Determining effect of a drug based on brain activity and a trained machine learning model

Abstract

A computer-implemented method for determining an effect of a drug on a subject, including: obtaining brain activity data of the subject; determining, using a trained machine learning model and the brain activity data of the subject, an effect of the drug administered to the subject, the trained machine learning model trained with brain activity data of subjects treated with at least one of the drug or a second drug and with brain activity data of subjects treated with a placebo; and providing an output indicative of the effect of the drug administered to the subject determined by the trained machine learning model.

Description

DETERMINING EFFECT OF A DRUG BASED ON BRAIN ACTIVITY ANDA TRAINED MACHINE LEARNING MODELCROSS-REFERENCE TO RELATED APPLICATION

[0001] This application claims priority to U.S. Provisional Patent Application No. 63 / 538,746, filed September 15, 2023, which is incorporated herein by reference in its entirety.BACKGROUND

[0002] According to the United Stated Centers for Disease Control and Prevention, adult epilepsy has an estimated prevalence of 3 million people in the United States, with approximately 31,000 patients diagnosed each year, the majority of whom have focal epilepsy. Focal epilepsy is characterized by localized (focal) areas of neuronal / network hyperexcitability that disrupt normal brain function by causing periods of abnormal network synchronization (seizure).

[0003] Sodium channel (NaV) blockers are commonly used to treat focal epilepsy, whereby these NaV blockers reduce pathologic neuronal hyperexcitability to prevent the development of seizures or terminate a seizure state. However, approved NaV blockers are not well tolerated, possibly due to an inability to selectively target hyperexcitable states. Currently available standard-of-care NaV blockers used to treat epilepsy are limited by a low therapeutic index and a requirement to titrate to an efficacious concentration while managing tolerability. Significant class-related adverse effects are central nervous system (CNS) related, including ataxia, drowsiness, and dizziness, which may be due to the excessive and long-lasting inhibition of peak INa resulting in an inability of existing NaV blockers to selectivity reduce hyperexcitability while sparing normal brain function.

[0004] As such, approximately 30% of those with diagnosed focal epilepsy are not adequately controlled using available therapeutic options due to lack of efficacy and / or tolerability issues.

[0005] In addition to focal epilepsy, many suffer from other neurological diseases, including other epilepsies (e.g., developmental and epileptic encephalopathy (DEE)) and essential tremor. DEE is a severe form of epilepsy characterized by frequent tonic seizures or spasms beginning in infancy. Essential tremor is the most common neurologic cause of postural or action tremor and may present as a bilateral postural tremor of the hands.Essential tremor is a disorder with high unmet patient needs. Treatment options are limited, with high discontinuation rates due to poor tolerability and modest efficacy.SUMMARY OF THE INVENTION

[0006] According to one or more embodiments described herein, for a computer- implemented method for determining an effect of a drug on a subject, the method may comprise: obtaining brain activity data of the subject; determining, using a trained machine learning model and the brain activity data of the subject, an effect of the drug administered to the subject, the trained machine learning model trained with brain activity data of subjects treated with at least one of the drug or a second drug and with brain activity data of subjects treated with a placebo; and providing an output indicative of the effect of the drug administered to the subject determined by the trained machine learning model.

[0007] For the computer-implemented method, the brain activity data of the subject may be based on electroencephalography (EEG) data of the subject obtained from a plurality of electrodes attached to the head of the subject and processed using quantitative electroencephalography (qEEG). For the computer-implemented method, the brain activity data of the subject may be based on band values for frequency bands of frequency-transformed EEG data of the subject separated into the frequency bands. For the computer- implemented method, the band values of the frequency bands may be based on a plurality of time periods of the EEG data of the subject. For the computer-implemented method, the band values of the frequency bands may be averages of band values for a plurality of electrodes attached to the head of the subject.

[0008] For the computer-implemented method, the trained machine learning model may output weights for a combination of the band values. For the computer-implemented method, the output indicative of the effect of the drug administered to the subject may be based on a weighted combination of the band values using the weights for the band values output by the trained machine learning model. For the computer-implemented method, the weighted combination of the band values may be processed using at least one of a negative log loss function, z-scoring, or subtracting of baseline values.

[0009] For the computer-implemented method, the trained machine learning model may be trained with brain activity data of subjects treated with the second drug and not trained with brain activity data of subjects treated with the drug. For the computer-implemented method, the trained machine learning model may be trained with brain activity data of subjects treated with the drug and brain activity data of subjects treated with the second drug. For the computer-implemented method, the trained machine learning model may be further trained with brain activity data of subjects treated with a third drug. For the computer- implemented method, the drug administered to the subject may not be one of the drugs used on subjects to obtain brain activity data to train the machine learning model.

[0010] For the computer-implemented method, the drug administered to the subject may be a drug to treat a brain-related disorder. For the computer-implemented method, the drug administered to the subject may be a drug to treat at least one of epilepsy or a movementdisorder. For the computer-implemented method, the drug administered to the subject may be a single medicinal drug or a combination of medicinal drugs.

[0011] For the computer-implemented method, the machine learning module may be trained with one of logistic regression or binary classification logistic regression. For the computer-implemented method, the machine learning module may be trained with one of linear regression, a random forest model, a gradient-boosted tree, a neural network, or a deep learning model. For the computer-implemented method, the method may further comprise training a machine learning model to obtain the trained machine learning model, wherein the machine learning model may be trained with the brain activity data of subjects treated with at least one of the drug or another drug and with the brain activity data of subjects treated with a placebo.

[0012] According to one or more embodiments described herein, a non-transitory processor readable medium may contain a set of instructions thereon for determining an effect of a drug on a subject, wherein when executed by one or more processors, the instructions may cause the one or more processors to perform a method comprising: obtaining brain activity data of the subject; determining, using a trained machine learning model and the brain activity data of the subject, an effect of the drug administered to the subject, the trained machine learning model trained with brain activity data of subjects treated with at least one of the drug or a second drug and with brain activity data of subjects treated with a placebo; and providing an output indicative of the effect of the drug administered to the subject determined by the trained machine learning model.

[0013] According to one or more embodiments described herein, for an apparatus for determining an effect of a drug on a subject, the apparatus may comprise: one or more processors; and memory accessible by the one or more processors. The memory may store instructions that when executed by the one or more processors, cause the apparatus toperform a method comprising: obtaining brain activity data of the subject; determining, using a trained machine learning model and the brain activity data of the subject, an effect of the drug administered to the subject, the trained machine learning model trained with brain activity data of subjects treated with at least one of the drug or a second drug and with brain activity data of subjects treated with a placebo; and providing an output indicative of the effect of the drug administered to the subject determined by the trained machine learning model.BRIEF DESCRIPTION OF THE DRAWINGS

[0014] FIG. 1 depicts an example process for determining the effect of a drug based on the brain activity of a subject according to one or more embodiments described herein.

[0015] FIG. 2 depicts a composite qEEG metric for a clinical trial.

[0016] FIGS. 3A and 3B provide data on the accuracy of a machine learning model.

[0017] FIG. 4 depicts a composite qEEG metric for a clinical trial.

[0018] FIGS. 5 A and 5B provide data on the accuracy of a machine learning model.

[0019] FIG. 6 depicts a composite qEEG metric for a clinical trial.

[0020] FIGS. 7 A and 7B provide data on the accuracy of a machine learning model.

[0021] FIG. 8 depicts a composite qEEG metric for a clinical trial.

[0022] FIGS. 9 A and 9B provide data on the accuracy of a machine learning model.

[0023] FIG. 10 depicts a composite qEEG metric for a clinical trial.

[0024] FIGS. 11 A and 1 IB provide data on the accuracy of a machine learning model.

[0025] FIG. 12 depicts an example computer apparatus according to one or more embodiments described herein.DESCRIPTION OF EMBODIMENTS

[0026] This application describes exemplary techniques for determining the effect of a drug, such as a drug to treat focal epilepsy, other epilepsies, including developmental and epileptic encephalopathy (DEE), or other brain-related disorders, including for example movement disorders, such as essential tremor, based on the brain activity of a subject. The exemplary techniques include using a trained machine learning model to make the determination.

[0027] There is a significant need for better tolerated NaV blockers to treat patients with focal epilepsy, as well as other drugs to treat patients with brain-related disorders. To help satisfy this significant need, techniques are disclosed herein for evaluating and determining the effect of a drug based on the brain activity of a subject. The inventive techniques involve training a machine learning model to recognize the effect of a drug on a subject based on the brain activity of the subject. The trained machine learning model can then be used to determine the effect of the drug, or another drug, on other subjects.

[0028] Further, using a single parameter (such as the signal from a single electrode attached to the head of the subject) can be underpowered and unreliable in predicting effects of a drug on a subject. Also, many parameters can be derived from (e.g., calculated from) a single electrode signal, and even more with multiple electrode signals, and it is difficult to ascertain which parameters to select and in what combination to use the parameters in order to predict the effect of a drug on a subject. Further, the available data set to analyze can be relatively small, have multicollinear (e.g., redundant) features in the data, and have interpretability difficulties. To address these problems, a machine learning module may be used to automatically learn optimal parameters to separate a subject treated with a drug and a subject treated with a placebo. By the inventive techniques, a composite of brain activity data of a subject may be determined using optimized parameters (e.g., weighting coefficients)determined by the machine learning module. Further, in exemplary embodiments, the composite of the brain activity data of a subject may be derived from a machine learning module employing binary classification logistic regression.

[0029] The techniques described herein provide a practical application to learn and determine the effect of a drug on a subject based on the brain activity of the subject. Due to the large amounts of data from recorded brain activity and the difficulty in determining any relationships in the data, machine learning may be useful to address these data complexities in determining the effect of a drug on a subject based on the brain activity of the subject. As such, manipulating the data and performing the required calculations cannot be performed in the human mind. Instead, an electronic computing device (e.g., a computer apparatus having at least one processor) is needed to implement the inventive machine learning techniques and thereby provide computational speed, computational accuracy, and data management.

[0030] FIG. 1 depicts an example process for determining the effect of a drug based on the brain activity of a subject according to one or more embodiments described herein. The drug may be, for example, a drug to treat focal epilepsy, another epilepsy, a movement disorder (e.g., essential tremor) or other brain-related disorders. The steps in the process are computer implemented.

[0031] In step 102, brain activity data of subjects treated with a drug and / or brain activity data of other subj ects treated with another drug are obtained, and brain activity data of further subjects treated with a placebo are obtained. Although the treated subjects may be treated with a single drug, the obtained brain activity data may include data for subjects treated with a first drug and other subjects treated with a second drug, and even include further subjects treated with a third drug and so on. Each “drug” administered to a subject may be a single medicinal drug or a combination of medicinal drugs (e.g., a drug cocktail).

[0032] The brain activity data may be stored on a processor readable medium (such as computer-readable memory) and then obtained from the processor readable medium. The obtained brain activity data may be electroencephalography (EEG) data of the subjects or processed EEG data of the subjects.

[0033] EEG data of the subjects may be data recorded from EEG signals received from the subjects treated with at least one of a drug or another drug and the subjects treated with a placebo. The EEG signals may be obtained from a number of electrodes attached to the head (or the scalp) of each subject. For example, electrodes may be placed on the subject according to the international 10-20 system or a higher resolution system, such as the modified combinatorial nomenclature (MCN) system. The number of electrodes used may be any number of electrodes, including in various embodiments, 20, 64, or 128 electrodes.

[0034] Processed EEG data of the subjects may be EEG data of the subjects processed for use in training a machine learning model. The EEG data may be processed using, for example, the techniques discussed below in the Experimental Results sections. The EEG data may be processed using, for example, quantitative electroencephalography (qEEG). For example, the EEG data from each electrode may be a time series of data and may be transformed into frequency data using, for example, a fast Fourier transform (FFT) or another time-to-frequency conversion technique.

[0035] The frequency data may be segmented into brain wave bands. For example, brain wave bands may be bands for gamma waves, beta waves, alpha waves, theta waves, and delta waves. For example, the brain wave bands may include a band from 1.5 Hz to less than 6.0 Hz, a band from 6.0 Hz to less than 8.5 Hz, a band from 8.5 Hz to less than 12.5 Hz, a band from 12.5 Hz to less than 30.0 Hz, and a band from 30.0 Hz to less than 40 Hz, where the bounds of each band may be exact or approximate (i.e., “about”). For example, the brain wave bands may include a band of less than 4 Hz (or between 0.0 Hz and 4 Hz), a bandbetween 4 Hz and 8 Hz, a band between 8 Hz and 12 Hz, a band between 12 Hz and 30 Hz, and a band greater than 30 Hz, where the bounds of each band may be exact or approximate (i.e., “about”). For example, the brain wave bands may include brain wave sub-bands, such as the brain wave sub-bands of low-range beta waves, mid-range beta waves, and high-range beta waves. For example, the brain wave bands may include a band between 12 Hz and 15 Hz, a band between 15 Hz and 20 Hz, and a band between 20 Hz and 30 Hz, where the bounds of each band may be exact or approximate (i.e., “about”). For example, the brain wave bands may have any number of bands, such as 5 bands, 7 bands, 10 bands, 20 bands, 30 bands, 40 bands, or any number greater, lesser, or in between.

[0036] Once the frequency data is segmented into bands, a band value may be assigned to each band. For example, the band value for a band may be the amount of signal energy in the band. For example, the band value for a band may be calculated as an area under the frequency curve for the band. For example, the band values for the bands may be based on spectral analysis of the frequency data.

[0037] Further, the brain activity data may be processed into band values over a time period or over a number of time periods for each electrode signal. The brain activity data may be processed into band values for each electrode signal or for a combination of the electrode signals. For example, an average band value for each band may be calculated from the band values of the electrode signals. The brain activity data may be processed into band values over a time period or over a number of time periods for a combination of the electrode signals, such as by averaging.

[0038] In step 104, a machine learning model is trained with obtained brain activity.The machine learning model may be trained with processed brain activity, such as with determined band values for brain activity. For example, the band values for brain activity may be the input to the machine learning model. The machine learning model may be trainedto output weights for a combination of the band values. As an example, the weighted combination of the band values may be a weighted linear combination of the band values, or a weighted nonlinear combination of the band values. The weighted combination of the band values may identify whether the subject was administered a drug or was administered a placebo.

[0039] The machine learning model may be, for example, based on a logistic regression algorithm or a binary classification logistic regression algorithm. Other machine learning models may be used, such as, for example, linear regression, random forest models, gradient- boosted trees, neural networks, or deep learning models.

[0040] The machine learning model may be a software program stored in memory and executable by one or more processors. The machine learning model may be a combination of hardware and / or software.

[0041] In step 106, brain activity data of a subject treated with a drug may be obtained. The brain activity data of the subject may be obtained and processed in the same manner as the brain activity data in step 102. The drug administered to the subject may be the same drug (or drugs) or a different drug (or drugs) than the drug (or drugs) used in the brain activity data obtained in step 102.

[0042] In step 108, using the trained machine learning model and the brain activity data of the subject, an effect of the drug administered to the subject is determined. For example, the band values for the brain activity data of the subject may be input to the trained machine learning model, and the trained machine learning model may provide an output, such as weights to combine the band values, or the calculating probability of a given input being part of a target class.

[0043] In step 110, an output indicative of the effect of the drug administered to the subject determined by the trained machine learning model may be provided. For example,the weighted combination of the band values may be an intermediate output. The output may be provided on an output device of a computer, such as a monitor, or may be provided to another computer or computing device for display thereon.

[0044] The weighted combination may be processed further by other functions, such as the negative log loss function, or by other techniques, such as z-scoring or subtracting out baseline values.

[0045] After training, the accuracy of the machine learning model may be evaluated using various metrics. For example, the accuracy of the machine learning model may be measured as a percentage of accurately classifying the training data. For example, the fl score for the machine learning model may be determined. An fl score combines precision (e.g., how many predicted treated subjects were truly treated subjects) and recall (i.e., how many true treated subjects were correctly predicted). An fl score of 1 is a perfect performance, and an fl score of 0 is a complete imperfect performance. For example, a comparison of the true positive rate and the false positive rate of the machine learning model may be determined, and an area under the receiver-operator curve (auROC) may be determined. For example, a comparison of the precision and recall may be determined, and an area under the precision-recall curve (auPRC) may be determined.

[0046] After training, if the machine learning model accurately models the drug effects, the machine learning model may be used to evaluate other subjects. If the weighted combination of the band values is closer to the weighted combination for the training data for the administered drug or drugs, the drug administered to the subject may have a significant effect on the subject. However, if the weighted combination of the band values is closer to the weighted combination for the placebo training data, the drug administered to the subject may not have a significant effect on the subject.

[0047] With the invention, the machine learning model may be trained with brain activity for subjects administered a drug (or a set of drugs) and may be used to test brain activity for other subjects administered the drug (or the set of drugs). In such use, the machine learning model may be used to determine the effectiveness of the drug on the other subjects compared to subjects used to train the machine learning model.

[0048] With the invention, the machine learning model may be trained with brain activity for subjects administered a first drug (or a set of first drugs) and may be used to test brain activity for subjects administered a second drug (or second set of drugs). In such use, the machine learning model may be used to determine the effectiveness of the second drug (or second set of drugs) compared to the first drug (or the first set of drugs).EXPERIMENTAL RESULTS: Compound 1

[0049] An experimental clinical trial was designed to evaluate a range of doses ofCompound 1 capsules in healthy participants. Compound l is a drug produced by PraxisPrecision Medicines, and the chemical structure of Compound 1 is:Compound l is a next-generation, functionally selective small molecule targeting the hyperexcitable state of sodium-channels in the brain. Compound 1 is being developed as a once daily, oral treatment for adult focal onset epilepsy.

[0050] The clinical trial was composed of two parts, namely Part A and Part B. Part A (randomized, double-blind, placebo controlled) was designed to assess the safety, tolerability, PK, and PD of single ascending oral doses (SAD) of Compound 1. Part B (randomized,double blind, placebo-controlled) was designed to assess the safety, tolerability, PK, and PD of multiple ascending doses (MAD) of Compound 1 (selected based on Part A data).

[0051] Randomization was employed in Parts A and B to minimize participant selection bias and to increase the likelihood that known and unknown participant attributes (e.g., demographics and baseline characteristics) were evenly balanced between the two treatments (Parts A and B) and also to enhance the validity of comparisons across treatments. Parts A and B were conducted in a double-blind manner, whereby both participants and investigator were blinded to treatment assignments to ensure the objectivity of safety and tolerability assessments for the primary endpoint.

[0052] The clinical trial population was composed of healthy participants, without confounding medical or psychiatric disorders that would jeopardize the safety of participants or the scientific validity of the clinical trial.

[0053] Eight participants per cohort were planned. The total number of participants required depended upon the number of escalation steps. In Part A, approximately 32 participants, or up to 56 participants, were planned to be administered Compound 1 or placebo. In Part B, approximately 24 participants, or up to 40 participants, were planned to be administered Compound 1 or placebo.

[0054] Table 1 and Table 2 provide dose, frequency, route, and duration for Part A and Part B, respectively.

[0055] Table 1. Investigational Products Used in the Clinical Trial for Part A.N / A=not applicable

[0056] Table 2. Investigational Products Used in the Clinical Trial for Part B.N / A=not applicable.

[0057] In Parts A and B of the clinical trial, participants were randomized in a 3 : 1 ratio to either Compound 1 or matching placebo. The randomization scheme was prepared by an unblinded statistician not otherwise involved in the clinical trial and provided to an unblinded clinical trial pharmacist, or approved unblinded designee, at the clinical site. The study drug was dispensed by the unblinded clinical trial pharmacist, or designee, in a blinded fashion to participants in Parts A and B. Once a randomization number was assigned to a participant, it was not reassigned to another participant.

[0058] ECGs were obtained using an ECG machine that automatically calculated heart rate and measured PR, QRS, QT, and QTc intervals.

[0059] In Parts A and B of the clinical trial, resting state EEGs were performed under eyes-open and eyes-closed conditions. Participants were semi-supine, and, at a minimum, recording took place with eyes open for 5 minutes followed by eyes closed for 5 minutes. Each participant was asked to refrain from talking, moving, and blinking during the recordings and to not fall asleep. Each resting-state EEG recording employed alternating periods with eyes open and closed with a duration of 64-seconds for each period. Subjects faced a featureless wall and were instructed not to stare, not to move their head and eyes, and to suppress eye blinks.

[0060] The EEG of each subject was continuously recorded using a 40-channel recording system. Electrodes were placed according to the international 10-20 system, but replacing electrodes placed at the earlobes (i.e., Al and A2) with electrodes placed at the mastoids (i.e., Ml and M2). The scalp electrode impedance was kept below 5k . The groundelectrode was placed at AFz. Additionally, to detect ocular artefacts, vertical and horizontal electrooculogram (EOG) were also recorded. Two Ag / AgCl electrodes were placed at the outer canthi of both eyes, and two Ag / AgCl electrodes were placed approximately 2 cm above and below the right eye. All signals were sampled at a sampling rate of 1024 Hz and were filtered prior to storage using a first order recursive high-pass filter with a cut-off frequency at 0.1 Hz. Digital markers were recorded by an amplifier indicating the start and end of each eye state.

[0061] Recorded channels were band-pass filtered using a third order Butterworth filter with cut-off frequencies at 0.5 Hz and 45.0 Hz. The filtered signals were then divided into four second epochs. Epochs containing ocular artifacts were removed for further analysis. A power spectrum density (PSD) was calculated for each epoch and averaged for each eye state. The resulting PSDs were then subdivided into five bands, and the total power per band was calculated. The five bands used are: delta band (1.5 Hz to less than 6.0 Hz), theta band (6.0 Hz to less than 8.5 Hz), alpha band (8.5 Hz to less than 12.5 Hz), beta band (12.5 Hz to less than 30.0 Hz), and gamma band (30.0 Hz to less than 40 Hz).

[0062] For qEEG, for the eyes open state and for each spectral power band (alpha, beta, theta, delta, and gamma), absolute power was computed. To remove baseline effects and normalize the data, qEEG data was scaled within each subject by subtracting the mean of each subject’s baseline qEEG measurements and dividing by the standard deviation of each subject’s baseline qEEG measurements. To evaluate the CNS activity of Compound 1, change from baseline was computed for each qEEG spectral band at specified time points for Parts A and B.

[0063] A machine learning method was developed to assess the composite change from baseline over all spectral bands. These composites consisted of linear combinations of scaled qEEG absolute power scores with optimized coefficients determined by a binaryclassification logistic regression algorithm. The algorithm was trained to discriminate the treatment from the placebo arm. The model was only trained on eyes closed data. The resulting composite score, an interpretable, optimized linear combination of the spectral powers, was then calculated.

[0064] The algorithm was trained on 75% of the scaled spectral power band data to fit coefficients and tested on 25% of the data for evaluation. The overall training and test sets were balanced in ratio of treatment to placebo measurements and included data from multiple time points to allow the model to make generalized predictions.

[0065] k-fold Cross-validation was used to select an optimal model. First, within the training set, the data was split into k smaller sets. Over k times, one set served as a testing set (i.e., a cross-validation testing set) and the rest served as a training set (i.e., a cross-validation training set). Once the model was fit on the cross-validation training set, the model was tested on the cross-validation testing set. This was repeated k times with each set serving as the cross-validation testing set once. The loss functions (e.g., a metric representing error of prediction) was calculated on each run and averaged across runs. The model with the minimum average loss from the cross-validation was selected and then tested and evaluated on the overall test set.

[0066] After computing the composite qEEG metric, mixed model repeated measures (MMRM) was used to assess the differences in the composite metric between placebo and treatment arms.

[0067] FIG. 2 depicts the composite qEEG metric for Part A of the clinical trial. The vertical axis is log odds of each composite qEEG metric, and the horizontal axis is the number of hours after administering the dose in Part A. The brain activity data collected was for eyes closed. The dark lines are the average of the composite qEEG metric for each subject, and the light lines are the composite qEEG metric for each subject. According to thecomposite qEEG metric, the three doses (5 mg, 15 mg, and 45 mg) exhibit a marked increase in effect from taking the dose (Visit 0 in FIG. 2) to the second hour (Visit 2 in FIG. 2). In addition, according to the composite qEEG metric, the three doses (5 mg, 15 mg, and 45 mg) exhibited decreased effectiveness at the fourth hour (Vi st 4 in FIG. 2) and at the eighth hour (Visit 8 in FIG. 2). Further, the composite qEEG metric correctly identified the test data for subjects administered the drug and the test data for subjects administered the placebo. In particular, the accuracy was 94.4%, and the fl score was 0.963. An fl score combines precision (i.e., how many predicted Compound 1 subjects were truly Compound 1 subjects) and recall (i.e., how many true Compound 1 subjects were correctly predicted). An fl score of 1 is a perfect performance, and an fl score of 0 is a complete imperfect performance.

[0068] FIGS. 3A and 3B provide further data on the accuracy of the machine learning model. FIG. 3A depicts a graph comparing the true positive rate and the false positive rate. For the logistic regression machine learning model, the area under the receiver-operator curve (auROC) was 0.98. FIG. 3B depicts a graph comparing precision and recall. For the logistic regression machine learning model, the area under the precision-recall curve (auPRC) was 0.99.

[0069] FIG. 4 depicts the composite qEEG metric for Part B of the clinical trial. The vertical axis is log odds of each composite qEEG metric, and the horizontal axis is the number of days administering the dose in Part B. The dark lines are the average of the composite qEEG metric for each subject, and the light lines are the composite qEEG metric for each subject. The brain activity data collected was for eyes closed. According to the composite qEEG metric, the two doses (20 mg and 30 mg) exhibit a marked increase in effect from taking the dose (Visit 0 in FIG. 2) to four days later (Visit 4 in FIG. 2). Also, according to the composite qEEG metric, the two doses (20 mg and 30 mg) continued to exhibit a continued effect between the fourth day (Visit 4 in FIG. 2) and the ninth day (Visit 9 in FIG.2). Further, the composite qEEG metric correctly identified the test data for subjects administered the drug and the test data for subjects administered the placebo. In particular, the accuracy was 84.4%, and the fl score was 0.898.

[0070] FIGS. 5 A and 5B provide further data on the accuracy of the machine learning model. FIG. 5A depicts a graph comparing the true positive rate and the false positive rate. For the logistic regression machine learning model, the auROC was 0.93. FIG. 5B depicts a graph comparing precision and recall. For the logistic regression machine learning model, the auPRC was 0.98.

[0071] Four pharmacodynamics (PD) findings in this clinical trial include the following.

[0072] First, administration of a single dose of Compound 1 (5-45 mg) resulted in significant changes in the qEEG composite metric for subjects administered Compound 1 versus placebo. The change was consistent with the PK profile of the drug.

[0073] Second, significant changes in the qEEG composite metric were observed across all Day 4 and Day 9 time points for Compound 1 20-mg and 30-mg doses when compared to placebo.

[0074] Third, the changes in the qEEG metrics are an indicator of brain penetrance and target engagement for Compound 1.

[0075] Fourth, the weights of the linear combinations comprising the composite metrics calculated indicate where key differences in neural activity patterns lay between placebo- treated and Compound 1 -treated subjects. In the SAD portion of the study, the high weighting of gamma, delta, alpha, and beta powers is consistent with findings that combining gamma with alpha and beta waves improves diagnostic power for focal epilepsy ([Archila-Melendez et al 2020]), along with the association of slow-wave powers with focal cerebral dysfunction ([Britton et al 2016]). In the MAD portion of the study, the high weighting of theta, alpha, and beta powers is consistent with the slow-wave association with focal cerebral dysfunctionas well as findings associating abnormal alpha wave activity with poor seizure control ([Abela et al 2018]).

[0076] The PD results from Part A (SAD) and Part B (MAD) indicate the expected target engagement of Compound 1 and underscore the drug’s capacity to modulate brain activity.

[0077] In the clinical trial, the composite qEEG metric comprising all spectral bands was created using a machine learning algorithm that optimized a linear combination of spectral powers to detect treatment. This composite metric showed significant differences between healthy volunteers treated with a placebo and with Compound 1 after first administration in Part A and over the majority of time points measured in Part B.

[0078] In addition, changes in the qEEG composite metric were seen consistently in all doses of Compound 1 in both the SAD and MAD portions and were consistent with the known pharmacokinetics (PK) parameters of the drug, further reinforcing its validity.

[0079] Biomarkers like the qEEG may play an important role in the early phases of drug development, offering potentially early and real-time insights into a drug's activity in the brain, which is crucial for central nervous system (CNS) drugs like Compound 1. The use of biomarkers like qEEG, especially in early study phases, accentuates the importance of confirming brain penetrance and target engagement, ensuring the drug's viability and therapeutic efficacy in subsequent stages of development.EXPERIMENTAL RESULTS: Compound 2

[0080] In a second experiment, the inventive techniques using the machine learning model were applied to an experimental clinical trial using Compound 2. Compound 2 is also a drug produced by Praxis Precision Medicines, and its chemical structure is:Compound 2 is a first-in-class small molecule in development for the treatment of developmental and epileptic encephalopathies (DEEs) as a preferential inhibitor of persistent sodium current, shown to be a key driver of seizure symptoms in early onset SCN2A-DEE and SCN8A-DEE. SCN2A-DEE is a monogenic epilepsy disorder caused by a variant in the SCN2A gene. SCN8A-DEE is a rare developmental and epileptic encephalopathy caused by a variant in the SCN8A gene.

[0081] Similar to the training and testing described above for Compound 1, the machine learning model was trained and tested based on the experimental clinical trial using Compound 2. The model was only trained on eyes closed data. Here, the clinical trial in MAD Part A evaluated the effects of 90 mg of Compound 2 over 28 days versus placebo. The clinical trial is further discussed in [Mahalingam 2023],

[0082] FIG. 6 depicts the composite qEEG metric for the clinical trial. The vertical axis is log odds of each composite qEEG metric, and the horizontal axis is the number of days administering the dose. The dark lines are the average of the composite qEEG metric for each subject, and the light lines are the composite qEEG metric for each subject. The brain activity data collected was for eyes closed. According to the composite qEEG metric, the administered drug exhibited a marked increase in effect from before taking the drug (Visit 0) to the first day (Visit 1). Also, according to the composite qEEG metric, the drug continued to exhibit a continued increased effect for each measurement. Further, the composite qEEG metric correctly identified the test data for subjects administered the drug and the test data for subjects administered the placebo. In particular, the accuracy was 86.1%, and the fl score was 0.91.

[0083] FIGS. 7 A and 7B provide further data on the accuracy of the machine learning model. FIG. 7A depicts a graph comparing the true positive rate and the false positive rate. For the logistic regression machine learning model, the auROC was 0.93. FIG. 7B depicts a graph comparing precision and recall. For the logistic regression machine learning model, the auPRC was 0.98.EXPERIMENTAL RESULTS: Compound 2 versus Another Compound

[0084] In a third experiment, the inventive techniques using the machine learning model were applied to an experimental clinical trial using Compound 2 compared to another compound, instead of compared to a placebo. The other compound was oxcarbazepine, and its chemical structure is:

[0085] Similar to the training and testing described above for Compound 1, the machine learning model was trained and tested based on the experimental clinical trial using Compound 2 compared to the other compound. The model was only trained on eyes closed data. Here, the clinical trial in MAD Part B evaluated the effects of 120 mg of Compound 2 over 28 days versus the other compound. The clinical trial is further discussed in [Praxis 2022],

[0086] FIG. 8 depicts the composite qEEG metric for the clinical trial. The vertical axis is log odds of each composite qEEG metric, and the horizontal axis is the number of days administering the dose. The dark lines are the average of the composite qEEG metric for each subject, and the light lines are the composite qEEG metric for each subject. Accordingto the composite qEEG metric, the administered drug exhibited a marked increase in effect and continued increase effect over the days of the trial. Further, the composite qEEG metric correctly identified the test data for subjects administered the drug and the test data for subjects administered the other compound. The accuracy of the logistic regression of Compound 2 versus the other compound was 84.6%, and the fl score was 0.882.

[0087] FIGS. 9 A and 9B provide further data on the accuracy of the machine learning model. FIG. 9A depicts a graph comparing the true positive rate and the false positive rate. For the logistic regression machine learning model, the auROC was 0.98. FIG. 7B depicts a graph comparing precision and recall. For the logistic regression machine learning model, the auPRC was 0.99.EXPERIMENTAL RESULTS: Compound 3

[0088] In a third experiment, the inventive techniques using the machine learning model were applied to an experimental clinical trial using Compound 3. Compound 3 is also a drug produced by Praxis Precision Medicines, and its chemical structure is:Compound 3 is a novel selective T-type calcium channel blocker currently in clinical development for the treatment of essential tremor.

[0089] Similar to the training and testing described above for Compound 1, the machine learning model was trained and tested based on the experimental clinical trial using Compound 3 compared to a placebo. The model was only trained on eyes closed data. Here, the clinical trial in MAD Part B evaluated the effects of increased dosages of Compound 3 over 31 days versus a placebo. The dosages for Compound 3 in the trial were 20 mg for days1-3, 40 mg for days 4-6, 60 mg on day 7 for 7 days, and continuing to increase the dose by 20 mg every 7 days until reaching 120 mg. The clinical trial is further discussed in [Scott 2022],

[0090] FIG. 10 depicts the composite qEEG metric for the clinical trial. The vertical axis is log odds of each composite qEEG metric, and the horizontal axis is the number of days administering the dose. The dark lines are the average of the composite qEEG metric for each subject, and the light lines are the composite qEEG metric for each subject. According to the composite qEEG metric, the administered drug exhibited a continued effect over the days of the trial. Further, the composite qEEG metric correctly identified the test data for subjects administered the drug and the test data for subjects administered the placebo. The accuracy of the logistic regression of Compound 3 versus the placebo was 95.7%, and the fl score was 0.97.

[0091] FIGS. 11 A and 1 IB provide further data on the accuracy of the machine learning model. FIG. 11 A depicts a graph comparing the true positive rate and the false positive rate. For the logistic regression machine learning model, the auROC was 1.00. FIG. 1 IB depicts a graph comparing precision and recall. For the logistic regression machine learning model, the auPRC was 1.00.ILLUSTRATIVE APPARATUS

[0092] FIG. 12 depicts an example computer apparatus for use with the one or more embodiments described herein. As an example, the apparatus 1200 may be a computer to implement certain inventive techniques disclosed herein, such as implementing the inventive training, testing, and using the machine learning model to determine the effect of a drug based on brain activity. For example, the steps of FIG. 1 may be performed by a computer, such as computer apparatus 1200. The apparatus 1200 may include one or more processors1202, memory 1203, one or more input devices 1205, and one or more output devices 1206.The apparatus 1200 may be a cloud-based computer apparatus.

[0093] Input to the apparatus 1200 may be provided by one or more input devices 1205, provided from one or more input devices in communication with the apparatus 1200 via link 1201 (e.g., a wired link or a wireless link; e.g., with a direct connection or over a network), and / or provided from another computer(s) in communication with the apparatus 1200 via link 1201.

[0094] Output for the apparatus 1200 may be provided by one or more output devices 1206, provided to one or more output devices in communication with the apparatus 1200 via link 1201, and / or provided from another computer(s) in communication with the apparatus 1200 via link 1201. The one or more output devices 1206 may include one more displays and one or more speakers. The output device(s) 1206 may provide the output as discussed in step 110 of FIG. 1 according to one or more embodiments described herein.

[0095] The one or more output devices 1205 may provide the status of the operation of the invention. The output device(s) 1205 may provide visualization data or numeric data according to certain embodiments of the invention. The apparatus 1200 may provide the output over a network to another computer apparatus.

[0096] In some embodiments, one or more input devices 1205 and one or more output devices 1206 may be combined into one or more unitary input / output devices (e.g., a touch screen).

[0097] In some embodiments, based on input from one or more input devices 1205 or input from outside the apparatus 1200 via the link 1201, the one or more processors 1202 may perform operations as described herein. As an example, user input may be received from the one or more input devices 1205. As an example, input may be from another computer incommunication with the apparatus 1200 via link 1201. As an example, input may be from one or more input devices in communication with the apparatus 1200 via link 1201.

[0098] In some embodiments, the one or more processors 1202 may perform operations as described herein and provide results of the operations as output. As an example, output may be provided to the one or more output devices 1206. As an example, output may be provided to another computer in communication with the apparatus 1200 via link 1201. As an example, output may be provided to one or more output devices in communication with the apparatus 1200 via link 1201.

[0099] The memory 1203 may be accessible by the one or more processors 1202 (e.g., via a link 1204) so that the one or more processors 1202 can read information from and write information to the memory 1203. The memory 1203 may store instructions that, when executed by the one or more processors 1202, implement one or more embodiments described herein. The memory 1203 may be a non-transitory processor readable medium (or a non-transitory processor readable medium) containing a set of instructions thereon for determining an effect of a drug on a subject, wherein when executed by a processor (such as one or more processors 1202), the instructions cause the processor to perform one or more methods discussed herein.

[0100] The apparatus 1200 may be an apparatus for determining an effect of a drug on a subject, the apparatus including: one or more processors (such as one or more processors 1202); and memory (such as memory 1203) accessible by the one or more processors, the memory storing instructions that when executed by the one or more processors, cause the apparatus to perform one or more methods disclosed herein.

[0101] The memory 1203 may be a non-transitory processor readable medium containing a set of instructions thereon for determining an effect of a drug on a subject, wherein when executed by one or more processors (such as one or more processors 1202), theinstructions cause the one or more processors to perform one or more methods described herein.REFERENCES

[0102] [Abela et al 2018] Abela E, Pawley AD, Tangwiriyasakul C, et al. Slower alpha rhythm associates with poorer seizure control in epilepsy. Ann Clin Transl Neurol.2018;6(2):333-343. Published 2018 Dec 18. doi: 10.1002 / acn3.710.

[0103] [Archila-Melendez et al 2020] Archila-Melendez ME, Valente G, Gommer ED, Correia JM, ten Oever S, Peters JC, et al. Combining Gamma With Alpha and Beta Power Modulation for Enhanced Cortical Mapping in Patients With Focal Epilepsy. Front. Hum. Neurosci. 2020;14:555054. Published 2020 Dec 21. doi: 10.3389 / fnhum.2020.555054.

[0104] [Britton et al 2016] Britton JW, Frey LC, Hopp JL, et al., eds.Electroencephalography (EEG): An Introductory Text and Atlas of Normal and Abnormal Findings in Adults, Children, and Infants. Chicago: American Epilepsy Society; 2016.

[0105] [Mahalingam 2023] Rajeshwari Mahalingam, Michael Oldham, Corey Puryear, Prashant Bansal, Balaji Sriram, Dharit Patel, Henry Jacotin, Marjie Hard, Million Arefayene, Bernard Ravina, Marcio Souza, Steven Petrou, “PRAX-562-102: A Phase 1 Trial Evaluating the Safety, Tolerability, Pharmacokinetics and Pharmacodynamics of PRAX-562 in Healthy Volunteers (P4-9.011),” Neurology Apr 2023, 100 (17 Supplement 2) 3192; DOI: 10.1212 / WNL.0000000000203090.

[0106] [Praxis 2022] Praxis Precision Medicines Provides Corporate Update and Reports Second Quarter 2022 Financial Results, August 8, 2022, investors.praxismedicines.com / news-releases / news-release-details / praxis-precision- medicines-provides-corporate-update-and-5.

[0107] [Scott 2022] Scott L, Puryear CB, Belfort GM, et al., Translational Pharmacology of PRAX-944, a Novel T-Type Calcium Channel Blocker in Development for the Treatment of Essential Tremor, Movement Disorders, 2022 Jun; 37(6): 1193-1201. www.ncbi.nlm.nih.gov / pmc / articles / PMC9310641 / .

[0108] Embodiments illustrated under any heading or in any portion of the disclosure may be combined with embodiments illustrated under the same or any other heading or other portion of the disclosure unless otherwise indicated herein or otherwise clearly contradicted by context. For example, and without limitation, embodiments described in dependent claim format for a given embodiment (e.g., the given embodiment described in independent claim format) may be combined with other embodiments (described in independent claim format or dependent claim format).

[0109] Numerous modifications, alterations, and changes to the described embodiments are possible without departing from the scope of the present invention defined in the claims. It is intended that the present invention need not be limited to the described embodiments, but that it has the full scope defined by the language of the following claims, and equivalents thereof.

Claims

CLAIMSWhat is claimed is:

1. A computer-implemented method for determining an effect of a drug on a subject, the method comprising: obtaining brain activity data of the subject; determining, using a trained machine learning model and the brain activity data of the subject, an effect of the drug administered to the subject, the trained machine learning model trained with brain activity data of subjects treated with at least one of the drug or a second drug and with brain activity data of subjects treated with a placebo; and providing an output indicative of the effect of the drug administered to the subject determined by the trained machine learning model.

2. The computer-implemented method of claim 1, wherein the brain activity data of the subject is based on electroencephalography (EEG) data of the subject obtained from a plurality of electrodes attached to the head of the subject and processed using quantitative electroencephalography (qEEG).

3. The computer-implemented method of claim 1, wherein the brain activity data of the subject is based on band values for frequency bands of frequency-transformed electroencephalography (EEG) data of the subject separated into the frequency bands.

4. The computer-implemented method of claim 3, wherein the band values of the frequency bands are based on a plurality of time periods of the EEG data of the subject.

5. The computer-implemented method of claim 3, wherein the band values of the frequency bands are averages of band values for a plurality of electrodes attached to the head of the subject.

6. The computer-implemented method of claim 3, wherein the trained machine learning model outputs weights for a combination of the band values.

7. The computer-implemented method of claim 6, wherein the output indicative of the effect of the drug administered to the subject is based on a weighted combination of the band values using the weights for the band values output by the trained machine learning model.

8. The computer-implemented method of claim 7, wherein the weighted combination of the band values is processed using at least one of a negative log loss function, z-scoring, or subtracting of baseline values.

9. The computer-implemented method of claim 1, wherein the trained machine learning model is trained with brain activity data of subjects treated with the second drug and not trained with brain activity data of subjects treated with the drug.

10. The computer-implemented method of claim 1, wherein the trained machine learning model is trained with brain activity data of subjects treated with the drug and brain activity data of subjects treated with the second drug.

11. The computer-implemented method of claim 10, the trained machine learning model is further trained with brain activity data of subjects treated with a third drug.

12. The computer-implemented method of claim 1, wherein the drug administered to the subject is not one of the drugs used on subjects to obtain brain activity data to train the machine learning model.

13. The computer-implemented method of claim 1, wherein the drug administered to the subject is a drug to treat a brain-related disorder.

14. The computer-implemented method of claim 1, wherein the drug administered to the subject is a drug to treat at least one of epilepsy or a movement disorder.

15. The computer-implemented method of claim 1, wherein the drug administered to the subject is a single medicinal drug or a combination of medicinal drugs.

16. The computer-implemented method of claim 1, wherein the machine learning module is trained with one of logistic regression or binary classification logistic regression.

17. The computer-implemented method of claim 1, wherein the machine learning module is trained with one of linear regression, a random forest model, a gradient-boosted tree, a neural network, or a deep learning model.

18. The computer-implemented method of claim 1, further comprising training a machine learning model to obtain the trained machine learning model,wherein the machine learning model is trained with the brain activity data of subjects treated with at least one of the drug or another drug and with the brain activity data of subjects treated with a placebo.

19. A non-transitory processor readable medium containing a set of instructions thereon for determining an effect of a drug on a subject, wherein when executed by one or more processors, the instructions cause the one or more processors to perform a method comprising: obtaining brain activity data of the subject; determining, using a trained machine learning model and the brain activity data of the subject, an effect of the drug administered to the subject, the trained machine learning model trained with brain activity data of subjects treated with at least one of the drug or a second drug and with brain activity data of subjects treated with a placebo; and providing an output indicative of the effect of the drug administered to the subject determined by the trained machine learning model.

20. An apparatus for determining an effect of a drug on a subject, the apparatus comprising: one or more processors; and memory accessible by the one or more processors, the memory storing instructions that when executed by the one or more processors, cause the apparatus to perform a method comprising: obtaining brain activity data of the subject; determining, using a trained machine learning model and the brain activity data of the subject, an effect of the drug administered to the subject, the trained machine learning model trained with brain activity data of subjects treated with at least one of the drug or a second drug and with brain activity data of subjects treated with a placebo; andproviding an output indicative of the effect of the drug administered to the subject determined by the trained machine learning model.

Images