Predicting blood-brain barrier permeability of molecules with a large language model and machine learning.

Predicting the blood-brain barrier (BBB) permeability of small-molecule compounds using a novel artificial intelligence platform is necessary for drug discovery. Machine learning and a large language model on artificial intelligence (AI) tools improve the accuracy and shorten the time for new drug development. The primary goal of this research is to develop artificial intelligence (AI) computing models and novel deep learning architectures capable of predicting whether molecules can permeate the human blood-brain barrier (BBB). The in silico (computational) and in vitro (experimental) results were validated by the Natural Products Research Laboratories (NPRL) at China Medical University Hospital (CMUH). The transformer-based MegaMolBART was used as the simplified molecular input line entry system (SMILES) encoder with an XGBoost classifier as an in silico method to check if a molecule could cross through the BBB. We used Morgan or Circular fingerprints to apply the Morgan algorithm to a set of atomic invariants as a baseline encoder also with an XGBoost classifier to compare the results. BBB permeability was assessed in vitro using three-dimensional (3D) human BBB spheroids (human brain microvascular endothelial cells, brain vascular pericytes, and astrocytes). Using multiple BBB databases, the results of the final in silico transformer and XGBoost model achieved an area under the receiver operating characteristic curve of 0.88 on the held-out test dataset. Temozolomide (TMZ) and 21 randomly selected BBB permeable compounds (Pred scores = 1, indicating BBB-permeable) from the NPRL penetrated human BBB spheroid cells. No evidence suggests that ferulic acid or five BBB-impermeable compounds (Pred scores < 1.29423E-05, which designate compounds that pass through the human BBB) can pass through the spheroid cells of the BBB. Our validation of in vitro experiments indicated that the in silico prediction of small-molecule permeation in the BBB model is accurate. Transformer-based models like MegaMolBART, leveraging the SMILES representations of molecules, show great promise for applications in new drug discovery. These models have the potential to accelerate the development of novel targeted treatments for disorders of the central nervous system.

A large language model for electronic health records.

There is an increasing interest in developing artificial intelligence (AI) systems to process and interpret electronic health records (EHRs). Natural language processing (NLP) powered by pretrained language models is the key technology for medical AI systems utilizing clinical narratives. However, there are few clinical language models, the largest of which trained in the clinical domain is comparatively small at 110 million parameters (compared with billions of parameters in the general domain). It is not clear how large clinical language models with billions of parameters can help medical AI systems utilize unstructured EHRs. In this study, we develop from scratch a large clinical language model-GatorTron-using >90 billion words of text (including >82 billion words of de-identified clinical text) and systematically evaluate it on five clinical NLP tasks including clinical concept extraction, medical relation extraction, semantic textual similarity, natural language inference (NLI), and medical question answering (MQA). We examine how (1) scaling up the number of parameters and (2) scaling up the size of the training data could benefit these NLP tasks. GatorTron models scale up the clinical language model from 110 million to 8.9 billion parameters and improve five clinical NLP tasks (e.g., 9.6% and 9.5% improvement in accuracy for NLI and MQA), which can be applied to medical AI systems to improve healthcare delivery. The GatorTron models are publicly available at: https://catalog.ngc.nvidia.com/orgs/nvidia/teams/clara/models/gatortron_og .

Sparsity and Biomechanics Inspired Integration of Shape and Speckle Tracking for Cardiac Deformation Analysis.

Cardiac motion analysis, particularly of the left ventricle (LV), can provide valuable information regarding the functional state of the heart. We propose a strategy of combining shape tracking and speckle tracking based displacements to calculate the dense deformation field of the myocardium. We introduce the use and effects of l1 regularization, which induces sparsity, in our integration method. We also introduce regularization to make the dense fields more adhering to cardiac biomechanics. Finally, we motivate the necessity of temporal coherence in the dense fields and demonstrate a way of doing so. We test our method on ultrasound (US) images acquired from six open-chested canine hearts. Baseline and post-occlusion strain results are presented for an animal, where we were able to detect significant change in the ischemic region. Six sets of strain results were also compared to strains obtained from tagged magnetic resonance (MR) data. Median correlation (with MR-tagging) coefficients of 0.73 and 0.82 were obtained for radial and circumferential strains respectively.

Radial basis functions for combining shape and speckle tracking in 4D echocardiography.

Quantitative analysis of left ventricular deformation can provide valuable information about the extent of disease as well as the efficacy of treatment. In this work, we develop an adaptive multi-level compactly supported radial basis approach for deformation analysis in 3D+time echocardiography. Our method combines displacement information from shape tracking of myocardial boundaries (derived from B-mode data) with mid-wall displacements from radio-frequency-based ultrasound speckle tracking. We evaluate our methods on open-chest canines (N=8) and show that our combined approach is better correlated to magnetic resonance tagging-derived strains than either individual method. We also are able to identify regions of myocardial infarction (confirmed by postmortem analysis) using radial strain values obtained with our approach.

Contour tracking in echocardiographic sequences via sparse representation and dictionary learning.

This paper presents a dynamical appearance model based on sparse representation and dictionary learning for tracking both endocardial and epicardial contours of the left ventricle in echocardiographic sequences. Instead of learning offline spatiotemporal priors from databases, we exploit the inherent spatiotemporal coherence of individual data to constraint cardiac contour estimation. The contour tracker is initialized with a manual tracing of the first frame. It employs multiscale sparse representation of local image appearance and learns online multiscale appearance dictionaries in a boosting framework as the image sequence is segmented frame-by-frame sequentially. The weights of multiscale appearance dictionaries are optimized automatically. Our region-based level set segmentation integrates a spectrum of complementary multilevel information including intensity, multiscale local appearance, and dynamical shape prediction. The approach is validated on twenty-six 4D canine echocardiographic images acquired from both healthy and post-infarct canines. The segmentation results agree well with expert manual tracings. The ejection fraction estimates also show good agreement with manual results. Advantages of our approach are demonstrated by comparisons with a conventional pure intensity model, a registration-based contour tracker, and a state-of-the-art database-dependent offline dynamical shape model. We also demonstrate the feasibility of clinical application by applying the method to four 4D human data sets.

Mining echocardiography workflows for disease discriminative patterns.

OBJECTIVE: To provide quick diagnostic insights to medical practitioners into echocardiograms by only analyzing the echocardiogram workflows (defined as the sequence of modalities examined). METHODS: We define a dictionary of workflows, called subflows, which are commonly encountered in echocardiography workflows but are mutually exclusive. We represent each workflow as a mixture of dictionary subflows and learn discriminative models for various cardiac diseases using Support Vector Machines. Using these discriminative models, we can predict occurrences of diseases for any, yet unseen, echocardiogram workflow. RESULTS: Working with a corpus of 2300 echocardiograms workflows, we build a dictionary of 172 subflows. Using the associated reports (expert created) we identify the ground-truth diagnoses. We then build discriminative models for 7 different cardiac diseases. Using just the workflow as input, these models can predict diseases on average with over 75% accuracy. CONCLUSIONS: Mining collection of echocardiography workflows, for the first time, we are able to predict diseases without even looking at the image contents.

A dynamical appearance model based on multiscale sparse representation: segmentation of the left ventricle from 4D echocardiography.

The spatio-temporal coherence in data plays an important role in echocardiographic segmentation. While learning offline dynamical priors from databases has received considerable attention, these priors may not be suitable for post-infarct patients and children with congenital heart disease. This paper presents a dynamical appearance model (DAM) driven by individual inherent data coherence. It employs multi-scale sparse representation of local appearance, learns online multiscale appearance dictionaries as the image sequence is segmented sequentially, and integrates a spectrum of complementary multiscale appearance information including intensity, multiscale local appearance, and dynamical shape predictions. It overcomes the limitations of database-driven statistical models and applies to a broader range of subjects. Results on 26 4D canine echocardiographic images acquired from both healthy and post-infarct subjects show that our method significantly improves segmentation accuracy and robustness compared to a conventional intensity model and our previous single-scale sparse representation method.