# US 2024/0203530 A1 · Machine learning techniques to determine base methylations

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

- Number: US 2024/0203530 A1
- Document ID: US20240203530A1
- Status: Published application
- Kind: A1
- Application number: 18/542,251
- Filed: 2023-12-15
- Published: 2024-06-20
- Earliest claimed: 2022-12-16
- Assignees: Centre for Novostics
- Applicants: Centre for Novostics
- Inventors: Yuk-Ming Dennis Lo; Rossa Wai Kwun Chiu; Kwan Chee Chan; Peiyong Jiang; Xi Hu
- Web page: https://patentgrep.com/patent/US20240203530A1

## CPC

- G16B 40/20: Supervised data analysis
- C12Q 1/6869: Methods for sequencing
- G06N 3/044: Recurrent networks, e.g. Hopfield networks
- G06N 3/045: Combinations of networks
- G06N 3/0464: Convolutional networks [CNN, ConvNet]
- G06N 3/048: Activation functions
- G06N 3/08: Learning methods
- G06N 3/084: Backpropagation, e.g. using gradient descent
- G06N 3/09: Supervised learning
- G06N 20/00: Machine learning
- G16B 20/20: Allele or variant detection, e.g. single nucleotide polymorphism [SNP] detection
- G16B 30/00: ICT specially adapted for sequence analysis involving nucleotides or amino acids
- G16B: BIOINFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR GENETIC OR PROTEIN-RELATED DATA PROCESSING IN COMPUTATIONAL MOLECULAR BIOLOGY
- C12Q: MEASURING OR TESTING PROCESSES INVOLVING ENZYMES, NUCLEIC ACIDS OR MICROORGANISMS (immunoassay G01N33/53); COMPOSITIONS OR TEST PAPERS THEREFOR; PROCESSES OF PREPARING SUCH COMPOSITIONS; CONDITION-RESPONSIVE CONTROL IN MICROBIOLOGICAL OR ENZYMOLOGICAL PROCESSES
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS

## Abstract

Systems and methods determine based methylations in analyzing nucleic acid molecules. Embodiments may use kinetic signals produced by a DNA polymerase during single-molecule sequencing. The use of kinetic signals of adaptor sequences may allow for determining the methylation patterns proximal to the ends of a DNA fragment. Deep learning models that can capture local and global signal patterns can be trained to detect the base methylations using these features, with improved performance. Deep learning models may include a convolutional neural network that preferentially captures the features of local signal patterns, with the integration of a transformer model that preferentially captures the features of global signal patterns. The improved performance in the determination of base methylation may lead to more accurate diagnoses of subjects. Accurate measurement of DNA methylation may have several other clinical applications.

## Claims

### Claim 1

A method for detecting a methylation of a nucleotide in a nucleic acid molecule, the method comprising:
receiving data acquired by sequencing a sample nucleic acid molecule by measuring pulses in a signal corresponding to nucleotides of the sample nucleic acid molecule and obtaining, from the data, values for one or more signal properties;
creating an input data structure, the input data structure comprising a window around a target position of the nucleotides sequenced in the sample nucleic acid molecule, wherein the input data structure includes, for each nucleotide within the window, one or more values for the one or more signal properties;
inputting the input data structure into a model, wherein the model is a machine learning model, the model trained by:
receiving a first plurality of first data structures, each first data structure of the first plurality of first data structures corresponding to a respective window around a respective target position of nucleotides sequenced in a respective nucleic acid molecule of a plurality of first nucleic acid molecules, wherein each of the first nucleic acid molecules is sequenced by measuring pulses in the signal corresponding to the nucleotides, wherein the methylation has a known first state in a nucleotide at the respective target position in each window of each first nucleic acid molecule, each first data structure comprising values for the same properties as the input data structure,
storing a plurality of first training samples, each including one of the first plurality of first data structures and a first label indicating the first state of the nucleotide at the respective target position,
filtering the first plurality of first data structures through one or more convolutional layers to obtain a plurality of convolutional matrices,
applying a transformer layer to the plurality of convolutional matrices to obtain transformer matrices, wherein applying the transformer layer to a convolutional matrix includes generating a plurality of attention scores that quantify a relevance among positions of the convolutional matrix,
generating methylation probabilities at respective target positions of the first plurality of first data structures using the transformer matrices,
determining outputs using the methylation probabilities, and
optimizing, using the plurality of first training samples, parameters of the model based on the outputs of the model matching or not matching corresponding labels of the first labels when the first plurality of first data structures is input to the model, wherein an output of the model specifies whether the nucleotide at the respective target position in the respective window has the methylation, wherein the parameters of the model include the plurality of attention scores,
determining, using the model, whether the methylation is present in the nucleotide at the target position within the window in the input data structure.

### Claim 2 (depends on 1)

The method of claim 1, wherein the signal is an optical signal or an electrical signal.

### Claim 3 (depends on 1)

The method of claim 1, wherein the one or more signal properties include:
for each nucleotide within the window:
an identity of the nucleotide.

### Claim 4 (depends on 3)

The method of claim 3, wherein the one or more signal properties further include:
for each nucleotide within the window:
a position of the nucleotide within the sample nucleic acid molecule,
a width of a pulse corresponding to the nucleotide, or
an interpulse duration representing a time between the pulse corresponding to the nucleotide and a pulse corresponding to a neighboring nucleotide.

### Claim 5 (depends on 1)

The method of claim 1, wherein generating the plurality of attention scores comprises using a plurality of multiple-head self-attentions.

### Claim 6 (depends on 1)

The method of claim 1, wherein generating the methylation probabilities comprises applying one or more neural network layers to the transformer matrices.

### Claim 7 (depends on 6)

The method of claim 6, wherein applying the one or more neural network layers comprises performing multiplication by weights or additions by biases.

### Claim 8 (depends on 1)

The method of claim 1, wherein the respective convolutional results have lower dimensionality than the respective first data structure.

### Claim 9 (depends on 1)

The method of claim 1, wherein the methylation is 5 mC (5-methylcytosine).

### Claim 10 (depends on 1)

The method of claim 1, wherein the methylation is 6 mA (N6-methyladenine).

### Claim 11 (depends on 1)

The method of claim 1, wherein the window comprises 13 consecutive nucleotides.

### Claim 12 (depends on 1)

The method of claim 1, wherein the window of the input data structure has a different number of consecutive nucleotides upstream of the nucleotide at the target position than the number of consecutive nucleotides downstream of the nucleotide at the target position.

### Claim 13 (depends on 1)

The method of claim 1, wherein the window of the input data structure comprises 21 consecutive nucleotides upstream of the nucleotide at the target position and 21 consecutive nucleotides downstream of the nucleotide at the target position.

### Claim 14 (depends on 1)

The method of claim 1, wherein:
the data is acquired by sequencing an extended nucleic acid molecule,
the extended nucleic acid molecule comprises the sample nucleic acid molecule and an adaptor,
the adaptor has a known sequence, and
the window of nucleotides includes at least one nucleotide in the adaptor.

### Claim 15 (depends on 1)

The method of claim 1, wherein determining whether the methylation is present comprises:
determining the methylation is present; and
determining the methylation is a first type from among a plurality of types.

### Claim 16 (depends on 15)

The method of claim 15, wherein each type of the plurality of types is selected from the group consisting of 5 mC, 5 hmC, and 6 mA.

### Claim 17 (depends on 1)

The method of claim 1, wherein the sample nucleic acid molecule is single-stranded.

### Claim 18 (depends on 17)

The method of claim 17, wherein the plurality of first nucleic acid molecules comprises single-stranded nucleic acid molecules.

### Claim 19

A method for detecting a methylation of a nucleotide in a nucleic acid molecule, the method comprising:
receiving a first plurality of first data structures, each first data structure of the first plurality of first data structures corresponding to a respective window of positions around a respective target position of nucleotides sequenced in a respective nucleic acid molecule of a plurality of first nucleic acid molecules, wherein each of the first nucleic acid molecules is sequenced by measuring pulses in a signal corresponding to the nucleotides, wherein the methylation has a known first state in a nucleotide at the respective target position in each window of each first nucleic acid molecule, each first data structure comprising values for one or more signal properties at positions within the respective window;
storing a plurality of first training samples, each first training sample including one of the first plurality of first data structures and a first label indicating the first state of the nucleotide at the respective target position; and
training a model by:
filtering the first plurality of first data structures through one or more convolutional layers to obtain a plurality of convolutional matrices,
applying a transformer layer to the plurality of convolutional matrices to obtain transformer matrices, wherein applying the transformer layer to a convolutional matrix includes generating a plurality of attention scores that quantify a relevance among positions of the convolutional matrices,
generating methylation probabilities at respective target positions of the first plurality of first data structures using the transformer matrices,
determining outputs using the methylation probabilities, and
optimizing, using the plurality of first training samples, parameters of the model based on the outputs of the model matching or not matching corresponding labels of the first labels when the first plurality of first data structures is input to the model, wherein an output of the model specifies whether the nucleotide at the respective target position in the respective window has the methylation, wherein the parameters of the model include the plurality of attention scores.

### Claim at position 20

20-52. (canceled)

### Claim 53

A method for detecting a methylation of a nucleotide in a nucleic acid molecule, the method comprising:
receiving a first plurality of first data structures, each first data structure of the first plurality of first data structures corresponding to a respective window around a respective target position of nucleotides sequenced in a respective nucleic acid molecule of a plurality of first nucleic acid molecules, wherein each of the first nucleic acid molecules is sequenced by measuring pulses in a signal corresponding to the nucleotides, wherein each first nucleic acid molecule comprises a training sample nucleic acid molecule and a first adaptor having a known sequence, wherein the methylation has a known first state in a nucleotide at the respective target position in a portion of each window of each first nucleic acid molecule corresponding to the training sample nucleic acid molecule, each first data structure comprising values for one or more signal properties;
storing a plurality of first training samples, each including one of the first plurality of first data structures and a first label indicating the first state of the nucleotide at the respective target position, and
training a model by optimizing, using the plurality of first training samples, parameters of the model based on outputs of the model matching or not matching corresponding labels of the first labels when the first plurality of first data structures is input to the model, wherein an output of the model specifies whether the nucleotide at the respective target position in the respective window has the methylation.

### Claim at position 22

54-85. (canceled)
