Accurate diagnosis of atopic dermatitis by combining transcriptome and microbiota data with supervised machine learning

Atopic dermatitis (AD) is a common skin disease in childhood whose diagnosis requires expertise in dermatology. Recent studies have indicated that host genes–microbial interactions in the gut contribute to human diseases including AD. We sought to develop an accurate and automated pipeline for AD diagnosis based on transcriptome and microbiota data. Using these data of 161 subjects including AD patients and healthy controls, we trained a machine learning classifier to predict the risk of AD. We found that the classifier could accurately differentiate subjects with AD and healthy individuals based on the omics data with an average F1-score of 0.84. With this classifier, we also identified a set of 35 genes and 50 microbiota features that are predictive for AD. Among the selected features, we discovered at least three genes and three microorganisms directly or indirectly associated with AD. Although further replications in other cohorts are needed, our findings suggest that these genes and microbiota features may provide novel biological insights and may be developed into useful biomarkers of AD prediction.

Automated Generation of Core Outcome Measures for Clinical Trials

Common outcome sets are vital for ensuring usability of clinical trial results and enabling inter-study comparisons. The task of identifying clinical outcomes for a particular field is cumbersome and time-consuming. The aim of this work was to develop an automated pipeline for identifying common outcomes by analyzing outcomes from relevant trials reported at ClinicalTrials.gov and to assess the pipeline accuracy. We validated the output of our pipeline by comparing the outcomes it identified for acute coronary syndromes and coronary artery disease with the set of outcomes recommended for these conditions by a panel of experts in a widely cited report. We found that our pipeline identified the same or similar outcomes for 100% of the outcomes recommended in the experts’ report. The coverage of the pipeline’s results dropped only slightly (to 21 out of 23 outcome domains, 91%) when we restricted the pipeline to trials posted before the publication of the report, indicating a great potential for this pipeline to be used in aiding and informing the future development of core outcome measures in clinical trials.

A Deep Learning Method for Fully Automatic Stomatal Morphometry and Maximal Conductance Estimation

Stomata are integral to plant performance, enabling the exchange of gases between the atmosphere and the plant. The anatomy of stomata influences conductance properties with the maximal conductance rate, gsmax, calculated from density and size. However, current calculations of stomatal dimensions are performed manually, which are time-consuming and error prone. Here, we show how automated morphometry from leaf impressions can predict a functional property: the anatomical gsmax. A deep learning network was derived to preserve stomatal morphometry via semantic segmentation. This forms part of an automated pipeline to measure stomata traits for the estimation of anatomical gsmax. The proposed pipeline achieves accuracy of 100% for the distinction (wheat vs. poplar) and detection of stomata in both datasets. The automated deep learning-based method gave estimates for gsmax within 3.8 and 1.9% of those values manually calculated from an expert for a wheat and poplar dataset, respectively. Semantic segmentation provides a rapid and repeatable method for the estimation of anatomical gsmax from microscopic images of leaf impressions. This advanced method provides a step toward reducing the bottleneck associated with plant phenotyping approaches and will provide a rapid method to assess gas fluxes in plants based on stomata morphometry.

findMySequence: a neural-network-based approach for identification of unknown proteins in X-ray crystallography and cryo-EM

Although experimental protein-structure determination usually targets known proteins, chains of unknown sequence are often encountered. They can be purified from natural sources, appear as an unexpected fragment of a well characterized protein or appear as a contaminant. Regardless of the source of the problem, the unknown protein always requires characterization. Here, an automated pipeline is presented for the identification of protein sequences from cryo-EM reconstructions and crystallographic data. The method's application to characterize the crystal structure of an unknown protein purified from a snake venom is presented. It is also shown that the approach can be successfully applied to the identification of protein sequences and validation of sequence assignments in cryo-EM protein structures.

A Toolkit for Analyzing Strategy Use in Phonemic Verbal Fluency Tasks preprint

The phonemic verbal fluency task is a common cognitive assessment of language and executive functioning which asks participants to list as many words as they can that begin with a given letter. Verbal fluency tasks are widely used to identify deficits in verbal fluency, which have been associated with disorders such as schizophrenia and dementia. Verbal fluency tasks are scored by the number of correct responses, however analysis of “clusters” of related words within a response list can give insights into the cognitive strategies used by participants. Unfortunately, manual word cluster analysis is time and labor intensive and inconsistent, since raters may cluster words differently depending on how they themselves have phonetically categorized the words. We present an automated pipeline for quantification of strategy use in the phonemic verbal fluency task, “LetterVF”. LetterVF is a python module (i.e., a script containing useful functions, which can be imported and used in other scripts) that uses a pronunciation dictionary to convert verbal fluency task data items into lists of phonemes, which can be analyzed to identify clusters of words that share similarities in any of several clustering categories. Additionally, LetterVF contains useful functions for identifying intrusions (words which do not follow the rules for the task), identifying perseverations (responses repeated within the same trial), counting the number of cluster switches in a list, and calculating the average size of clusters for a list. Analysis of data from 50 participants’ verbal fluency task responses indicated that analysis using LetterVF yields accuracy and consistency on par with manual analysis. Our hope is that this tool will allow researchers to get more out of their datasets, and explore new topics related to cognitive strategy use, such as how strategies change with age and differences in strategies between experimental groups.

DiagnosisQA: A semi-automated pipeline for developing clinician validated diagnosis specific QA datasets.

Question answering (QA) is one of the oldest research areas of AI and Compu- national Linguistics. QA has seen significant progress with the development of state-of-the-art models and benchmark datasets over the last few years. However, pre-trained QA models perform poorly for clinical QA tasks, presumably due to the complexity of electronic healthcare data. With the digitization of healthcare data and the increasing volume of unstructured data, it is extremely important for healthcare providers to have a mechanism to query the data to find appropriate answers. Since diagnosis is central to any decision-making for the clinicians and patients, we have created a pipeline to develop diagnosis-specific QA datasets and curated a QA database for the Cerebrovascular Accident (CVA). CVA, also commonly known as Stroke, is an important and commonly occurring diagnosis amongst critically ill patients. Our method when compared to clinician validation achieved an accuracy of 0.90(with 90% CI [0.82,0.99]). Using our method, we hope to overcome the key challenges of building and validating a highly accurate QA dataset in a semiautomated manner which can help improve performance of QA models.

VALIS: Virtual Alignment of pathoLogy Image Series

Spatial analyses can reveal important interactions between and among cells and their microenvironment. However, most existing staining methods are limited to a handful of markers per slice, thereby limiting the number of interactions that can be studied. This limitation is frequently overcome by registering multiple images to create a single composite image containing many markers. While there are several existing image registration methods for whole slide images (WSI), most have specific use cases. Here, we present the Virtual Alignment of pathoLogy Image Series (VALIS), a fully automated pipeline that opens, registers (rigid and/or non-rigid), and saves aligned slides in the ome.tiff format. VALIS has been tested with 273 immunohistochemistry (IHC) samples and 340 immunofluorescence (IF) samples, each of which contained between 2-69 images per sample. The registered WSI tend to have low error and are completed within a matter of minutes. In addition to registering slides, VALIS can also using the registration parameters to warp point data, such as cell centroids previously determined via cell segmentation and phenotyping. VALIS is written in Python and requires only few lines of code for execution. VALIS therefore provides a free, opensource, flexible, and simple pipeline for rigid and non-rigid registration of IF and/or IHC that can facilitate spatial analyses of WSI from novel and existing datasets.