Medical decision support system using protein and DNA language models

Abstract

A medical decision support system using protein and DNA language models. Both the DNA language model and the protein language model are used to generate scores for medical decision support based on the patient's available genomic data. These scores are used to predict the health of the patient. Decision support is provided, for example, by scoring the expression of genes and the functionality of the encoded proteins, and inputting the scores into a predictive model for predicting the health of the patient.

Description

Background Art

[0001] The present embodiment relates to medical decision support based on genomic data. The genetic composition of cells contributes to or even determines the cause of disease. For most common complex diseases, polygenic inheritance, involving many common small-effect genetic variants, plays a greater role than rare single gene mutations, thereby causing a wide range of individual disease risks. In cancer, the focus has been on driver mutations in marker oncogenes and tumor suppressor genes, which damage the function of encoded proteins and eventually transform normal cells into malignant cells. In both cases, genotyping techniques such as next-generation sequencing are increasingly adopted to support clinical decision making. However, the size of the genome domain is combined with a small patient cohort and inherent genetic heterogeneity (i.e., different genotypes that cause the same phenotype), making it difficult to appropriately train a clinical decision support system that considers all relevant genomic information. For example, the impact of non-coding mutations in regulatory sequences (such as promoter and enhancer sequences) that determine the expression of genes (rather than the functionality of encoded proteins) has not yet been considered on a large scale.

[0002] In order to fully characterize the genetic component of a disease, and thus improve diagnosis, prognosis, and prediction, for example, with respect to normal tissue complications or toxicity caused by treatments such as radiation therapy, it is essential to consider the combined effects of protein-coding and non-coding mutations, including both inherited (i.e., germline) and acquired (i.e., somatic) mutations, on both protein functionality and expression. Although the expression of genes and their encoded proteins can also be measured directly using transcriptomic and proteomic approaches, genomics is by far the most clinically recognized and most often the only molecular modality available in clinical practice.

[0003] For common complex diseases, such as coronary artery disease, genome-wide polygenic risk scores (PRS) identify high-risk individuals and, therefore, enable enhanced screening or preventive therapy, for example, using the LDPred algorithm. However, PRS typically does not consider the effects of variants in the context of genes and gene regulatory sequences, let alone the interactions of encoded proteins. In cancer, the standard approach to developing prediction models based on gene mutations is to preselect task-specific gene / mutation sets that have been previously identified as being independently associated with phenotypes, and to limit the training of basic prediction models to this limited set, for example, learning Cox regression models from mutations in 11 genes to predict responses to immunotherapy. Combinations of somatic mutations that predict cancer survival and immunotherapy benefits have also been identified using unsupervised restricted Boltzmann machines. Genomic alterations determined using a more comprehensive sequencing panel targeting up to 468 cancer-associated genes and features inferred therefrom have been used to construct random forest classifiers to predict tumor origin sites. Regulatory sequences, such as promoters and enhancer elements, are generally not considered, with a few exceptions, such as TERT promoter mutations.

[0004] Additionally, there are methods to integrate genomic data with more functional readouts such as gene and protein expression data. For example, neural network models have been trained from multiple omics modalities to support clinical decision making. Some of these methods also integrate prior knowledge on direct or indirect protein-protein interactions. However, all of these methods require access to data from multiple modalities, including transcriptomics and proteomics, during training and in clinical practice. Summary of the invention

[0005] Provide systems, methods and instructions on a computer-readable medium for medical decision support. Based on the available genomic data of the patient, both the DNA language model and the protein language model are used to generate scores. These scores are used to predict the health of the patient. For example, by scoring the expression of the gene and the functionality of the encoded protein and inputting the score into the prediction model to provide decision support.

[0006] In a first aspect, a method for decision support in a medical system is provided. A genomic sequence of a patient is obtained. A first score of protein function is generated by a protein language model implemented by a processor based on an input derived from the genomic sequence. A second score of gene expression is generated by a regulatory DNA language model implemented by a processor based on an input derived from the genomic sequence. A machine learning model is used to predict the health of the patient in response to input of the first score and the second score to the machine learning model. The health as predicted by the machine learning model is displayed.

[0007] In a second aspect, a medical decision support system is provided. A memory is configured to store information of a patient's genomic sequence, a first language model for expression of a gene, a second language model for functionality of an encoded protein, and a neural network for prediction of a medical condition or disease. A processor is configured to apply the first language model to at least some of the information, apply the second language model to at least some of the information, and apply a neural network to outputs of the first and second language models, the neural network being configured to output decision support information for the patient from application of the neural network. A display is configured to display the decision support information or information derived from the decision support information.

[0008] In a third aspect, a method for analyzing genomic sequence data is provided. The genomic sequence data of a patient is obtained. The protein coding gene sequence contained in the genomic sequence data is translated into the encoded protein sequence. A protein language model is applied to the encoded protein sequence. The application results in a first metric score for evaluating the functionality of the corresponding protein. A DNA sequence that exerts a regulatory effect on the expression of the protein coding gene is extracted from the genomic sequence data. The regulatory DNA language model is applied to the extracted DNA sequence. The application results in a second metric score for evaluating the expression of the corresponding protein coding gene. A diagnosis, prognosis or prediction output about a disease or medical condition is derived by a machine-trained prediction model in response to the input of the first and second metric scores to the machine-trained prediction model. The diagnosis, prediction or prediction output is displayed.

[0009] Any one or more of the aspects or concepts described above or in the illustrative embodiments below may be used alone or in combination. An aspect or concept described for one embodiment or aspect may be used in other embodiments or aspects. An aspect or concept described for a method or system may be used in other methods or systems in a system, method, or non-transitory computer-readable storage medium.

[0010] These and other aspects, features, embodiments and advantages will become apparent from the following detailed description of the preferred embodiments, which is to be read in conjunction with the accompanying drawings. The present invention is defined by the following claims, and nothing in this section should be construed as a limitation on those claims. Further aspects and advantages of the present invention are discussed below in conjunction with the preferred embodiments, and may be claimed later, either independently or in combination. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] The components and the figures are not necessarily to scale, emphasis instead being placed on illustrating the principles of the embodiments. Moreover, in the figures, like reference numerals designate corresponding parts throughout the different views.

[0012] Figure 1is a flow chart of one embodiment of a method for medical decision support using a language model and genomic data;

[0013] Figure 2 illustrates an example graph neural network for making predictions from language model scores; and

[0014] Figure 3 is a block diagram of one embodiment of a system for medical decision support. DETAILED DESCRIPTION

[0015] For genomics-based decision support, various features can be useful. For example, depending exclusively on genome sequence data as input enables clinical applicability. As another example, consider all genome sequences together to determine the expression of genes and the functional variants of encoded proteins. As another example, consider the prior knowledge of the interaction between proteins to enable robust model training on a small patient cohort. Use one or more of these features in a general system to have the ability to make many diseases and clinically relevant downstream tasks adjustable, to improve patient management, such as the prognosis or prediction of therapeutic response.

[0016] Features can be provided by using a molecular language model. A combination of estimates of both the expression of a gene and the functionality of the encoded protein is input to a predictive model. As a result, improved clinical decision support for a disease or medical condition can be provided based on genomic sequence data alone, based on gene expression and protein functionality predicted therefrom, and in the case of providing protein-protein interactions.

[0017] Molecular sequences, such as protein and DNA sequences, can be scored by transformer architectures configured as language models. For example, protein language models are trained on large databases of protein sequences (e.g., UniRef90) with or without supervision from multiple protein sequence alignments and experimental measurements of function to capture the functional effects of protein sequence variation. Similarly, DNA regulatory sequences (including distal enhancers, repressors, and / or insulators) that can affect gene expression from sites greater than 20 kb away from the transcription start site, a limitation of convolutional network-based algorithms for predicting gene expression, can be effectively integrated by the increased receptive field of up to 100 kb transformer layers in DNA language models.

[0018] Based on (for example, individually or in part based on) genome sequence data, clinical decision making for a subject suffering from a disease or medical condition is improved. In order to consider the expression and functional variants of protein coding genes determined together in the genome sequence, multiple language models are applied to the corresponding genome sequence. The protein language model determines a metric score corresponding to the degree of functionality of each protein based on the protein sequence of each protein. The regulatory DNA language model determines a metric score corresponding to the degree of expression of each gene based on the regulatory sequence of each gene. The metric score is used as an input to a prediction model (for example, a neural network), which has been trained, or alternatively trained in the context of a disease or medical condition about diagnosis, prognosis or prediction conclusion.

[0019] Figure 1 It is a flow chart of an embodiment of a method for decision support in a medical system. The method includes analysis of genomic sequence data. Multiple language models score from genome-related data, such as scoring protein function, gene expression (including RNA and / or protein expression), RNA function, protein interaction or other molecular characteristics. The score is used to predict the patient's health, such as disease, medical condition, prognosis, therapy response and / or diagnosis.

[0020] The method is performed in the order shown (e.g., from top to bottom or numerical order), but other orders may also be used. For example, action pairs 130, 135 and 140, 145 are performed in any order. As another example, action 120 is performed after action 135 and / or 145 or before and after both.

[0021] Additional, different or fewer actions may be provided. For example, actions 120, 140, 160 and / or 170 are not performed. As another example, an action for extracting different gene sequences, such as functional RNA encoding genes, is provided.

[0022] The method is performed by a medical system, workstation, server or computer. A DNA sequencer or other genotyping instrument can be used to obtain genome sequence data. A processor performs various actions to predict health from genome sequence data. A display displays the prediction. Other devices, circuits or equipment can be used.

[0023] In action 100, the processor obtains the patient's genome sequence. The patient's genome sequence data is obtained from the patient by a sequencer or other genotyping instrument. The genome sequence data can be from a tumor sample (i.e., nucleic acid can be extracted from a tumor, tissue, urine, blood, semen, liquid, plasma, serum, umbilical cord or other sample of the subject). Liquid biopsy samples derived from blood, urine or other body fluids can be used. Alternatively, the genome sequence data is obtained from a memory or transmitted on a computer network, such as obtained from a previously performed sample assay.

[0024] Only genomic sequence data is used for prediction. In other approaches, genomic sequence data and other data, such as from the patient's electronic medical records, are acquired and used for prediction.

[0025] Genomic sequence data include targeted gene panel sequencing, whole exome sequencing, whole genome sequencing, whole genome genotyping microarray, and / or imputation of unobserved genotypes such as gene (i.e., DNA) information of a selected genome (panel). Nucleotide sequences of amplified clones of individual nucleic acid molecules or individual nucleic acid molecules are obtained. The panel and resulting genomic sequence data can be specific to protein-coding genes, corresponding regulatory DNA sequences, and / or diseases or medical conditions of related genes of interest. For imputation, unobserved genotypes (e.g., genotypes that have not yet been directly assayed) are inferred from genotype reference sequences or genotype sequence populations.

[0026] Genome sequence data is the sequence data obtained by any technology that is suitable for providing the sequence data of organism genome.Genomic sequence can include any interested genomic sequence fragment for determining the existence of mutation.Encoding sequence and / or non-coding segment (for example, regulatory sequence, regulatory active intron sequence etc.) can be obtained.Genomic sequence data can mainly include exon sequence or exon sequence.Genomic sequence data can further be limited to specific chromosome, chromosome region, gene cluster, gene family or gene.

[0027] In action 120, the processor selects from the genomic sequence data. The entire sequence can be used, such as the whole genome sequence of the patient or all gene sequences provided from a panel. In action 120, only a portion of the available genomic sequence data is used. For example, any protein coding gene, corresponding regulatory DNA sequence, and / or gene associated with a disease or medical condition is selected.

[0028] The genome sequence data to be used as input to one or more language models are selected. Some of the available genome sequence data are not selected. Different genome sequence data can be selected for different language models (e.g., different selections for regulatory DNA language models and protein language models).

[0029] The selection occurs in the genome sequence data. The selected data can be transcribed into functional RNA and / or translated into protein sequence. For functional RNA or protein, the selection can occur after being transcribed / translated into RNA sequence or protein sequence from DNA.

[0030] Specific genes whose meaning is known in disease etiology can be selected. For example, in the context of cancer, genes annotated in the COSMIC cancer gene census and carrying one or more non-synonymous mutations can be selected. Preferably, for example, in the context of cancer, genes encoding TP53, KRAS, KEAP1, STK11, EGFR, NF1, BRAF or STED2 can be pre-selected and corresponding genomic sequence data obtained. Further specific examples of genes involved in different classes of diseases will be known to the technician or can be derived from a suitable database.

[0031] In an alternative or additional method, the selection of action 120 has the score output from action 135 and / or 145. The output of the language model is selected for the input of the prediction to action 150. Fewer than all scores and corresponding genes, functional RNAs and / or proteins are selected. For example, a threshold score is applied. In action 145, a pre-selection of genomic sequence data about protein coding genes is obtained by applying a minimum threshold to the metric score obtained from the DNA language model.

[0032] Select genomic sequence data or data derived therefrom (e.g., protein sequence) for use by two or more language models. Sequence data (e.g., DNA sequence and / or protein sequence) is encoded (e.g., a combination of four bases ACGT of DNA, four bases ACGU of RNA and twenty amino acids of protein). Coding is considered a language. Language model predicts characteristics from sequence data, such as gene expression and protein function or interaction. In order to consider the variation of expression and functionality of protein coding genes determined together in the genomic sequence, a protein language model for scoring the functionality of each protein and a regulatory DNA language model for scoring the degree to which genes are expressed are used. Other language models for scoring other characteristics may also be used or used instead. Genomic sequence data (or information derived therefrom) is used to score by a language model, and the score can then be used to predict health.

[0033] In action 130, the processor translates the protein coding gene sequence contained in the genome sequence data into the encoded protein sequence. Any translation program can be used, such as based on a database that relates gene sequences to protein sequences. An in silico translation code performed on a computer is provided. The obtained genome sequence data is converted into a protein sequence. The conversion is performed from a nucleotide sequence to a protein's encoded amino acid sequence.

[0034] The genomic and / or protein sequence data may be modified or processed, for example, the sequence data may be converted to a new format, analyzed for features of interest, or combined in any suitable manner.

[0035] In action 135, the processor generates one or more scores for protein function. One or more scores may be generated for each of one or more proteins or protein combinations. In one approach, a score is provided for each of a plurality of proteins or protein-protein interactions.

[0036] The score is a measure of the property(s) of interest. For example, a score is provided within a range. In a protein, the range may be [0, 1], where 1 represents no mutation (e.g., fully functional) and 0 represents the highest impact mutation(s) (e.g., non-functional), with intermediate values ​​representing different levels of mutation or function. Other ranges and / or representations may be used.

[0037] The score is generated by a protein language model implemented by a processor in response to an input derived from a genomic sequence. The translated gene is input as a protein sequence to the protein language model, which generates an output score in response. The protein language model determines a metric score corresponding to the degree to which each protein is functional based on the protein sequence of each protein. By inputting different, specific and / or all protein sequences of the patient, a score of protein function is generated. The protein language model is applied to the protein sequence to obtain a metric score for evaluating the functionality of the corresponding protein.

[0038] Any protein language model may be used. The protein language model may be a machine learning model, such as a transformer neural network previously trained to score based on input of a protein sequence. In one embodiment, the protein language model is an evolutionary scale model (ESM), such as an ESM-b, ESM-v1, or ESM-2 model. Other ESMs or other protein language models may be used.

[0039] Transformer architectures can be trained to operate on molecular sequences such as proteins and DNA sequences. For example, protein language models have been trained on large databases of protein sequences (e.g., UniRef90) to capture the functional effects of protein sequence variation, with or without relying on multiple protein sequence alignments and supervision from experimental measurements of function.

[0040] In action 140, the processor extracts DNA sequences from the genomic sequence data. DNA sequences that exert regulatory effects on the expression of protein-coding genes are extracted. Other sequences may also be extracted or extracted instead. Coding and / or non-coding (e.g., promoter) sequences may be extracted.

[0041] The extraction can be based on the expression of genes of interest. For example, genes known to be associated with overall health, therapy response, prognosis, diagnosis, medical conditions and / or diseases are identified from a database or study. DNA sequences within the range of these genes in the genomic sequence are extracted as having regulatory effects. For example, the score is to be generated using input derived from a genomic sequence within 20, 50, 100 or more kb around the gene, such as around the transcription start site of the gene sequence. DNA regulatory sequences (including distal enhancers, repressors and insulators) that can affect gene expression far greater than 20kb away from the transcription start site, the limitations of algorithms based on convolutional networks for predicting gene expression, are extracted and can be effectively integrated by the increased receptive field of up to 100kb or more transformer layers. A smaller range can be provided.

[0042] The transcription start site is derived from a database or study. In one method, the transcription start site is found from a map, such as the FANTOM5 map. Other maps or resources may also be used.

[0043] In action 145, the processor generates a score for gene expression. One or more scores may be generated for each of one or more genes or gene combinations. In one method, a score is provided for each of a plurality of regulatory DNA sequences.

[0044] The score is a measure of the characteristic (one or more) of interest. For example, a score within a range is provided. In a gene, the score can be in a positive range, where 0 represents no expression, and values ​​greater than 0 represent varying degrees of expression. Other ranges and / or representations can be used.

[0045] The control DNA language model implemented by the processor generates a score based on the input derived from the genome sequence. The control DNA sequence of the extracted gene of interest is input. For example, the input is the DNA sequence of 100kb or other scopes around the transcription start site of each protein coding gene of the genome sequence. A score is generated for each DNA sequence corresponding to each protein coding gene. Other genes or DNA (for example, gene of interest) can also be input or can be replaced by the control DNA sequence. The control DNA language model determines the metric score corresponding to the degree of expression of each gene based on the control sequence of each gene. The control DNA language model is applied to the control sequence, and the metric score of the expression of the corresponding protein coding gene of the evaluation is obtained thus.

[0046] Any regulatory DNA language model can be used. For example, a language model using machine learning, such as a model based on a transformer neural network. The model is trained to generate a metric score in response to the input of the extracted regulatory DNA sequence. In one embodiment, the regulatory DNA language model is an Enformer model. Other models can be used.

[0047] Other language models can be used to generate scores. Patient's genomic sequence data (e.g., extraction) or its derivation (e.g., translation) is input to other language models to generate scores for sequences of interest. For example, scores are generated based on functional RNA coding genes and protein coding genes. Metric scores for functional RNA coding genes such as long non-coding RNA (lncRNA) are incorporated. Other characteristics, DNA, RNA or protein can be used for scoring.

[0048] In action 150, the processor predicts the health of the patient. The health of the patient can be a diagnosis, prognosis, or prediction output about a disease or medical condition. The prediction is provided by a prediction model trained on a diagnosis, prognosis, or prediction conclusion in the context of a disease or medical condition corresponding to genomic sequence data from a cohort of subjects affected by the disease or medical condition.

[0049] Health can be reflected by classification, such as indicating the type, level, or scale of a medical condition or disease. Health can be reflected as a value of a clinical decision scale. A metric score is estimated, where the score represents the level of a disease or medical condition. For example, the scale is a range (e.g., [0,1]), where one extreme of the range maps to one extreme of the prediction, and the other extreme of the range maps to the other extreme of the prediction (e.g., 1 is 100% probability of responding to treatment, and 0 is 0% probability of responding to treatment). Any range or clinical decision scale can be used.

[0050] In another method, health is disease subtype classification (e.g., genotype of cancer), prognostic trend assessment (e.g., 5-year survival rate) and / or treatment response prediction. As an example, a prediction model is trained to predict the toxicity risk after radiotherapy treatment of lung cancer patients based on the genome and outcome data generated from the REQUITE queue. In this data set, whole genome genotyping data has been attributed from Illumina Infinium OncoArray-500kBeadChip data, and radiotherapy toxicity has been evaluated using the Common Terminology Criteria for Adverse Events (CTCAE).

[0051] The disease or condition may be cancer or a genetic defect, such as an inherited genetic disease. Examples of diseases or medical conditions include cancer, such as gastric cancer, colon cancer, rectal cancer, liver cancer, pancreatic cancer, lung cancer, breast cancer, cervical cancer, uterine corpus cancer, ovarian cancer, prostate cancer, testicular cancer, bladder cancer, kidney cancer, brain cancer / CNS cancer, head and neck cancer, laryngeal cancer, Hodgkin lymphoma, non-Hodgkin lymphoma, multiple myeloma, leukemia, melanoma skin cancer, non-melanoma skin cancer, acute lymphocytic leukemia, acute myeloid leukemia, Ewing's sarcoma, small cell lung cancer, choriocarcinoma, rhabdomyosarcoma, Wilms tumor, neuroblastoma, hairy cell leukemia, oral / pharyngeal cancer, esophageal cancer, pharyngeal cancer, kidney cancer, lymphoma, or any subtype thereof; or a nervous system disease, such as Alzheimer's disease, multiple sclerosis, amyotrophic lateral sclerosis (ALS) or ataxia, or a cardiovascular disease, such as coronary heart disease; or a metabolic disease, such as diabetes, metabolic syndrome, iron metabolism disorders, lipid metabolism disorders, calcium metabolism disorders, etc. Further, the disease or condition may be a predisposition to the disease reflected by the presence of a combination of genetic variants or genetic modifications.

[0052] The prediction is responsive to input to the prediction model. The score generated by the language model in actions 135, 145 and / or any other language model score is input. The extracted and / or translated sequence of interest is used to generate a score that is input to the prediction model to predict the health of the patient. In response to the input of the metric score to the prediction model, a diagnosis, prognosis, or prediction output regarding a disease or medical condition is predicted by the prediction model.

[0053] In one approach, scores such as those provided by a language model using only genomic sequence data are used alone to make predictions. In other approaches, additional information from the genomic data is used alone in predictions, such as inputting extracted or translated sequence or genomic information in addition to scores from a language model, without an intermediate language model.

[0054] In yet another approach, additional inputs from sources other than genomic sequence data are used. For example, other inputs to the prediction model for prediction include age, sex, race, characteristics of the disease phenotype (e.g., histological characteristics), the stage of development of the disease, detectable molecular changes (e.g., at the level of the transcriptome, proteome, metabolome, or lipidome), and / or biomarkers (e.g., derived from medical images, pathological images, and / or laboratory tests). Different or additional inputs may be used.

[0055] The prediction model is a machine learning model. The metric score is used as an input to a prediction model (e.g., a neural network) that has been machine trained in the context of a disease or medical condition for a diagnosis, prognosis, or prediction conclusion. Any machine learning model can be used. For example, a support vector machine, clustering, a neural network, or another model is provided for generating an estimate of health based on the input.

[0056] Various architectures can be used. The architecture defines the arrangement of layers, units or other machine learning components. For a neural network, the architecture can include one or more convolutional layers, where deep learning trains (one or more) kernels for filtering (convolution). Fully connected layers, maximum pooling layers, batch normalization and / or other layers can be included. The network is a deep architecture, which can include a convolutional neural network (CNN) or a deep belief network (DBN). Other deep networks can be used.

[0057] A neural network is defined as a plurality of serialized feature units or layers. Within a unit or layer, any number of nodes is provided. The nodes of each layer or unit may be connected to all nodes or only a subset of nodes of the previous layer or subsequent layer or unit. The features of the nodes are learned by the machine using any building blocks. For example, an autoencoder (AE) or restricted Boltzmann machine (RBM) method is used. In other embodiments, at least one unit is a convolution with ReLU activation, or a batch normalization with ReLU activation, followed by a convolution layer (BN+LeakyRU+convolution). Max pooling, upsampling, downsampling and / or softmax layers or units may be used. Different units may be of the same or different types.

[0058] In training, the values ​​of learnable parameters (features) are learned. Optimization is performed using many samples of paired inputs and outputs using an objective criterion or ground truth (e.g., unsupervised or supervised). Any optimization, such as ADAM, and corresponding losses, such as L1 norm, L2 norm, or cross entropy, can be used.

[0059] In one embodiment, the prediction model of machine learning is a graph neural network. The architecture of the neural network is formed as a graph that at least partially represents protein-protein interactions. The protein-protein interaction network incorporates prior knowledge about protein-protein interactions. This prior knowledge is incorporated into the prediction model in the form of a graph neural network to consider proteins in the context of biological complexes, functions and / or pathways. For example, the edges between the nodes of different proteins are weights representing the likelihood of the physical association of the connected protein pairs. The weights correspond to a continuous measure of the physical association of each connected protein pair. In the case where the relationship is known, the weights can be fixed. Alternatively, the weights are initialized using a known but then learnable or learnable relationship. The existence of the edge and / or the value of the weight are incorporated into the prior knowledge.

[0060] Figure 2 An example graph neural network and further network architecture for making predictions from scores 200, 202 are illustrated. The scores 200, 202 from the language model are input to a graph neural network 210. The graph neural network 210 represents protein-protein interactions, where each node G corresponds to a gene / encoded protein and is assigned a tuple corresponding to the score 200, 202. Each undirected edge corresponds to an interaction between two proteins, where the arrows illustrate the messages passed between proteins G1-5 when the graph convolutional layers are applied. The readout of the graph neural network 210 produces a graph embedding 220. Additional information 224 (e.g., age, race, or gender) can be input as an encoding and concatenated with the graph embedding 220. The resulting layer 222 is provided to a prediction head, such as a multilayer perceptron 230, to produce a predicted label. Other arrangements with additional, different, or fewer layers may be used.

[0061] Graph neural networks are based on neighborhood aggregation or message passing schemes, where the representation vector of a node is computed by recursively aggregating and transforming the representation vectors of its neighboring nodes, i.e., each node aggregates the feature vectors of its neighbors to compute its new feature vector. In general, after a certain number of iterations of aggregation, the node is represented by its transformed feature vector, which captures information within the node's neighborhood. Then, the representation of the entire graph can be obtained, for example, by pooling, for example, by summing the representation vectors of all nodes in the graph.

[0062] Therefore, graph neural networks should be understood accordingly as optimizable transformations on all properties of a graph, e.g., nodes, edges, and global context that preserve graph symmetries. Typically, graph neural networks provide a model of a protein-protein interaction network in the form of a graph.

[0063] An adjacency matrix, which is a square matrix, is used to represent a finite graph. The elements of the matrix indicate whether a pair of nodes is adjacent in the graph. Additionally, or alternatively, an adjacency list may be generated, wherein the list describes the connectivity of the edges between nodes as tuples.

[0064] The interactions or relationships that form the graph neural network can be determined from databases or studies. Examples of databases include STRING and IntAct. Examples of studies include the Integrated Association Strictness (IAS) score network or the OncoPPi network.

[0065] In one embodiment, the graph neural network is a graph convolutional network, a graph isomorphic network, or a graph attention network. Graph convolutional networks share filter parameters at all locations in the graph. In another embodiment, the graph neural network is a graph isomorphic network. Graph isomorphic networks implement aggregation schemes for general functions on multiple sets representing nodes and their neighbors. Graph isomorphic networks typically train separate multi-layer perceptrons on each component of the graph, resulting in graph isomorphic network layers. Therefore, for each node vector, a multi-layer perceptron is applied and a node feature vector is generated. In yet another embodiment, the graph neural network is a graph attention network. Graph attention networks operate on graph-structured data. The attention mechanism processes inputs of variable size, focusing on the most relevant parts of the input to make decisions. The graph attention network computes the hidden representation of each node in the graph by attending to its neighbors following a self-attention strategy.

[0066] The trained graph neural network is or has been trained based on genomic sequence data from a cohort of subjects affected by a disease or medical condition. Thus, the genomic data is associated with a specific cohort or group of people known to be affected by a disease or medical condition. The group may be affected by the same disease or condition, or by a family or group of similar diseases or conditions. Additionally, genomic data of one or more healthy subjects may be additionally input and used for reference and comparison purposes.

[0067] The disease or medical condition is, for example, reflected by the genome sequence of the subject in the form of a detectable mutation. Such a mutation can be an insertion, deletion or substitution of one or more nucleotides. In a further embodiment, such a mutation can cause a change in the primary amino acid sequence of a protein. After training, the prediction model of machine learning responds to the input of the score from the language model with or without other inputs, and outputs the health of the patient about the disease or medical condition. Therefore, the trained prediction model generates the output of the input score from the genome sequence data in the context of the learning task, thereby providing the diagnosis, prognosis and / or prediction conclusions about the disease or medical condition initially associated with the training data of the ill subject cohort in the context of the genome sequence data of the target subject. In other words, the prediction model of machine learning generates the output linking the diagnosis, prognosis and / or therapy instructions to the genome sequence data of the patient based on the language model score.

[0068] exist Figure 1 In action 160, the processor determines the uncertainty of health as predicted by the machine learning model. The machine learning model can output health (e.g., a metric score for a clinical decision) as a probability. The probability reflects uncertainty. As another example, similar patients are identified. One or more patients with the closest combination of metric scores from a language model and / or a genome or patient characteristics can be identified. The results for those patients can be analyzed for variance to indicate the uncertainty in the predictions made by the machine learning prediction model. In another example, the output prediction is mapped to a specific class (e.g., yes or no for 5-year survival rate). A majority vote of different prediction models based on different genes and / or proteins is used to determine uncertainty. Using bootstrapping in training, different prediction models can be trained from the same training data. The outputs of different prediction models can be used to determine uncertainty.

[0069] In one method, uncertainty is a statistical factor that reflects the accuracy of prediction by considering the similarity between the current molecular situation of the target subject and the molecular situation of the subject queue that produces the training data set. The statistical factor can be derived, for example, from the measurement of the similarity between the potential representation of the genomic data of the target subject and the potential representation of the genomic data of the patient queue. In a further embodiment, the statistical factor can be derived from the measurement of patient similarity, for example, about age, gender, race, health records, etc. The measurement of similarity can be performed, for example, using a taxi or Manhattan geometry, where the distance between two points is measured as the sum of the absolute differences of Cartesian coordinates. For example, in the case where the genomic sequence data of the target subject produces a mutation pattern-the mutation pattern is reflected by the mutation data that can be derived from the subject of the training queue, and more preferably by the data of the subject that has been successfully treated-the likelihood of the positive therapy response of the target subject is very high. Therefore, uncertainty will be very low.

[0070] In action 170, the display shows health as predicted by the machine learning model. The processor generates an image, which is displayed on the display. The image indicates a diagnosis, prognosis, or prediction output regarding a disease and / or medical condition. The image can be output to a display, a patient medical record, and / or a report.

[0071] Health can be an annotation, alphanumeric text, graph, or chart. For example, survival is displayed as a graph of probability of survival as a function of time. As another example, stratified results are displayed. Prediction scores are displayed for a physician. For example, the output can be provided in a report or as an alert.

[0072] Other information can be displayed along with the predicted health, such as input information (e.g., genomic sequence data) that is strongly linked to the prediction and / or uncertainty. More than one prediction can be output, such as (1) predictions of treatment outcomes for different therapies, or (2) diagnosis and prognosis.

[0073] The output health can be presented with or on the image of the patient (e.g., overlaid or annotated). For example, a slice image through a tumor region is displayed. The output health is provided on a display screen with an image of an anatomical structure of interest (such as a lung or total tumor volume).

[0074] The display may include information from one or more similar patients.The language model scores and / or other information may be used to find other or past patients with similar score arrangements.

[0075] In one embodiment, a metric score on a clinical decision scale output by a prediction model is output. The metric score is a ranking along a grading scheme. The grading scheme is provided as a metric classification embedded in a clinical or therapy decision scale. For example, the severity of the disease can be classified on a scale from 1 to 10, for example, 1 is very slightly severe, and 10 is strongly severe. The diagnosis can then lead to a corresponding therapy decision. Alternatively, the output can be a value of the likelihood of, for example, 6 months survival (for example, after a mutation causing cancer is detected). In a further example, the output can be a time prediction until the disease relapses. Similarly, the prediction output can be provided as a metric score, for example, a specific dose can be provided in a metric scale from 1 to 10 times the normalized dose 1, or radiotherapy can be suggested based on a normalized standard intensity 5 with a metric grading from intensity 1 to 10.

[0076] Based on the displayed health, the doctor assists the patient. For example, a diagnosis and / or prognosis can be confirmed and transmitted. As another example, the possible results of one or more treatments can be analyzed and a treatment can be selected for the patient. The doctor uses information about the patient and the output health to select a treatment or not. Stratification, survival and / or other results are used to select a treatment for the patient.

[0077] Figure 3 A medical decision support system is shown. The system generates an estimate of a patient's health. Genomic sequence data and language models 302, 304 are used to predict the patient's health by a prediction model 306.

[0078] The medical decision support system includes a display 320, a memory 300, and a processor 310. A sequencer 330 may be included to obtain genome sequence data of a patient. The display 320, the processor 310, and the memory 300 may be part of a computer, a server, a workstation, or other system for medical diagnosis, prognosis, and / or treatment of a patient.

[0079] Additional, different or fewer components may be provided. For example, a computer network is included for remote prediction based on locally captured sequencing data. As another example, a user input device (e.g., a keyboard, button, slider, dial, trackball, mouse or other device) is provided for user interaction with prediction. In yet another example, a sequencer 330 is not provided, such as in which the memory 300 stores genomic sequence data or data derived therefrom (e.g., extracted regulatory sequences and translated protein sequences).

[0080] The memory 300 is configured by formatting and / or by the processor 310 to store information of the patient's genome sequence. The memory 300 also or alternatively stores a protein language model 302, a DNA language model 304 and / or a prediction model 306 (e.g., a neural network for the prediction of a medical condition or disease). In one embodiment, the protein language model 302 is used for the functionality of the encoded protein, and the DNA language model 304 is used for the expression of the gene. For example, the DNA language model is a transformer as a machine learning of a regulatory DNA language model. The regulatory DNA language model operates on 20kb, 50kb, 100kb or more DNA around the transcription start site in the genome sequence of a specific gene of interest for a medical condition or disease. As another example, the protein language model 302 is a transformer as a machine learning of a model for determining protein functionality based on a protein sequence.

[0081] The prediction model 306 as stored in the memory 300 is a machine learning model, such as a neural network, a graph neural network, or another architecture or type, for predicting health at least in part from the scores output by the language models 302, 304, ... As a graph neural network, at least a portion of the network incorporates prior knowledge based on protein-protein interactions.

[0082] The memory 300 is an external storage device, RAM, ROM, database and / or local memory (e.g., a solid state drive or hard drive). The same or different non-temporary computer readable media can be used for instructions and other data. The memory 300 can be implemented using a database management system (DBMS) and resides on a memory such as a hard drive, RAM or removable media. Alternatively, the memory 300 is inside the processor 310 (e.g., a cache). The memory 300 is formed by a device or a collection of devices, such as different memories storing different types of data.

[0083] Instructions for implementing the training or application processes (i.e., predictions for patients), methods, and / or techniques discussed herein by the processor 310 are provided on a non-transitory computer-readable storage medium or memory, such as a cache, buffer, RAM, removable media, hard drive, or other computer-readable storage medium (e.g., memory 300). Computer-readable storage media include various types of volatile and non-volatile storage media. In response to one or more instruction sets stored in or on a computer-readable storage medium, the functions, actions, or tasks illustrated in the figures or described herein are performed. The functions, actions, or tasks are independent of a particular type of instruction set, storage medium, processor, or processing strategy, and can be operated by software, hardware, integrated circuits, firmware, microcode, and the like, alone or in combination.

[0084] In one embodiment, the instructions are stored on a removable media device for reading by a local or remote system. In other embodiments, the instructions are stored in a remote location for transmission over a computer network. In yet another embodiment, the instructions are stored in a given computer, CPU, GPU, or system. Because some of the constituent system components and method steps depicted in the accompanying drawings can be implemented in software, the actual connection between the system components (or process steps) can be different depending on the embodiment in the manner in which it is programmed.

[0085] Processor 310 is a control processor, a general purpose processor, a digital signal processor, a graphics processing unit, an application specific integrated circuit, a field programmable gate array, an artificial intelligence processor or accelerator, a digital circuit, an analog circuit, and a combination thereof, or other devices now known or later developed for processing language models, genomic sequence data, and machine learning prediction models. Processor 310 is a single device, multiple devices, or a network. For more than one device, parallel or sequential processing partitioning can be used. Different devices that make up processor 310 can perform different functions. Processor 310 operates according to stored instructions, hardware, and / or firmware to perform various actions described herein.

[0086] In one embodiment, processor 310 is configured to train one or more machine learning networks, such as transformers forming language models 302, 304 and / or neural networks forming prediction model 306. Based on the network architecture and training data provided by a user or other source, processor 310 learns features or values ​​of learnable parameters to train the network.

[0087] Alternatively or additionally, the processor 310 is configured to apply one or more machine learning models. For example, the processor 310 is configured to translate the genome sequence data into a protein sequence, and the protein language model 302 is applied to at least some of the translated information. As another example, the processor 310 is configured to extract DNA or gene sequences (e.g., regulatory DNA) from the genome sequence data, and the DNA language model 304 is applied to at least some of the extracted information. In yet another example, the processor 310 is configured to apply a prediction model 306 (e.g., a neural network) to the output of the language models 302, 304. The prediction model 306 is configured by previous training to output the decision support information (e.g., health) of the patient from the application of the prediction model 306. For example, the health of the patient (e.g., diagnosis, prognosis and / or therapy response) is output.

[0088] The processor 310 is configured to generate an image. An image showing the predicted information is generated. The information about the disease or medical condition may be displayed together with an image of the inside of the patient, such as a computed tomography image, or displayed separately.

[0089] The predicted information is displayed for decision support. The image shows decision support information or information derived from decision support information (e.g., a recommended treatment given a prediction of success of different treatments output by the prediction model 306, including an explanation of the results, e.g., providing feature attributes).

[0090] The display 320 is a CRT, LCD, projector, plasma display, printer, tablet computer, smart phone or other display device now known or later developed for displaying decision support information or information derived from decision support information. The display 320 is configured by loading an image into a display buffer or plane, which is then displayed on a screen. Other configurations may be provided, such as being configured for display by printing. The display 320 is configured to display decision support information or information derived from decision support information.

[0091] Various illustrative embodiments are listed below.

[0092] Illustrative embodiments 1. A method for decision support in a medical system, the method comprising: obtaining a patient's genomic sequence; a protein language model implemented by a processor generates a first score of protein function based on an input derived from the genomic sequence; a regulatory DNA language model implemented by the processor generates a second score of gene expression based on an input derived from the genomic sequence; in response to input of the first score and the second score to a machine learning model, predicting the patient's health by the machine learning model; and displaying the health as predicted by the machine learning model.

[0093] Illustrative embodiment 2. The method of illustrative embodiment 1, wherein predicting comprises predicting health as a value in a clinical decision scale for a disease or medical condition.

[0094] Illustrative embodiment 3. The method of any of illustrative embodiments 1-2, further comprising determining uncertainty in health as predicted by the machine learning model.

[0095] Illustrative embodiment 4. A method according to any of illustrative embodiments 1-3, wherein generating a first score includes generating a first score by a protein language model implemented by a processor including an evolutionary scale model (ESM), and wherein generating a second score includes generating a second score by a regulatory DNA language model implemented by a processor including an Enformer model.

[0096] Illustrative embodiment 5. A method according to any of illustrative embodiments 1-4, wherein generating the second score comprises generating the second score using an input derived from a genomic sequence, wherein the genomic sequence is 100 kb or more surrounding a transcription start site of a protein-coding gene of the genomic sequence.

[0097] Illustrative Example 6. The method according to Illustrative Example 5, further comprising deriving the transcription start site from the map.

[0098] Illustrative embodiment 7. The method according to any one of illustrative embodiments 5-6, wherein generating the second score further comprises generating based on functional RNA encoding genes and protein encoding genes.

[0099] Illustrative embodiment 8. A method according to any of illustrative embodiments 1-7, wherein predicting includes making predictions when the machine learning model includes a neural network.

[0100] Illustrative embodiment 9. A method according to illustrative embodiment 8, wherein the neural network comprises a graph neural network representing a protein-protein interaction network.

[0101] Illustrative embodiment 10. A method according to illustrative embodiment 9, wherein the edges of the graph neural network include weights representing the likelihood of physical association of the connected protein pairs.

[0102] Illustrative embodiment 11. A method according to any of illustrative embodiments 1-10, wherein predicting includes making a prediction in response to inputs, wherein the inputs further include age, gender, race, histological characteristics, stage of disease development, detectable molecular alterations, biomarkers derived from medical images, pathological images and / or laboratory tests.

[0103] Illustrative embodiment 12. The method according to any of illustrative embodiments 1-11, further comprising selecting from the genomic sequence with respect to protein-coding genes, corresponding regulatory DNA sequences, and / or diseases or medical conditions.

[0104] Illustrative embodiment 13. A method according to illustrative embodiment 12, wherein the input derived from the genomic sequence that regulates the DNA language model and / or the protein language model is derived from the selection.

[0105] Illustrative embodiment 14. The method of any of illustrative embodiments 12-13, wherein selecting comprises selecting based on a threshold of the first or second score.

[0106] Illustrative embodiment 15. The method of any of illustrative embodiments 1-14, wherein obtaining comprises obtaining from gene panel sequencing, whole exome sequencing, whole genome sequencing, whole genome genotyping microarray, and / or imputation of unobserved genotypes.

[0107] Illustrative embodiment 16. A medical decision support system comprising: a memory configured to store information of a patient's genomic sequence, a first language model for expression of genes, a second language model for functionality of encoded proteins, and a neural network for prediction of a medical condition or disease; a processor configured to apply the first language model to at least some of the information, apply the second language model to at least some of the information, and apply the neural network to outputs of the first and second language models, the neural network configured to output decision support information for the patient from application of the neural network; and a display configured to display the decision support information or information derived from the decision support information.

[0108] Illustrative embodiment 17. A medical decision support system according to illustrative embodiment 16, wherein the first language model includes a regulatory DNA language model, wherein at least some of the information to which the first language model is applied includes 100 kb or more surrounding a transcription start site in a genomic sequence.

[0109] Illustrative Embodiment 18. The medical decision support system of any of Illustrative Embodiments 16-17, wherein the neural network comprises a graph neural network arranged based on protein-protein interactions.

[0110] Illustrative embodiment 19. A method for analyzing genomic sequence data, the method comprising: obtaining genomic sequence data of a patient; translating a protein-coding gene sequence contained in the genomic sequence data into an encoded protein sequence; applying a protein language model to the encoded protein sequence, the application resulting in a first metric score for evaluating the functionality of the corresponding protein; extracting a DNA sequence that exerts a regulatory effect on the expression of the protein-coding gene from the genomic sequence data; applying a regulatory DNA language model to the extracted DNA sequence, the application resulting in a second metric score for evaluating the expression of the corresponding protein-coding gene; deriving a diagnosis, prognosis or prediction output regarding a disease or medical condition from the machine-trained prediction model in response to input of the first and second metric scores to a machine-trained prediction model; and displaying the diagnosis, prognosis or prediction output.

[0111] Illustrative embodiment 20. A method according to illustrative embodiment 19, wherein the deriving includes deriving by a machine-trained predictive model including a graph neural network, wherein the graph represents protein-protein interactions.

[0112] The various improvements described herein may be used together or separately. Although illustrative embodiments of the present invention have been described herein with reference to the accompanying drawings, it should be understood that the invention is not limited to those precise embodiments and that various other changes and modifications therein may be effected by those skilled in the art without departing from the scope or spirit of the invention.

Claims

1. A method for decision support in a medical system, the method comprising: Obtaining the patient's genome sequence; A protein language model implemented by a processor generates a first score of protein function based on an input derived from a genomic sequence; The regulatory DNA language model implemented by the processor generates a second score for gene expression based on input derived from the genomic sequence; In response to inputting the first score and the second score into the machine learning model, predicting the health of the patient by the machine learning model; as well as Displays health as predicted by the machine learning model.

2. The method of claim 1, wherein predicting comprises predicting health as a value in a clinical decision scale for a disease or medical condition.

3. The method of claim 1 , further comprising determining uncertainty in health as predicted by the machine learning model.

4. The method of claim 1 , wherein generating a first score comprises generating a first score by a protein language model implemented by a processor including an evolutionary scaling model (ESM), and wherein generating a second score comprises generating a second score by a regulatory DNA language model implemented by a processor including an Enformer model.

5. The method of claim 1, wherein generating the second score comprises generating the second score using an input derived from a genomic sequence, the genomic sequence being 100 kb or more surrounding a transcription start site of a protein-coding gene of the genomic sequence.

6. The method of claim 5, further comprising deriving the transcription start site from the map.

7. The method of claim 5, wherein generating the second score further comprises generating based on functional RNA encoding genes and protein encoding genes.

8. The method of claim 1, wherein predicting comprises predicting where the machine learning model includes a neural network.

9. The method of claim 8, wherein the neural network comprises a graph neural network representing a protein-protein interaction network.

10. The method of claim 9, wherein the edges of the graph neural network include weights representing the likelihood of a physical association of the connected protein pairs.

11. The method of claim 1 , wherein predicting comprises predicting in response to inputs, wherein the inputs further comprise age, sex, race, histological characteristics, stage of disease progression, detectable molecular alterations, biomarkers derived from medical images, pathological images, and / or laboratory tests.

12. The method of claim 1, further comprising selecting from the genomic sequence with respect to protein coding genes, corresponding regulatory DNA sequences, and / or diseases or medical conditions.

13. The method of claim 12, wherein input derived from a genomic sequence that regulates a DNA language model and / or a protein language model is derived from a selection. The method of claim 12 , wherein selecting comprises selecting based on a threshold of the first or second score.

15. The method of claim 1, wherein obtaining comprises obtaining from gene panel sequencing, whole exome sequencing, whole genome sequencing, whole genome genotyping microarray, and / or imputation of unobserved genotypes.

16. A medical decision support system comprising: A memory configured to store information of a patient's genomic sequence, a first language model for predicting expression of a gene, a second language model for predicting functionality of an encoded protein, and a neural network for prediction of a medical condition or disease; a processor configured to apply a first language model to at least some of the information, to apply a second language model to at least some of the information, and to apply a neural network to outputs of the first and second language models, the neural network configured to output decision support information for the patient from application of the neural network; as well as A display is configured to display the decision support information or information derived from the decision support information.

17. The medical decision support system of claim 16, wherein the first language model comprises a regulatory DNA language model, wherein at least some of the information to which the first language model is applied comprises 100 kb or more surrounding a transcription start site in a genomic sequence.

18. The medical decision support system of claim 16, wherein the neural network comprises a graph neural network arranged based on protein-protein interactions.

19. A method for analyzing genomic sequence data, the method comprising: Obtaining the patient's genomic sequence data; translating protein-coding gene sequences contained in the genomic sequence data into encoded protein sequences; applying a protein language model to the encoded protein sequence, said application resulting in a first metric score assessing functionality of the corresponding protein; Extracting DNA sequences that exert regulatory effects on the expression of protein-coding genes from genomic sequence data; applying a regulatory DNA language model to the extracted DNA sequences, said applying resulting in a second metric score assessing expression of corresponding protein-coding genes; deriving, by the machine-trained predictive model, a diagnostic, prognostic, or predictive output regarding a disease or medical condition in response to inputting the first and second metric scores to the machine-trained predictive model; and Display diagnostic, prognostic, or predictive output.

20. The method of claim 19, wherein deriving comprises deriving by a machine trained predictive model comprising a graph neural network, wherein the graph represents protein-protein interactions.

Images