Leveraging GPT-4 for food effect summarization to enhance product-specific guidance development via iterative prompting.

Food effect summarization from New Drug Application (NDA) is an essential component of product-specific guidance (PSG) development and assessment, which provides the basis of recommendations for fasting and fed bioequivalence studies to guide the pharmaceutical industry for developing generic drug products. However, manual summarization of food effect from extensive drug application review documents is time-consuming. Therefore, there is a need to develop automated methods to generate food effect summary. Recent advances in natural language processing (NLP), particularly large language models (LLMs) such as ChatGPT and GPT-4, have demonstrated great potential in improving the effectiveness of automated text summarization, but its ability with regard to the accuracy in summarizing food effect for PSG assessment remains unclear. In this study, we introduce a simple yet effective approach,iterative prompting, which allows one to interact with ChatGPT or GPT-4 more effectively and efficiently through multi-turn interaction. Specifically, we propose a three-turn iterative prompting approach to food effect summarization in which the keyword-focused and length-controlled prompts are respectively provided in consecutive turns to refine the quality of the generated summary. We conduct a series of extensive evaluations, ranging from automated metrics to FDA professionals and even evaluation by GPT-4, on 100 NDA review documents selected over the past five years. We observe that the summary quality is progressively improved throughout the iterative prompting process. Moreover, we find that GPT-4 performs better than ChatGPT, as evaluated by FDA professionals (43% vs. 12%) and GPT-4 (64% vs. 35%). Importantly, all the FDA professionals unanimously rated that 85% of the summaries generated by GPT-4 are factually consistent with the golden reference summary, a finding further supported by GPT-4 rating of 72% consistency. Taken together, these results strongly suggest a great potential for GPT-4 to draft food effect summaries that could be reviewed by FDA professionals, thereby improving the efficiency of the PSG assessment cycle and promoting generic drug product development.

PharmBERT: a domain-specific BERT model for drug labels.

Human prescription drug labeling contains a summary of the essential scientific information needed for the safe and effective use of the drug and includes the Prescribing Information, FDA-approved patient labeling (Medication Guides, Patient Package Inserts and/or Instructions for Use), and/or carton and container labeling. Drug labeling contains critical information about drug products, such as pharmacokinetics and adverse events. Automatic information extraction from drug labels may facilitate finding the adverse reaction of the drugs or finding the interaction of one drug with another drug. Natural language processing (NLP) techniques, especially recently developed Bidirectional Encoder Representations from Transformers (BERT), have exhibited exceptional merits in text-based information extraction. A common paradigm in training BERT is to pretrain the model on large unlabeled generic language corpora, so that the model learns the distribution of the words in the language, and then fine-tune on a downstream task. In this paper, first, we show the uniqueness of language used in drug labels, which therefore cannot be optimally handled by other BERT models. Then, we present the developed PharmBERT, which is a BERT model specifically pretrained on the drug labels (publicly available at Hugging Face). We demonstrate that our model outperforms the vanilla BERT, ClinicalBERT and BioBERT in multiple NLP tasks in the drug label domain. Moreover, how the domain-specific pretraining has contributed to the superior performance of PharmBERT is demonstrated by analyzing different layers of PharmBERT, and more insight into how it understands different linguistic aspects of the data is gained.

Fine-tuning BERT for automatic ADME semantic labeling in FDA drug labeling to enhance product-specific guidance assessment.

Product-specific guidances (PSGs) recommended by the United States Food and Drug Administration (FDA) are instrumental to promote and guide generic drug product development. To assess a PSG, the FDA assessor needs to take extensive time and effort to manually retrieve supportive drug information of absorption, distribution, metabolism, and excretion (ADME) from the reference listed drug labeling. In this work, we leveraged the state-of-the-art pre-trained language models to automatically label the ADME paragraphs in the pharmacokinetics section from the FDA-approved drug labeling to facilitate PSG assessment. We applied a transfer learning approach by fine-tuning the pre-trained Bidirectional Encoder Representations from Transformers (BERT) model to develop a novel application of ADME semantic labeling, which can automatically retrieve ADME paragraphs from drug labeling instead of manual work. We demonstrate that fine-tuning the pre-trained BERT model can outperform conventional machine learning techniques, achieving up to 12.5% absolute F1 improvement. To our knowledge, we were the first to successfully apply BERT to solve the ADME semantic labeling task. We further assessed the relative contribution of pre-training and fine-tuning to the overall performance of the BERT model in the ADME semantic labeling task using a series of analysis methods, such as attention similarity and layer-based ablations. Our analysis revealed that the information learned via fine-tuning is focused on task-specific knowledge in the top layers of the BERT, whereas the benefit from the pre-trained BERT model is from the bottom layers.

Predicting dementia from spontaneous speech using large language models.

Language impairment is an important biomarker of neurodegenerative disorders such as Alzheimer's disease (AD). Artificial intelligence (AI), particularly natural language processing (NLP), has recently been increasingly used for early prediction of AD through speech. Yet, relatively few studies exist on using large language models, especially GPT-3, to aid in the early diagnosis of dementia. In this work, we show for the first time that GPT-3 can be utilized to predict dementia from spontaneous speech. Specifically, we leverage the vast semantic knowledge encoded in the GPT-3 model to generate text embedding, a vector representation of the transcribed text from speech, that captures the semantic meaning of the input. We demonstrate that the text embedding can be reliably used to (1) distinguish individuals with AD from healthy controls, and (2) infer the subject's cognitive testing score, both solely based on speech data. We further show that text embedding considerably outperforms the conventional acoustic feature-based approach and even performs competitively with prevailing fine-tuned models. Together, our results suggest that GPT-3 based text embedding is a viable approach for AD assessment directly from speech and has the potential to improve early diagnosis of dementia.

Artificial Intelligence-Enabled End-To-End Detection and Assessment of Alzheimer's Disease Using Voice.

There is currently no simple, widely available screening method for Alzheimer's disease (AD), partly because the diagnosis of AD is complex and typically involves expensive and sometimes invasive tests not commonly available outside highly specialized clinical settings. Here, we developed an artificial intelligence (AI)-powered end-to-end system to detect AD and predict its severity directly from voice recordings. At the core of our system is the pre-trained data2vec model, the first high-performance self-supervised algorithm that works for speech, vision, and text. Our model was internally evaluated on the ADReSSo (Alzheimer's Dementia Recognition through Spontaneous Speech only) dataset containing voice recordings of subjects describing the Cookie Theft picture, and externally validated on a test dataset from DementiaBank. The AI model can detect AD with average area under the curve (AUC) of 0.846 and 0.835 on held-out and external test set, respectively. The model was well-calibrated (Hosmer-Lemeshow goodness-of-fit p-value = 0.9616). Moreover, the model can reliably predict the subject's cognitive testing score solely based on raw voice recordings. Our study demonstrates the feasibility of using the AI-powered end-to-end model for early AD diagnosis and severity prediction directly based on voice, showing its potential for screening Alzheimer's disease in a community setting.

Information Extraction From FDA Drug Labeling to Enhance Product-Specific Guidance Assessment Using Natural Language Processing.

Towards the objectives of the UnitedStates Food and Drug Administration (FDA) generic drug science and research program, it is of vital importance in developing product-specific guidances (PSGs) with recommendations that can facilitate and guide generic product development. To generate a PSG, the assessor needs to retrieve supportive information about the drug product of interest, including from the drug labeling, which contain comprehensive information about drug products and instructions to physicians on how to use the products for treatment. Currently, although there are many drug labeling data resources, none of them including those developed by the FDA (e.g., Drugs@FDA) can cover all the FDA-approved drug products. Furthermore, these resources, housed in various locations, are often in forms that are not compatible or interoperable with each other. Therefore, there is a great demand for retrieving useful information from a large number of textual documents from different data resources to support an effective PSG development. To meet the needs, we developed a Natural Language Processing (NLP) pipeline by integrating multiple disparate publicly available data resources to extract drug product information with minimal human intervention. We provided a case study for identifying food effect information to illustrate how a machine learning model is employed to achieve accurate paragraph labeling. We showed that the pre-trained Bidirectional Encoder Representations from Transformers (BERT) model is able to outperform the traditional machine learning techniques, setting a new state-of-the-art for labelling food effect paragraphs from drug labeling and approved drug products datasets.

Virtual screening FDA approved drugs against multiple targets of SARS-CoV-2.

The outbreak of the novel coronavirus severe acute respiratory syndrome-coronavirus 2 (SARS-CoV-2), the causative agent of coronavirus disease 2019 (COVID-19) respiratory disease, led to a global pandemic with high morbidity and mortality. Despite frenzied efforts in therapeutic development, there are currently no effective drugs for treatment, nor are there vaccines for its prevention. Drug repurposing, representing as an effective drug discovery strategy from existing drugs, is one of the most practical treatment options against the outbreak. In this study, we present a novel strategy for in silico molecular modeling screening for potential drugs that may interact with multiple main proteins of SARS-CoV-2. Targeting multiple viral proteins is a novel drug discovery concept in that it enables the potential drugs to act on different stages of the virus' life cycle, thereby potentially maximizing the drug potency. We screened 2631 US Food and Drug Administration (FDA)-approved small molecules against 4 key proteins of SARS-CoV-2 that are known as attractive targets for antiviral drug development. In total, we identified 29 drugs that could actively interact with 2 or more target proteins, with 5 drugs (avapritinib, bictegravir, ziprasidone, capmatinib, and pexidartinib) being common candidates for all 4 key host proteins and 3 of them possessing the desirable molecular properties. By overlaying docked positions of drug candidates onto individual host proteins, it has been further confirmed that the binding site conformations are conserved. The drugs identified in our screening provide potential guidance for experimental confirmation, such as in vitro molecular assays and in vivo animal testing, as well as incorporation into ongoing clinical studies.

Hippocampal Ripple Coordinates Retrosplenial Inhibitory Neurons during Slow-Wave Sleep.

The hippocampus and retrosplenial cortex (RSC) play indispensable roles in memory formation, and importantly, a hippocampal oscillation known as ripple is key to consolidation of new memories. However, it remains unclear how the hippocampus and RSC communicate and the role of ripple oscillation in coordinating the activity between these two brain regions. Here, we record from the dorsal hippocampus and RSC simultaneously in freely behaving mice during sleep and reveal that the RSC displays a pre-ripple activation associated with slow and fast oscillations. Immediately after ripples, a subpopulation of RSC putative inhibitory neurons increases firing activity, while most RSC putative excitatory neurons decrease activity. Consistently, optogenetic stimulation of this hippocampus-RSC pathway activates and suppresses RSC putative inhibitory and excitatory neurons, respectively. These results suggest that the dorsal hippocampus mainly inhibits RSC activity via its direct innervation of RSC inhibitory neurons, which overshadows the RSC in supporting learning and memory functions.

Improved prediction of brain age using multimodal neuroimaging data.

Brain age prediction based on imaging data and machine learning (ML) methods has great potential to provide insights into the development of cognition and mental disorders. Though different ML models have been proposed, a systematic comparison of ML models in combination with imaging features derived from different modalities is still needed. In this study, we evaluate the prediction performance of 36 combinations of imaging features and ML models including deep learning. We utilize single and multimodal brain imaging data including MRI, DTI, and rs-fMRI from a large data set with 839 subjects. Our study is a follow-up to the initial work (Liang et al., 2019. Human Brain Mapping) to investigate different analytic strategies to combine data from MRI, DTI, and rs-fMRI with the goal to improve brain age prediction accuracy. Additionally, the traditional approach to predicting the brain age gap has been shown to have a systematic bias. The potential nonlinear relationship between the brain age gap and chronological age has not been thoroughly tested. Here we propose a new method to correct the systematic bias of brain age gap by taking gender, chronological age, and their interactions into consideration. As the true brain age is unknown and may deviate from chronological age, we further examine whether various levels of behavioral performance across subjects predict their brain age estimated from neuroimaging data. This is an important step to quantify the practical implication of brain age prediction. Our findings are helpful to advance the practice of optimizing different analytic methodologies in brain age prediction.

Highly Flexible Precisely Braided Multielectrode Probes and Combinatorics for Future Neuroprostheses.

The braided multielectrode probe (BMEP) is an ultrafine microwire bundle interwoven into a precise tubular braided structure, which is designed to be used as an invasive neural probe consisting of multiple microelectrodes for electrophysiological neural recording and stimulation. Significant advantages of BMEPs include highly flexible mechanical properties leading to decreased immune responses after chronic implantation in neural tissue and dense recording/stimulation sites (24 channels) within the 100-200 µm diameter. In addition, because BMEPs can be manufactured using various materials in any size and shape without length limitations, they could be expanded to applications in deep central nervous system (CNS) regions as well as peripheral nervous system (PNS) in larger animals and humans. Finally, the 3D topology of wires supports combinatoric rearrangements of wires within braids, and potential neural yield increases. With the newly developed next generation micro braiding machine, we can manufacture more precise and complex microbraid structures. In this article, we describe the new machine and methods, and tests of simulated combinatoric separation methods. We propose various promising BMEP designs and the potential modifications to these designs to create probes suitable for various applications for future neuroprostheses.

Spike-field Granger causality for hybrid neural data analysis.

Neurotechnological innovations allow for simultaneous recording at various scales, ranging from spiking activity of individual neurons to large neural populations' local field potentials (LFPs). This capability necessitates developing multiscale analysis of spike-field activity. A joint analysis of the hybrid neural data is crucial for bridging the scales between single neurons and local networks. Granger causality is a fundamental measure to evaluate directional influences among neural signals. However, it is mainly limited to inferring causal influence between the same type of signals-either LFPs or spike trains-and not well developed between two different signal types. Here we propose a model-free, nonparametric spike-field Granger causality measure for hybrid data analysis. Our measure is distinct from existing methods in that we use "binless" spikes (precise spike timing) rather than "binned" spikes (spike counts within small consecutive time windows). The latter clearly distort the information in the mixed analysis of spikes and LFP. Therefore, our spectral estimate of spike trains is directly applied to the neural point process itself, i.e., sequences of spike times rather than spike counts. Our measure is validated by an extensive set of simulated data. When the measure is applied to LFPs and spiking activity simultaneously recorded from visual areas V1 and V4 of monkeys performing a contour detection task, we are able to confirm computationally the long-standing experimental finding of the input-output relationship between LFPs and spikes. Importantly, we demonstrate that spike-field Granger causality can be used to reveal the modulatory effects that are inaccessible by traditional methods, such that spike→LFP Granger causality is modulated by the behavioral task, whereas LFP→spike Granger causality is mainly related to the average synaptic input.NEW & NOTEWORTHY It is a pressing question to study the directional interactions between local field potential (LFP) and spiking activity. In this report, we propose a model-free, nonparametric spike-field Granger causality measure that can be used to reveal directional influences between spikes and LFPs. This new measure is crucial for bridging the scales between single neurons and neural networks; hence it represents an important step to explicate how the brain orchestrates information processing.

Investigating systematic bias in brain age estimation with application to post-traumatic stress disorders.

Brain age prediction using machine-learning techniques has recently attracted growing attention, as it has the potential to serve as a biomarker for characterizing the typical brain development and neuropsychiatric disorders. Yet one long-standing problem is that the predicted brain age is overestimated in younger subjects and underestimated in older. There is a plethora of claims as to the bias origins, both methodologically and in data itself. With a large neuroanatomical dataset (N = 2,026; 6-89 years of age) from multiple shared datasets, we show this bias is neither data-dependent nor specific to particular method including deep neural network. We present an alternative account that offers a statistical explanation for the bias and describe a simple, yet efficient, method using general linear model to adjust the bias. We demonstrate the effectiveness of bias adjustment with a large multi-modal neuroimaging data (N = 804; 8-21 years of age) for both healthy controls and post-traumatic stress disorders patients obtained from the Philadelphia Neurodevelopmental Cohort.

A Copula-Based Granger Causality Measure for the Analysis of Neural Spike Train Data.

In systems neuroscience, it is becoming increasingly common to record the activity of hundreds of neurons simultaneously via electrode arrays. The ability to accurately measure the causal interactions among multiple neurons in the brain is crucial to understanding how neurons work in concert to generate specific brain functions. The development of new statistical methods for assessing causal influence between spike trains is still an active field of neuroscience research. Here, we suggest a copula-based Granger causality measure for the analysis of neural spike train data. This method is built upon our recent work on copula Granger causality for the analysis of continuous-valued time series by extending it to point-process neural spike train data. The proposed method is therefore able to reveal nonlinear and high-order causality in the spike trains while retaining all the computational advantages such as model-free, efficient estimation, and variability assessment of Granger causality. The performance of our algorithm can be further boosted with time-reversed data. Our method performed well on extensive simulations, and was then demonstrated on neural activity simultaneously recorded from primary visual cortex of a monkey performing a contour detection task.

Granger-Geweke causality: Estimation and interpretation.

In a recent PNAS article1, Stokes and Purdon performed numerical simulations to argue that Granger-Geweke causality (GGC) estimation is severely biased, or of high variance, and GGC application to neuroscience is problematic because the GGC measure is independent of 'receiver' dynamics. Here, we use the same simulation examples to show that GGC measures, when properly estimated either via the spectral factorization-enabled nonparametric approach or the VAR-model based parametric approach, do not have the claimed bias and high variance problems. Further, the receiver-independence property of GGC does not present a problem for neuroscience applications. When the nature and context of experimental measurements are taken into consideration, GGC, along with other spectral quantities, yield neurophysiologically interpretable results.

Synergistic Processing of Visual Contours across Cortical Layers in V1 and V2.

Visual cortical areas are interconnected via layer-specific feedforward and feedback projections. Such intricate connections are thought to be essential for parsing complex visual images, but the synergy among different layers in different cortical areas remains unclear. By simultaneously mapping neuronal activities across cortical depths in V1 and V2 of behaving monkeys, we identified spatiotemporally dissociable processes for grouping contour fragments and segregating background components. These processes generated and amplified contour signals within specific layers in V1 and V2. Contour-related inter-areal interactions, measured as Granger causality, were also dominant between these cortical layers within a time window when the contour signals were rapidly augmented. The grouping process became much faster for isolated contour elements compared with visual contours embedded in a complex background. Our results delineate the mode whereby image components are grouped and segmented through synergistic inter-laminar and inter-areal processes that are dynamically adjusted during interpretation of sensory inputs.

Interactions between feedback and lateral connections in the primary visual cortex.

Perceptual grouping of line segments into object contours has been thought to be mediated, in part, by long-range horizontal connectivity intrinsic to the primary visual cortex (V1), with a contribution by top-down feedback projections. To dissect the contributions of intraareal and interareal connections during contour integration, we applied conditional Granger causality analysis to assess directional influences among neural signals simultaneously recorded from visual cortical areas V1 and V4 of monkeys performing a contour detection task. Our results showed that discounting the influences from V4 markedly reduced V1 lateral interactions, indicating dependence on feedback signals of the effective connectivity within V1. On the other hand, the feedback influences were reciprocally dependent on V1 lateral interactions because the modulation strengths from V4 to V1 were greatly reduced after discounting the influences from other V1 neurons. Our findings suggest that feedback and lateral connections closely interact to mediate image grouping and segmentation.

Structure-Function Network Mapping and Its Assessment via Persistent Homology.

Understanding the relationship between brain structure and function is a fundamental problem in network neuroscience. This work deals with the general method of structure-function mapping at the whole-brain level. We formulate the problem as a topological mapping of structure-function connectivity via matrix function, and find a stable solution by exploiting a regularization procedure to cope with large matrices. We introduce a novel measure of network similarity based on persistent homology for assessing the quality of the network mapping, which enables a detailed comparison of network topological changes across all possible thresholds, rather than just at a single, arbitrary threshold that may not be optimal. We demonstrate that our approach can uncover the direct and indirect structural paths for predicting functional connectivity, and our network similarity measure outperforms other currently available methods. We systematically validate our approach with (1) a comparison of regularized vs. non-regularized procedures, (2) a null model of the degree-preserving random rewired structural matrix, (3) different network types (binary vs. weighted matrices), and (4) different brain parcellation schemes (low vs. high resolutions). Finally, we evaluate the scalability of our method with relatively large matrices (2514x2514) of structural and functional connectivity obtained from 12 healthy human subjects measured non-invasively while at rest. Our results reveal a nonlinear structure-function relationship, suggesting that the resting-state functional connectivity depends on direct structural connections, as well as relatively parsimonious indirect connections via polysynaptic pathways.

Data on copula modeling of mixed discrete and continuous neural time series.

Copula is an important tool for modeling neural dependence. Recent work on copula has been expanded to jointly model mixed time series in neuroscience ("Hu et al., 2016, Joint Analysis of Spikes and Local Field Potentials using Copula" [1]). Here we present further data for joint analysis of spike and local field potential (LFP) with copula modeling. In particular, the details of different model orders and the influence of possible spike contamination in LFP data from the same and different electrode recordings are presented. To further facilitate the use of our copula model for the analysis of mixed data, we provide the Matlab codes, together with example data.

Joint analysis of spikes and local field potentials using copula.

Recent technological advances, which allow for simultaneous recording of spikes and local field potentials (LFPs) at multiple sites in a given cortical area or across different areas, have greatly increased our understanding of signal processing in brain circuits. Joint analysis of simultaneously collected spike and LFP signals is an important step to explicate how the brain orchestrates information processing. In this contribution, we present a novel statistical framework based on Gaussian copula to jointly model spikes and LFP. In our approach, we use copula to link separate, marginal regression models to construct a joint regression model, in which the binary-valued spike train data are modeled using generalized linear model (GLM) and the continuous-valued LFP data are modeled using linear regression. Model parameters can be efficiently estimated via maximum-likelihood. In particular, we show that our model offers a means to statistically detect directional influence between spikes and LFP, akin to Granger causality measure, and that we are able to assess its statistical significance by conducting a Wald test. Through extensive simulations, we also show that our method is able to reliably recover the true model used to generate the data. To demonstrate the effectiveness of our approach in real setting, we further apply the method to a mixed neural dataset, consisting of spikes and LFP simultaneously recorded from the visual cortex of a monkey performing a contour detection task.

Copula regression analysis of simultaneously recorded frontal eye field and inferotemporal spiking activity during object-based working memory.

Inferotemporal (IT) neurons are known to exhibit persistent, stimulus-selective activity during the delay period of object-based working memory tasks. Frontal eye field (FEF) neurons show robust, spatially selective delay period activity during memory-guided saccade tasks. We present a copula regression paradigm to examine neural interaction of these two types of signals between areas IT and FEF of the monkey during a working memory task. This paradigm is based on copula models that can account for both marginal distribution over spiking activity of individual neurons within each area and joint distribution over ensemble activity of neurons between areas. Considering the popular GLMs as marginal models, we developed a general and flexible likelihood framework that uses the copula to integrate separate GLMs into a joint regression analysis. Such joint analysis essentially leads to a multivariate analog of the marginal GLM theory and hence efficient model estimation. In addition, we show that Granger causality between spike trains can be readily assessed via the likelihood ratio statistic. The performance of this method is validated by extensive simulations, and compared favorably to the widely used GLMs. When applied to spiking activity of simultaneously recorded FEF and IT neurons during working memory task, we observed significant Granger causality influence from FEF to IT, but not in the opposite direction, suggesting the role of the FEF in the selection and retention of visual information during working memory. The copula model has the potential to provide unique neurophysiological insights about network properties of the brain.