrep.com/patent/US11308101B2

## CPC

- G06F 16/2458: Special types of queries, e.g. statistical queries, fuzzy queries or distributed queries
- G06F 16/2462: {Approximate or statistical queries}
- G06N 3/045: Combinations of networks
- G06N 3/0455: Auto-encoder networks; Encoder-decoder networks
- G06N 3/0464: Convolutional networks [CNN, ConvNet]
- G06N 3/08: Learning methods
- G06N 3/09: Supervised learning
- G16B 5/20: Probabilistic models
- G16B 20/20: Allele or variant detection, e.g. single nucleotide polymorphism [SNP] detection
- G16B 40/00: ICT specially adapted for biostatistics; ICT specially adapted for bioinformatics-related machine learning or data mining, e.g. knowledge discovery or pattern finding
- G16B 40/20: Supervised data analysis
- G16H 50/20: for computer-aided diagnosis, e.g. based on medical expert systems
- G06N 3/048: Activation functions
- G06N 3/126: Evolutionary algorithms, e.g. genetic algorithms or genetic programming
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G16B: BIOINFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR GENETIC OR PROTEIN-RELATED DATA PROCESSING IN COMPUTATIONAL MOLECULAR BIOLOGY
- G16H: HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA

## Abstract

An activity of interest is modeled by a non-stationary discrete stochastic process, such as a pattern of mutations across a cancer genome. Initially, input genomic data is used to train a model to predict rate parameters and their associated uncertainty estimation for each of a set of process regions. For any arbitrary set of indexed positions of the stochastic process that are identified in an information query, the rate parameters and their associated estimation uncertainties are scaled using the model to obtain a distribution of the events of interest and their associated estimation uncertainties for the set of indexed positions. In one practical application, and in response to a search query associated with one or more base-pairs, a result is then returned. The result, which represents deviations between the estimated and observed mutation rates, is used to identify genomic elements that have more mutations than expected and therefore constitute previously unknown driver mutations.

## Claims

### Claim 1

A method of information search and retrieval associated with a non-stationary discrete stochastic process representing a pattern of mutations across a genome of a cancer of interest, wherein events of interest are cancer-specific mutations that occur within the process approximately independently at units i within a region R, and with an unknown rate λR that is approximately constant across R, wherein λR has an associated estimation uncertainty defined by a set of parameters that include expectation μR and variance σR2, and wherein the non-stationary discrete stochastic process comprising a plurality of regions, wherein at least one region is a non-coding region of the genome, comprising:
responsive to a set of input data having a first dimensionality, training a model for the cancer of interest a single time to predict rate parameters and their associated estimation uncertainty for each region of the plurality of regions, the model being trained by:
processing the input data through a first process that maps patterns in the input data to a reduced data set having a second dimensionality that is at least an order of magnitude less than the first dimensionality, and
processing the reduced set through a second process to output a distribution of the events of interest and their associated estimation uncertainty for each region of the plurality of regions;
receiving a query associated with any arbitrary set of indexed positions within the plurality of regions;
responsive to receipt of the query, and by scaling existing predictions of the rate parameters and their associated estimation uncertainties from the trained model for any region of the plurality of regions, obtaining a response to the query, the response being a distribution of the events of interest and their associated estimation uncertainties for the set of indexed positions; and
providing the response for use in a target application.

### Claim 2 (depends on 1)

The method as described in claim 1 wherein the query is processed using a function computed in constant time, wherein the function represents a closed form solution to a marginalization of a multi-variate probability distribution for the set of indexed positions.

### Claim 3 (depends on 2)

The method as described in claim 2 wherein the closed form solution to the marginalization of the multi-variate probability distribution is a variant of a Poisson-Gamma (PG) distribution.

### Claim 4 (depends on 1)

The method as described in claim 1 wherein the first process is a Convolutional Neural Network (CNN), and the second process is a Gaussian Process (GP).

### Claim 5 (depends on 1)

The method as described in claim 1 wherein the first process is one of: a Convolutional Neural Network (CNN), an Auto-Encoder (AE), Principal Component Analysis (PCA), and Uniform Manifold Approximation and Projection (UMAP) learning.

### Claim 6 (depends on 1)

The method as described in claim 1 wherein the second process is one of: a Gaussian Process, a Random Forest (RF) regression, a Negative Binomial Regression (NBR); and a Binomial Regression (BR).

### Claim 7 (depends on 1)

The method as described in claim 1 wherein the input data is a set of genomic feature data.

### Claim 8 (depends on 7)

The method as described in claim 7 wherein patterns in the input data comprise epigenetic tracks.

### Claim 9 (depends on 8)

The method as described in claim 8 wherein an epigenetic track comprises a chromatin track that measures an abundance of a particular chromatin mark genome-wide.

### Claim 10 (depends on 4)

The method as described in claim 4 wherein the CNN generates the reduced data set by performing a non-linear feature space reduction over information in the local epigenetic patterns.

### Claim 11 (depends on 1)

The method as described in claim 1 wherein estimating the distribution of events includes formulating a closed-form generative model for a number of events of interest at any position within a particular region.

### Claim 12 (depends on 11)

The method as described in claim 11 wherein the closed-form generative model is formulated using predicted regional rate parameters and their associated estimation uncertainty, together with local context information.

### Claim 13 (depends on 12)

The method as described in claim 12 wherein the local context information is an observed genetic context of the particular region.

### Claim 14

An information search and retrieval apparatus for use with an input dataset that is a non-stationary discrete stochastic process representing a pattern of mutations across a genome of a cancer of interest, wherein events of interest are cancer-specific mutations that occur within the process approximately independently at units i within a region R, and with an unknown rate λR that is approximately constant across R, wherein λR has an associated estimation uncertainty defined by a set of parameters that include expectation μR and variance σR2, and wherein the non-stationary discrete stochastic process comprising a plurality of regions, wherein at least one region is a non-coding region of the genome, comprising:
a hardware processor; and
computer memory holding computer program code executed by the hardware processor, the computer program code comprising:
a neural network that uses the input dataset to one-time train a model for the cancer of interest to predict rate parameters and their associated estimation uncertainty for each region of the plurality of regions; and
a query function that receives a query associated with any arbitrary set of indexed positions within the plurality of regions and, in response, scales existing predictions of the rate parameters and their associated estimation uncertainties from the trained model for any region of the plurality of regions to obtain a response to the query, the response being a distribution of the events of interest and their associated estimation uncertainties for the set of indexed positions; and
an output function that applies data derived from the response to a target application.

### Claim 15

A computer program product in a non-transitory computer readable medium for use in a data processing system, the computer program product holding computer program instructions which, when executed by the data processing system, are operative to perform information search and retrieval with respect to a non-stationary discrete stochastic process representing a pattern of mutations across a genome for a cancer of interest, wherein events of interest are cancer-specific mutations that occur within the process approximately independently at units i within a region R, and with an unknown rate λR that is approximately constant across R, wherein λR has an associated estimation uncertainty defined by a set of parameters that include expectation μR and variance σR2, and wherein the non-stationary discrete stochastic process comprises a plurality of regions, wherein at least one region is a non-coding region of the genome, the computer program instructions comprising program code configured to:
train a model for the cancer of interest a single time to predict rate parameters and their associated estimation uncertainty for each region of the plurality of regions;
receiving a set of one or more queries, wherein a query is associated with any arbitrary set of indexed positions within the plurality of regions;
responsive to receipt of each query, scale existing predictions of the rate parameters and their associated estimation uncertainties from the trained model for any region of the plurality of regions to obtain a response to the query, the response being a distribution of the events of interest and their associated estimation uncertainties for the set of indexed positions; and
provide the response for use in a target application.

### Claim 16 (depends on 15)

The computer program product as described in claim 15 wherein the program code configured to train the model comprises a neural network and a Gaussian process.