Systems and methods for diagnosing, prognosticating, selecting, or treating a subject based on embedding in a generative neural model feature space

Abstract

Disclosed herein are methods of determining a therapeutic for a patient with a psychiatric or neurological disorder, based upon the patient's functional brain imaging data through use of a generative model that predicts brain activity.

Description

SYSTEMS AND METHODS FOR DIAGNOSING, PROGNOSTICATING, SELECTING, OR TREATING A SUBJECT BASED ON EMBEDDING IN A GENERATIVE NEURAL MODEL FEATURE SPACECROSS REFERENCE

[0001] This application claims the benefit of U.S. Provisional Application 63 / 565,397, filed on March 14, 2024, which is incorporated by reference herein in its entirety.FIELD OF THE INVENTION

[0002] The present invention relates to methods of determining a therapeutic agent for a patient with a psychiatric or neurological disorder, based upon the patient’s functional brain imaging data (e.g., functional MRI data) through use of a generative model that predicts brain activity. In particular, the present invention relates to the development, training, and application of a set of algorithms that allow for quantitative representation of an individual in the generative model and classification from that representation to predict the individual’s response to a therapeutic.BACKGROUND

[0003] Many existing therapeutic agents for psychiatric and neurological disorders have limited efficacy, in part because of heterogeneity of patients’ responses as well as a complete absence of a response for some patients with the same diagnosis. That is, a given therapeutic agent may be effective in one patient but not in another, both with the same psychiatric or neurological diagnosis in part due to different disease mechanisms. Furthermore, this issue limits the discovery of novel therapeutic agents because there is currently no neuroimaging informed method to enrich CNS clinical trials with patients that may be maximally responsive. This heterogeneity impacts the process of determining a therapeutic agent for a patient and can result in selecting ineffective treatments for a patient. In turn, this heterogeneity hinders clinical trials for psychiatric and neurological disorders, because heterogeneity can degrade efficacy at thelevel of the clinical group and therefore reduce effect size. Understanding the neural underpinnings of clinical and therapeutic heterogeneity could improve therapeutic agent efficacy as well as probability of clinical trial success by optimizing the selection of patients that are most likely to respond. In addition, such an understanding would aid in the selection of the therapeutic agent most likely to exhibit efficacy for a given person.

[0004] In particular, precision medicine for psychiatric and neurological disorders can be advanced through methods which use an individual patient’s functional brain imaging data to predict the patient’s response to a therapeutic agent.SUMMARY

[0005] There is a need for novel methods and strategies for prediction of a patient’s response to a therapeutic agent and application of such predictive methods to decision making in the development and use of therapeutic agents. In particular, disclosed herein are novel methods which use generative models of an individual patient’s brain activity for detecting, diagnosing, prognosticating, or treating a subject based on embedding in a generative neural model feature space. In certain embodiments, the methods disclosed herein may be used to facilitate the identification of therapeutic agents to which a patient is likely to respond, the selection of patients for inclusion in clinical trials for therapeutic agents, and the selection of clinical groups which are likely to respond to a therapeutic agent. The present invention is directed toward further solutions to address this need, in addition to having other desirable characteristics.

[0006] To meet the needs set forth above, three methods are disclosed.

[0007] A first method is disclosed for identifying a therapeutic agent for treating a patient with a psychiatric or neurological disorder. The method comprises the following steps: (a) receiving, from a user interface of a computing device, phenotypic data of a patient comprising or corresponding to at least one of a biological process, a psychiatric symptom, a neurological symptom, and / or a cognitive status of the patient; (b) processing, by one or more processor of thecomputing device, the phenotypic data of the patient into phenotypic features; (c) receiving, by the one or more processor of the computing device, neural data of the patient comprising or corresponding to functional imaging of brain activity of the patient; (d) processing, by the one or more processor of the computing device, the neural data of the patient into a parcellated time series, wherein the parcellated time series comprises numerical values of brain activity of a set of brain regions at multiple time points, and wherein the brain regions are represented as two- dimensional cortical surfaces and subcortical volumetric structures; (e) retrieving, by the one or more processor of the computing device, a generative model for parcellated time series of brain activity, wherein the generative model represents neural data and phenotypic data for a plurality of subjects as a set of points in a multi-dimensional latent space with each point corresponding to one subject, and wherein the latent space is trained to separate representation of differences across the plurality of subjects from representation of differences across time of brain states within each subject; (f) generating, by the one or more processor of the computing device, a set of values for the coordinates of the point in the latent space of the generative model to which the patient’s processed neural data and processed phenotypic data are mapped; (g) retrieving, by the one or more processor of the computing device, one or more multivariate classifiers, each having inputs and outputs and each associated with a therapeutic agent, wherein the inputs to each multivariate classifier are the values of coordinates in the latent space of the generative model and the outputs to each multivariate classifier are one or more features that describe the predicted response to its associated therapeutic agent; (h) generating, by the one or more processor of the computing device, the predicted response of the patient to the therapeutic agent based on the outputs of the one or more multivariate classifiers associated with that therapeutic agent, wherein the outputs from each multivariate classifier are computed by inputting into each multivariate classifier the patient’s coordinates in the latent space of the generative model; (i) determining, by the one or more processor of the computing device, a composite score as a scalar measure of the patient’s predicted response from the one or more multivariate classifier outputs; (j) selecting a therapeutic agent based on the composite scores; and (k) treating the patient with the selected therapeutic agent.

[0008] Also provided is a method of selecting a candidate patient based on their predicted response to a therapeutic agent. The method is carried out by: (a) receiving, from a user interface of a computing device, phenotypic data of the patient comprising or corresponding to at least one of a biological process, a psychiatric symptom, a neurological symptom, and / or a cognitive status of the patient; (b) processing, by one or more processor of the computing device, the phenotypic data of the patient into phenotypic features; (c) receiving, by the one or more processor of the computing device, neural data of the patient comprising or corresponding to functional imaging of brain activity of the patient; (d) processing, by the one or more processor of the computing device, the neural data of the patient into a parcellated time series, wherein the parcellated time series comprises numerical values of brain activity at multiple time points assigned to a set of brain regions, and wherein the brain regions are represented as two- dimensional cortical surfaces and subcortical volumetric structures; (e) retrieving, by the one or more processor of the computing device, a generative model for parcellated time series of brain activity, wherein the generative model represents neural data and phenotypic data for a plurality of subjects as a set of points in a multi-dimensional latent space with each point corresponding to one subject, and wherein the latent space is trained to separate representation of differences across the plurality of subjects from representation of differences across time of brain states within each subject; (f) generating, by the one or more processor of the computing device, a set of values for the coordinates of the point in the latent space of the generative model to which the patient’s processed neural data and processed behavioral data are mapped; (g) retrieving, by the one or more processor of the computing device, one or more multivariate classifiers, each having inputs and outputs and each associated with a therapeutic agent, wherein the inputs to each multivariate classifier are the values of coordinates in the latent space of the generative model and the outputs to each multivariate classifier are one or more features that describe the predicted response to its associated therapeutic agent; (h) generating, by the one or more processor of the computing device, the predicted response of the patient to the therapeutic agent based on the outputs of the multivariate classifiers associated with the therapeutic agent, wherein the multivariate classifier outputs are computed by inputting into a multivariate classifier the patient’s coordinates in the latent space of the generative model; (i) determining, by the one or more processor of the computing device, a composite score as a scalar measure of the subject’spredicted response from the multivariate classifier outputs; and (j) selecting the patient predicted to respond to the therapeutic agent based on the composite score.

[0009] Further disclosed is a method of selecting a psychiatric or neurological indication with the associated symptom profile based on a predicted response of patients with that indication to a therapeutic agent. The method includes the steps of: (a) receiving, from a user interface of a computing device, phenotypic data of one or more groups of patients, comprising or corresponding to at least one of a psychiatric symptom, a neurological symptom, and / or a cognitive status of the patients, wherein each group of patients is representative of a psychiatric or neurological indication; (b) processing, by one or more processor of the computing device, the phenotypic data of each patient, into phenotypic features; (c) receiving, by the one or more processor of the computing device, neural data of the one or more groups of patients, comprising or corresponding to functional imaging of brain activity of the patients; (d) processing, by the one or more processor of the computing device, the neural data of each patient into a parcellated time series, wherein the parcellated time series comprises numerical values of brain activity at multiple time points assigned to a set of brain regions, and wherein the brain regions are represented as two-dimensional cortical surfaces and subcortical volumetric structures; (e) retrieving, by the one or more processor of the computing device, a generative model for parcellated time series of brain activity, wherein the generative model represents neural data and phenotypic data for a plurality of subjects as a set of points in a multi-dimensional latent space with each point corresponding to one subject, wherein the latent space is trained to separate representation of differences across the plurality of subjects from representation of differences across time of brain states within each subject; (f) generating, by the one or more processor of the computing device, a set of values for the coordinates of the point in the latent space of the generative model to which each patient’s processed neural data and processed phenotypic data are mapped; (g) retrieving, by the one or more processor of the computing device, a neural modulation map, corresponding to the therapeutic agent, wherein the map comprises numerical values assigned to a set of brain regions, and wherein the brain regions are represented as two- dimensional cortical surfaces and subcortical volumetric structures; (h) simulating, by the one or more processor of the computing device, artificial neural data for each patient in each patientgroup, using the generative model with modulation of simulated neural activity in each brain region, wherein the strength of modulation for each brain region is determined by the scalar value of the neural modulation map; (i) generating, by the one or more processor of the computing device, a set of values for the coordinates of the point in the latent space of the generative model to which each patient’s simulated neural data is mapped, for each patient group; (j) retrieving, by the one or more processor of the computing device, a set of latent space coordinates of a plurality of healthy subjects; (k) generating, by the one or more processor of the computing device, a numerical similarity score quantifying the similarity between the set of latent space coordinates of healthy subjects in the generative model and the simulated neural data patient group, for each patient group; and (1) selecting a psychiatric or neurological indication based on the patient groups’ similarity scores. In certain embodiments, the neural modulation map is derived from a positron emission tomography (PET) map. In certain embodiments, the neural modulation map is derived from a gene expression map. In other embodiments, the neural modulation map is derived from functional brain imaging of a baseline or placebo condition to a therapeutically modified condition.

[0010] The above discussed, and many other features and attendant advantages of the present inventions will become better understood by reference to the following detailed description of the invention.BRIEF DESCRIPTION OF THE FIGURES

[0011] These and other characteristics of the present invention will be more fully understood by reference to the following detailed description in conjunction with the attached drawings, in which:

[0012] FIG. 1 represents a diagram of the constituent parts and steps of one embodiment of the generative model to simulate brain activity.

[0013] FIG. 2 represents a diagram of the fractionation of the latent space of the generative model. This improves the coordinates of one subspace (Person) to be predictive of a subject’s response to a therapeutic agent.

[0014] FIG. 3 represents a diagram of the multivariate classifier approach to determining scores for evaluating predicted responses to therapeutic agents for a subject.

[0015] FIG. 4 represents a diagram of the method of simulating neural activity during modulation by a therapeutic agent by use of a topographic neural modulation map.DETAILED DESCRIPTION

[0016] An illustrative embodiment of the present invention relates to the training and application of a generative model of an individual patient’s brain activity to predict the patient’s response to a therapeutic agent. Here, a generative model refers to a computation model which is trained through machine learning methods to predict multivariate time series data as outputs from multivariate time series data as inputs, such as predicting future or missing time series data. The generative model is first trained using one or more reference datasets, each comprising neural data and phenotypic data for a plurality of subjects. The generative model is trained to predict a subject’s brain activity pattern at a time point, based on the subject’s neural data and, in some embodiments, the subject’s phenotypic data. The generative model can simulate brainactivity time series data that reflects individual differences across subjects. The generative model represents the patient’s neural data and phenotypic data as a single point in a multi-dimensional latent space. A “latent space” refers to a multi-dimensional space used as an intermediate representation in the processing of input data to predict output data. The coordinates of the latent space are inputs to multivariate classifiers trained to predict the patient’s response to a therapeutic agent. A score derived from the classifier outputs can be used to select a therapeutic agent for the patient, or to select the patient for inclusion or exclusion in a clinical trial associated with the therapeutic agent.

[0017] FIG. 1 represents a diagram of the constituent parts and steps of one embodiment of the generative model to simulate brain activity. Input data are neural spatio-temporal time series and phenotypic, in this example specifically behavioral, measures (FIG. 1, label 1). The neural spatio-temporal time series data are structured as a matrix organized by brain regions and by time points, with the numerical value of each element representing the activity level of the brain region at the time point. The behavioral data are structured as a vector organized by behavioral features, with the numerical value of each element representing a quantified measure of that behavioral feature, such as a psychiatric or neurological clinical rating scale.

[0018] In some embodiments, the architecture of the generative model can be one of a variety of architectures used in the field of machine learning for generative modeling, including transformers, vision transformers, recurrent neural networks, and graph neural networks. The generative model architecture includes an encoder, which is a computational function based on neural networks (FIG. 1, label 2). The encoder maps the input data for an individual subject to a point in the multi-dimensional latent space. The latent space is a multi-dimensional embedding of the input data, of lower dimensionality than the input data (FIG. 1, label 3). As described below and shown in FIG. 2, here the latent space is fractionated into distinct subspaces, with the objective being to separate neural and phenotypic variation across individuals from neural variation across time within an individual. The decoder is a computational function, based on neural networks, that maps from a point in the latent space and generates as output predicted neural activity for brain regions at one or more time points (FIG. 1, label 4). The outputs of thegenerative model are therefore simulated neural activity for brain regions across time points (FIG. 1, label 5). The generative model can simulate neural data time series by iteratively predicting the pattern of brain activity at each subsequent time point.

[0019] The coordinates of the latent space of the generative model provide features which can be used for multivariate classification of individual differences, such as individual differences in response to a therapeutic agent. The latent space embedding of the generative model represents differences across subjects and differences across time within subjects, as both properties are useful for prediction of neural time series. To improve isolation of a representation of differences across subjects, the invention organizes the latent space as fractionated into distinct subspaces. FIG. 2 represents a diagram of the fractionation of the latent space of the generative model. In some embodiments, the fractionated latent subspace contains three subspaces. The Spatio-temporal subspace represents the momentary state of brain activity, and captures fluctuations in brain activity dynamics (FIG. 2, label 1). The Person subspace represents individual differences in neural and phenotypic features, and the subspace coordinates are used as inputs to multivariate classifiers (FIG. 2., label 2). The Therapeutic subspace represents the identity and status of a therapeutic agent administered to a subject, and is used for training the model on neuroimaging datasets that include administration of therapeutic agents (FIG. 2, label 3).

[0020] In some embodiments, the invention may be used for treating a patient with a psychiatric or neurological disorder, based on the predicted response of the patient to one or more therapeutic agents. The coordinates of the latent space, in particular the Person subspace, are used as inputs to a multivariate classifier which are trained to generate as outputs the predicted response of the patient to the therapeutic agent. The multivariate classifier can use a variety of machine learning methods, including multivariate regression, random forests, and neural networks. The predicted response can be multi-dimensional, and include quantified measures phenotypic features, including psychiatric and neurological clinical scores. The predicted response can also include predicted neural features in response to the therapeutic agent, including the Person subspace coordinates. For training data to the classifier, neural features inresponse to the therapeutic agent can be derived from functional neuroimaging datasets in which the patient’s brain activity is imaged both before and after administration of the therapeutic agent.

[0021] FIG. 3 represents a diagram of the multivariate classifier approach to determining scores for evaluating predicted responses to therapeutic agent for a subject. Each subject is represented as a point in the Person subspace of the latent space, or the entire latent space. For each therapeutic agent, a multivariate classifier is trained to predict the subject’s response to the therapeutic agent (FIG. 3, label 2) from the subject’s coordinates in the latent space (FIG. 3, label 1). The outputs of each classifier are predicted responses, one for each phenotypic or neural feature of interest (FIG. 3, label 3). A composite score for each therapeutic agent is then calculated for each set of predicted responses (FIG. 3, label 4). The composite score can implement a weighting of different response features by their importance for clinical outcome, including relative prioritization of clinical symptom scale items. Selection of a therapeutic agent is based on the set of composite scores, such as selecting the therapeutic agent with the highest score among the set of candidate therapeutic agents if the score is above a specified threshold value.

[0022] In some embodiments, the invention may be used for selection of a candidate subject based on their predicted response to a therapeutic agent, including selection for inclusion or exclusion in a clinical trial for the therapeutic agent. This method uses the multivariate classifier that predicts a subject’s response to the therapeutic agent using the coordinates of the latent space of the generative model as inputs. The training dataset for the multivariate classifier includes, for a plurality of subjects, neural data, which can be measured before administration of the therapeutic agent, and phenotypic data measured before and after administration of the therapeutic agent. For each subject, a composite score is calculated from the response to the therapeutic agent. The composite score can implement a weighting of different response features by their importance for clinical outcome, including relative prioritization of clinical symptom scale items. Each candidate subject is thereby assigned a composite score value. Selection of candidate subjects is based on the composite scores. Methods for selection include selectingthose subjects with the highest composite scores, with a specified threshold value or based on the target of selecting a specified percentage of subjects.

[0023] In some embodiments, the invention may be used for selection of a psychiatric or neurological indication with the associated symptom profile based on a predicted response of patients with that indication to a therapeutic agent, which does not require training a multivariate classifier using an empirical dataset comprising patients’ responses to the therapeutic agent. For the therapeutic agent, an associated neural modulation map can be derived. The neural modulation map assigns a numerical value to each brain region which describes the putative relative strength of modulation of each brain region by the therapeutic agent. In different embodiments, the neural modulation map can be derived from distinct experimental modalities. In particular, the neural modulation map can be derived from a brain- wide positron emission tomography (PET) map of receptor occupancy for the therapeutic agent, a brain-wide gene expression map for one or more gene expression targets associated with the therapeutic agent, or functional brain imaging of a baseline or placebo condition to a therapeutically modified condition for the therapeutic agent.

[0024] FIG. 4 represents a diagram of the method of simulating neural activity during modulation by a therapeutic agent by use of the neural modulation map. Using a trained generative model, first, for a subject, their neural and phenotypic data are used to determine coordinates in the Person subspace. The generative model can then generate simulated neural activity. In particular, for simulated time times up to time point t (FIG. 4, label 1), the generative model simulates activity for the next time point t+1 (FIG. 4, label 3). The neural modulation map is used to modulate the topographic activity simulated at time point t+1 with the strength of that modulation for each brain region scaled by the value of the neural modulation map for the brain region (FIG. 2, label 4). This method thereby generates simulated neural activity under a therapeutic modulation.

[0025] With the simulated neural time series under modulation by the therapeutic agent as inputs, the generative model represents the time series as a point in the latent space. Simulatedneural activity under therapeutic modulation for a patient group is thereby represented as a set of points in the Person subspace of the latent space. A next step is quantification of the similarity, as a similarity score, between the coordinates in the Person subspace of the simulated data under therapeutic modulation for a patient group, with a group of healthy subjects. In some embodiments, this similarity score can be defined as the inverse of the mean Euclidean distance between each point for the simulated data of the patient group and each point for the healthy group. The invention thereby generates a similarity score for each patient group which describes the degree to which a simulated effect in brain activity by the therapeutic agent results in brain activity, which is similar, in terms of latent space embedding, to the brain activity of healthy subjects. Selection of a patient group can then be based on which group has the highest similarity score, i.e., which group is predicted to best resemble healthy brain activity states under modulation by the therapeutic agent.DEFINITIONS

[0026] As used herein, the terms “drug” and “therapeutic agent” are used to refer to any substance that is not food or part of a food and provides medical and / or health benefits, including prevention and / or treatment of disease.

[0027] As used herein, the term “generative model” is used to refer to a computational model which is trained through machine learning methods to predict multivariate time series data as outputs from multivariate time series data as inputs, such as predicting future or missing time series data.

[0028] As used herein, the term “latent space” is used to refer to a multi-dimensional space in a generative model as an intermediate representation in the processing of input data to predict output data.

[0029] As used herein, the term “multivariate classifier” is used to refer to a mapping, trained by machine learning methods, to predict a measure of interest, such as a patient’s response to a therapeutic agent, from data points in a multivariate space, such as the coordinates of anembedding space of a generative model in which a patient’s neural data is localized.

[0030] As used herein, the term “phenotypic data” is used to refer to data collected from one or more organisms on a manifested trait or characteristic that can be observed and measured, such as symptom severity measured through a questionnaire, cognitive performance measured through a behavioral task, analyte levels measured by a blood test, or the presence or absence of a gene measured through genetic testing.EXAMPLES

[0031] Numerous modifications and alternative embodiments of the present invention will be apparent to those skilled in the art in view of the foregoing description. Accordingly, this description is to be construed as illustrative only and is for the purpose of teaching those skilled in the art the best mode for carrying out the present invention. Details of the structure may vary substantially without departing from the spirit of the present invention, and exclusive use of all modifications that come within the scope of the appended claims is reserved. Within this specification embodiments have been described in a way which enables a clear and concise specification to be written, but it is intended and will be appreciated that embodiments may be variously combined or separated without parting from the invention. It is intended that the present invention be limited only to the extent required by the appended claims and the applicable rules of law.

[0032] As one example, consider training the model with data recorded from multiple runs for each subject. Data from each run is represented in the latent space by a set of tokens each covering specific brain regions and time points in the recording. For each run, an additional “class” token is included in the set of tokens to represent the run as a whole. Removing a subset of tokens (not including the “class” token) at random, the model is trained to recover the removed tokens and to represent each subject’s identity in the “class” token. Such a model can be used to impute missing data from specific brain regions or time points, and to generate new simulated data e.g. for additional subjects.

[0033] As another example, consider training the model with data from multiple runs for each subject, where some runs for each subject were recorded in the presence of a brain-state altering drug and the remaining runs were recorded without administering such a substance. The model is trained as in the previous example, except that that “class” token is now trained to also distinguish between the presence and absence of a drug in each run. This could be achieved, for example, by treating runs from the same subject in the presence or absence of drug as if these runs originated from different subjects, or by training a drug-presence prediction model on top of the “class” tokens. Such a model can be used to predict whether a subject was under the effect of the drug during a particular recording, or to predict what the effect of the drug would be on a particular subject given a recording where the subject was under the influence of the drug.

[0034] As another example, consider training the model to predict one or more phenotypic variables from “class” tokens, in addition to a training curriculum focusing on time series recovery and subject representation in the “class” token (or any other time-series focused training curriculum). This will imbue each subject’s “class” token with additional phenotypic knowledge. Such a model can be used to predict phenotypic features given a recording of the time series data from a subject.

[0035] It is also to be understood that the following claims are to cover all generic and specific features of the invention described herein, and all statements of the scope of the invention which, as a matter of language, might be said to fall there between.

Claims

CLAIMSWhat is claimed is:

1. A method for identifying a therapeutic agent for treating a patient with a psychiatric or neurological disorder, the method comprising: a. receiving, from a user interface of a computing device or a laboratory test, phenotypic data of a patient comprising or corresponding to at least one of a biological analyte, psychiatric symptom, a neurological symptom, and / or a cognitive status of the patient; b. processing, by one or more processor of the computing device, the phenotypic data of the patient into phenotypic features; c. receiving, by the one or more processor of the computing device, neural data of the patient comprising or corresponding to a functional time series of brain activity of the patient; d. processing, by the one or more processor of the computing device, the neural data of the patient into a parcellated time series, i. wherein the parcellated time series comprises numerical values of brain activity of a set of one or more brain regions at multiple time points, ii. wherein the one or more brain regions are represented as two-dimensional cortical surfaces and / or subcortical volumetric structures; e. retrieving, by the one or more processor of the computing device, a generative model for parcellated time series of brain activity, i. wherein the generative model represents neural data and, optionally, phenotypic data for a plurality of subjects as a set of points in a multidimensional latent space with each point corresponding to one subject, ii. wherein the latent space is trained to separate the representation of differences across the plurality of subjects from representation of differences across time of brain states within each subject;f. generating, by the one or more processor of the computing device, a set of values for the coordinates of the point in the latent space of the generative model to which the patient’s processed neural data and processed phenotypic data are mapped; g. retrieving, by the one or more processor of the computing device, one or more multivariate classifiers, each having inputs and outputs and each associated with a therapeutic agent, i. wherein the inputs to each multivariate classifier are the values of coordinates in the latent space of the generative model and the outputs to each multivariate classifier are one or more features that describe the predicted response to its associated therapeutic agent; h. generating, by the one or more processor of the computing device, the predicted response of the patient to the therapeutic agent based on the outputs of the one or more multivariate classifiers associated with that therapeutic agent; i. wherein the outputs from each multivariate classifier are computed by inputting into each multivariate classifier the patient’s coordinates in the latent space of the generative model; i. determining, by the one or more processor of the computing device, a composite score as a scalar measure of the patient’s predicted response from the one or more multivariate classifier outputs; j. selecting a therapeutic agent based on the composite scores; and k. treating the patient with the selected therapeutic agent.

2. The method of claim 1 , wherein the phenotypic data comprises or corresponds to at least one of a psychiatric symptom, a neurological symptom, and / or a cognitive status of the patient.

3. The method of claim 1, wherein the phenotypic data comprises or corresponds to analyte data corresponding to a biological process, such as a level of a blood-based molecule or a cerebrospinal fluid-based molecule.A method of selecting a candidate patient based on their predicted response to a therapeutic agent, the method comprising: a. receiving, from a user interface of a computing device or a laboratory test, phenotypic data of the patient comprising or corresponding to at least one of a psychiatric symptom, a neurological symptom, and / or a cognitive status of the patient; b. processing, by one or more processor of the computing device, the phenotypic data of the patient into phenotypic features; c. receiving, by the one or more processor of the computing device, neural data of the patient comprising or corresponding to functional time series of brain activity of the patient; d. processing, by the one or more processor of the computing device, the neural data of the patient into a parcellated time series, i. wherein the parcellated time series comprises numerical values of brain activity at multiple time points assigned to a set of one or more brain regions, ii. wherein the one or more brain regions are represented as two-dimensional cortical surfaces and / or subcortical volumetric structures; e. retrieving, by the one or more processor of the computing device, a generative model for parcellated time series of brain activity, i. wherein the generative model represents neural data and, optionally, phenotypic data for a plurality of subjects as a set of points in a multidimensional latent space with each point corresponding to one subject, ii. wherein the latent space is trained to separate the representation of differences across the plurality of subjects from representation of differences across time of brain states within each subject; f. generating, by the one or more processor of the computing device, a set of values for the coordinates of the point in the latent space of the generative model towhich the patient’s processed neural data and processed phenotypic data are mapped; g. retrieving, by the one or more processor of the computing device, one or more multivariate classifiers, each having inputs and outputs and each associated with a therapeutic agent, i. wherein the inputs to each multivariate classifier are the values of coordinates in the latent space of the generative model and the outputs to each multivariate classifier are one or more features that describe the predicted response to its associated therapeutic agent; h. generating, by the one or more processor of the computing device, the predicted response of the patient to the therapeutic agent based on the outputs of the multivariate classifiers associated with the therapeutic agent; i. wherein the multivariate classifier outputs are computed by inputting into a multivariate classifier the patient’s coordinates in the latent space of the generative model; i. determining, by the one or more processor of the computing device, a composite score as a scalar measure of the subject’s predicted response from the multivariate classifier outputs; and j . selecting the patient predicted to respond to the therapeutic agent based on the composite score.

5. The method of claim 4, wherein the phenotypic data comprises or corresponds to least one of a psychiatric symptom, a neurological symptom, and / or a cognitive status of the patient.

6. The method of claim 4, wherein the phenotypic data comprises or corresponds to analyte data corresponding to a biological process, such as a level of a blood-based molecule or a cerebrospinal fluid-based molecule.A method of selecting a psychiatric or neurological indication with the associated symptom profile based on a predicted response of patients with that indication to a therapeutic agent, the method comprising: a. receiving, from a user interface of a computing device or a laboratory test, phenotypic data of one or more groups of patients, comprising or corresponding to at least one of a psychiatric symptom, a neurological symptom, and / or a cognitive status of the patients, i. wherein each group of patients is representative of a psychiatric or neurological indication; b. processing, by one or more processor of the computing device, the phenotypic data of each patient, into phenotypic features; c. receiving, by the one or more processor of the computing device, neural data of the one or more groups of patients, comprising or corresponding to functional time series of brain activity of the patients; d. processing, by the one or more processor of the computing device, the neural data of each patient into a parcellated time series, i. wherein the parcellated time series comprises numerical values of brain activity at multiple time points assigned to a set of one or more brain regions, ii. wherein the one or more brain regions are represented as two-dimensional cortical surfaces and / or subcortical volumetric structures; e. retrieving, by the one or more processor of the computing device, a generative model for parcellated time series of brain activity, i. wherein the generative model represents neural data and, optionally, phenotypic data for a plurality of subjects as a set of points in a multidimensional latent space with each point corresponding to one subject, ii. wherein the latent space is trained to separate the representation of differences across the plurality of subjects from representation of differences across time of brain states within each subject;f. generating, by the one or more processor of the computing device, a set of values for the coordinates of the point in the latent space of the generative model to which each patient’s processed neural data and processed phenotypic data are mapped; g. retrieving, by the one or more processor of the computing device, a neural modulation map, corresponding to the therapeutic agent, i. wherein the map comprises numerical values assigned to the same set of brain regions as represented by the parcellated time series, ii. wherein the brain regions are represented as two-dimensional cortical surfaces and / or subcortical volumetric structures; h. simulating, by the one or more processor of the computing device, artificial neural data for each patient in each patient group, using the generative model with modulation of simulated neural activity in each brain region, i. wherein the strength of modulation for each brain region is determined by the scalar value of the neural modulation map; i. generating, by the one or more processor of the computing device, a set of values for the coordinates of the point in the latent space of the generative model to which each patient’s simulated neural data is mapped, for each patient group; j. retrieving, by the one or more processor of the computing device, a set of latent space coordinates of a plurality of healthy subjects; k. generating, by the one or more processor of the computing device, a numerical similarity score quantifying the similarity between the set of latent space coordinates of healthy subjects in the generative model and the simulated neural data of the patient group, for each patient group; and l. selecting a psychiatric or neurological indication based on the patient groups’ similarity scores.

8. The method of claim 7, wherein the neural modulation map is derived from a positron emission tomography (PET) map.

9. The method of claim 7, wherein the neural modulation map is derived from a gene expression map.

10. The method of claim 7, wherein the neural modulation map is derived from functional brain imaging of a baseline or placebo condition to a therapeutically modified condition.

11. The method of claim 7, wherein the phenotypic data comprises or corresponds to at least one of a psychiatric symptom, a neurological symptom, and / or a cognitive status of the patient.

12. The method of claim 7, wherein the phenotypic data comprises or corresponds to analyte data corresponding to a biological process, such as a level of a blood-based compound or a cerebrospinal fluid-based compound.

Images