Machine Learning Methods for Endocrine Disrupting Potential Identification Based on Single-Cell Data

Humans are continuously exposed to a variety of toxicants and chemicals which is exacerbated during and after environmental catastrophes such as floods, earthquakes, and hurricanes. The hazardous chemical mixtures generated during these events threaten the health and safety of humans and other living organisms. This necessitates the development of rapid decision-making tools to facilitate mitigating the adverse effects of exposure on the key modulators of the endocrine system, such as the estrogen receptor alpha (ERα), for example. The mechanistic stages of the estrogenic transcriptional activity can be measured with high content/high throughput microscopy-based biosensor assays at the single-cell level, which generates millions of object-based minable data points. By combining computational modeling and experimental analysis, we built a highly accurate data-driven classification framework to assess the endocrine disrupting potential of environmental compounds. The effects of these compounds on the ERα pathway are predicted as being receptor agonists or antagonists using the principal component analysis (PCA) projections of high throughput, high content image analysis descriptors. The framework also combines rigorous preprocessing steps and nonlinear machine learning algorithms, such as the Support Vector Machines and Random Forest classifiers, to develop highly accurate mathematical representations of the separation between ERα agonists and antagonists. The results show that Support Vector Machines classify the unseen chemicals correctly with more than 96% accuracy using the proposed framework, where the preprocessing and the PCA steps play a key role in suppressing experimental noise and unraveling hidden patterns in the dataset.

Normalization and possibility of classification analysis using the optimal warping paths of dynamic time warping in gait analysis

The purpose of this study was to verify classification performance and the difference analysis between gender using optimal warping paths of dynamic time warping (DTW) and to examine the usefulness of root mean square error (RMSE) represented by the perpendicular distance from the optimal warping path to the diagonal. A 3-dimensional motion analysis experiment was performed with 24 healthy adults (male=12, female=12) in their 20s of age without gait-related diseases or injuries for the past 6 months to collect gait data. This study performed a DTW 132 times in total (male=62, female=62) for the flexion angle of the right leg's hip, knee, and ankle joints. Then, the global cost and the RMSE of the optimal warping paths were calculated and normalized. The difference analysis was performed by independent t-test. Machine learning was performed to test the classification performance using the neural network, support vector machine, and logistic regression model among the supervised models. Results analyzed using global cost and RMSE for hip, knee, and ankle joints showed a statistically significant difference between genders in global cost and RMSE for hip and knee joints but not for ankle joints using RMSE. Considering both area under the receiver operating characteristic curve and F1-score, the logistic regression model has been evaluated as the most suitable for gender classification using the global cost or RMSE. This study demonstrated that optimal warping paths could be used for statistical difference analysis and classification analysis.

Chemometrics analysis for the detection of dental caries via UV absorption spectroscopy

Caries is one of the non-communicable diseases that has a high prevalence trend. The current methods used to detect caries require sophisticated laboratory equipment, professional inspection, and expensive equipment such as X-ray imaging device. A non-invasive and economical method is required to substitute the conventional methods for the detection of caries. UV absorption spectroscopy coupled with chemometrics analysis has emerged as a good potential candidate for such an application. Data preprocessing methods such as mean centre, autoscale and Savitzky-Golay smoothing were implemented to enhance the signal-to-noise ratio of spectra data. Various classification algorithms namely K-nearest neighbours (KNN), logistic regression (LR) and linear discriminant analysis (LDA) were implemented to classify the severity of dental caries into International Caries Detection and Assessment System (ICDAS) scores. The performance of the prediction model was measured and comparatively analysed based on the accuracy, precision, sensitivity, and specificity. The LDA algorithm combined with the Savitzky-Golay preprocessing method had shown the best result with respect to the validation data accuracy, precision, sensitivity and specificity, where each had values of 0.90, 1.00, 0.86 and 1.00 respectively. The area under the curve of the ROC plot computed for the LDA algorithm was 0.95, which indicated that the prediction algorithm was capable of differentiating normal and caries teeth excellently.

Interpretability of selected variables and performance comparison of variable selection methods in a polyethylene and polypropylene NIR classification task

Near infrared (NIR) spectra are collected as a high amount of absorption values which usually greatly exceeds the sample size. Variable selection methods are employed in NIR spectroscopy to avoid "curse of dimensionality" related issues. In this paper, we examined the interpretability of selected variables, that is, how much selected wavelengths are related to the chemical structure of the materials studied, and if the relation is important for classification performance. Additionally, we examined classification performance in dependence on the number of selected variables. For this purpose, relative standard deviation (RSD), successive projection algorithm (SPA), stepwise decorrelation of variables (SELECT), genetic algorithm (GA), principal component analysis (PCA), and random (RANDOM) variable selection were applied in two-class classification modelling using linear discriminant analysis (LDA) or a support vector machine (SVM). Different pre-treatments and sample sizes were considered. Variable selection improved classification performance and variables selected by a majority of the methods considered were conveniently related to chemical structure. Interpretability and performance increase/decrease depend greatly on the number of selected variables, however. Since selected variables reveal great chemical interpretability, some variable selection methods could be employed to determine material characteristic absorption bands. SELECT and SPA displayed the best properties among the methods considered. To avoid faulty results, optimization of the number of selected variables should become the crucial stage in the variable selection process.

Identification of active and inactive agonists/antagonists of estrogen receptor based on Tox21 10K compound library: Binomial analysis and structure alert

Endocrine disrupting chemicals can mimic, block, or interfere with hormones in organisms and subsequently affect their development and reproduction, which has raised significant public concern over the past several decades. To investigate (quantitative) structure-activity relationship, 8280 compounds were compiled from the Tox21 10K compound library. The results show that 50% activity concentrations of agonists are poorly related to that of antagonists because many compounds have considerably different activity concentrations between the agonists and antagonists. Analysis on the chemical classes based on mode of action (MOA) reveals that estrogen receptor (ER) is not the main target site in the acute toxicity to aquatic organisms. Binomial analysis of active and inactive ER agonists/antagonists reveals that ER activity of compounds is dominated by octanol/water partition coefficient and excess molar refraction. The binomial equation developed from the two descriptors can classify well active and inactive ER chemicals with an overall prediction accuracy of 73%. The classification equation developed from the molecular descriptors indicates that estrogens react with the receptor through hydrophobic and π-n electron interactions. At the same time, molecular ionization, polarity, and hydrogen bonding ability can also affect the chemical ER activity. A decision tree developed from chemical structures and their applications reveals that many hormones, proton pump inhibitors, PAHs, progestin, insecticides, fungicides, steroid and chemotherapy medications are active ER agonists/antagonists. On the other hand, many monocyclic/nonaromatic chain compounds and herbicides are inactive ER compounds. The decision tree and binomial equation developed here are valuable tools to predict active and inactive ER compounds.

Normal appearing brain white matter changes in relapsing multiple sclerosis: Texture image and classification analysis in serial MRI scans

OBJECTIVE

There is a clinical interest in identifying normal appearing white matter (NAWM) areas in brain T2-weighted (T₂W) MRI scans in multiple sclerosis (MS) subjects. These areas are susceptible to disease development and areas need to be studied in order to find potential associations between texture feature changes and disease progression.

METHODS

The subjects investigated had a first demyelinating event (Clinically Isolated Syndrome-CIS) at baseline (Time₀), and the NAWM₀ (i.e. NAWM at Time₀) of the brain tissue was subsequently converted to demyelinating plaques (as evaluated in a follow up MRI at Time_6-12). 38 untreated subjects that had developed a CIS, had brain MRI scans within an interval of 6-12 months (Time_6-12 at follow-up). An experienced MS neurologist manually delineated the demyelinating lesions at Time₀ (L₀) and at Time_6-12 (L_6-12). Areas in the Time_6-12 MRI scans, where new lesions had been developed, were mapped back to their corresponding NAWM areas on the Time₀ MR scans (ROIS₀). In addition, contralateral ROIs of similar size and shape were segmented on the same images at Time₀ (ROIS_C0) to form an intra-subject control group. Following that, texture features were extracted from all prescribed areas and MS lesions.

RESULTS

Texture features were used as input into Support Vector Machine (SVM) models to differentiate between the following: NAWM₀ vs ROIS_C0, NAWM₀ vs NAWM_6-12, NAWM₀ vs L₀, NAWM_6-12 vs L_6-12, ROIS₀ vs L₀, ROIS₀ vs L_6-12 and ROIS₀ vs ROIS_C0, where the corresponding % correct classifications scores were 89%, 95%, 98%, 92%, 85%, 90% and 65% respectively.

CONCLUSIONS

Texture features may provide complementary information for following up the development and progression of MS disease. Future work will investigate the proposed method on more subjects.

Classification of ischemic and non-ischemic cardiac events in Holter recordings based on the continuous wavelet transform

Holter recordings are widely used to detect cardiac events that occur transiently, such as ischemic events. Much effort has been made to detect early ischemia, thus preventing myocardial infarction. However, after detection, classification of ischemia has still not been fully solved. The main difficulty relies on the false positives produced because of non-ischemic events, such as changes in the heart rate, the intraventricular conduction or the cardiac electrical axis. In this work, the classification of ischemic and non-ischemic events from the long-term ST database has been improved, using novel spectral parameters based on the continuous wavelet transform (CWT) together with temporal parameters (such as ST level and slope, T wave width and peak, R wave peak, QRS complex width). This was achieved by using a nearest neighbour classifier of six neighbours. Results indicated a sensitivity and specificity of 84.1% and 92.9% between ischemic and non-ischemic events, respectively, resulting a 10% increase of the sensitivity found in the literature. Extracted features based on the CWT applied on the ECG in the frequency band 0.5-4 Hz provided a substantial improvement in classifying ischemic and non-ischemic events, when comparing with the same classifier using only temporal parameters. Graphical Abstract In this work it is improved the classification of ischemic and non-ischemic events. The main difficulty of ischemic detectors relies on the false positives produced because of non-ischemic events. After a preprocessing stage, temporal and spectral parameters are extracted from events of the Long Term ST Database. The novel parameters proposed in this work are extracted from the Continuous Wavelet Transform. A nearest Neighbor Classifier is used, obtaining a sensitivity and specificity of 84.1% and 92.9%, respectively.

A comparison of statistical and machine-learning techniques in evaluating the association between dietary patterns and 10-year cardiometabolic risk (2002-2012): the ATTICA study

Statistical methods are usually applied in examining diet-disease associations, whereas factor analysis is commonly used for dietary pattern recognition. Recently, machine learning (ML) has been also proposed as an alternative technique in health classification. In this work, the predictive accuracy of statistical v. ML methodologies as regards the association of dietary patterns on CVD risk was tested. During 2001-2002, 3042 men and women (45 (sd 14) years) were enrolled in the ATTICA study. In 2011-2012, the 10-year CVD follow-up was performed among 2020 participants. Item Response Theory was applied to create a metric of combined 10-year cardiometabolic risk, the 'Cardiometabolic Health Score', that incorporated incidence of CVD, diabetes, hypertension and hypercholesterolaemia. Factor analysis was performed to extract dietary patterns, on the basis of either foods or nutrients consumed; linear regression analysis was used to assess their association with the cardiometabolic score. Two ML techniques (k-nearest-neighbor's algorithm and random-forests decision tree) were applied to evaluate participants' health based on dietary information. Factor analysis revealed five and three factors from foods and nutrients, respectively, explaining 54 and 65 % of the total variation in intake. Nutrient and food pattern regression models showed similar accuracy in correctly classifying an individual according to the cardiometabolic risk (R 2=9·6 % and R 2=8·3 %, respectively). ML techniques were superior compared with linear regression in correct classification of the individuals according to the Health Score (accuracy approximately 38 v. 6 %, respectively), whereas the two ML methods showed equal classification ability. Conclusively, ML methods could be a valuable tool in the field of nutritional epidemiology, leading to more accurate disease-risk evaluation.

Decoding fMRI activity in the time domain improves classification performance

Most current functional Magnetic Resonance Imaging (fMRI) decoding analyses rely on statistical summaries of the data resulting from a deconvolution approach: each stimulation event is associated with a brain response. This standard approach leads to simple learning procedures, yet it is ill-suited for decoding events with short inter-stimulus intervals. In order to overcome this issue, we propose a novel framework that separates the spatial and temporal components of the prediction by decoding the fMRI time-series continuously, i.e. scan-by-scan. The stimulation events can then be identified through a deconvolution of the reconstructed time series. We show that this model performs as well as or better than standard approaches across several datasets, most notably in regimes with small inter-stimuli intervals (3-5s), while also offering predictions that are highly interpretable in the time domain. This opens the way toward analyzing datasets not normally thought of as suitable for decoding and makes it possible to run decoding on studies with reduced scan time.

Classification and quantification of palm oil adulteration via portable NIR spectroscopy

Short wave near infrared spectroscopy (NIR) method was used to detect the presence of lard adulteration in palm oil. MicroNIR was set up in two different scan modes to study the effect of path length to the performance of spectral measurement. Pure and adulterated palm oil sample were classified using soft independent modeling class analogy (SIMCA) algorithm with model accuracy more than 0.95 reported for both transflectance and transmission modes. Additionally, by employing partial least square (PLS) regression, the coefficient of determination (R²) of transflectance and transmission were 0.9987 and 0.9994 with root mean square error of calibration (RMSEC) of 0.5931 and 0.6703 respectively. In order to remove the uninformative variables, variable selection using cumulative adaptive reweighted sampling (CARS) has been performed. The result of R² and RMSEC after variable selection for transflectance and transmission were improved significantly. Based on the result of classification and quantification analysis, the transmission mode has yield better prediction model compared to the transflectance mode to distinguish the pure and adulterated palm oil.

Sound identification in human auditory cortex: Differential contribution of local field potentials and high gamma power as revealed by direct intracranial recordings

High gamma power has become the principal means of assessing auditory cortical activation in human intracranial studies, albeit at the expense of low frequency local field potentials (LFPs). It is unclear whether limiting analyses to high gamma impedes ability of clarifying auditory cortical organization. We compared the two measures obtained from posterolateral superior temporal gyrus (PLST) and evaluated their relative utility in sound categorization. Subjects were neurosurgical patients undergoing invasive monitoring for medically refractory epilepsy. Stimuli (consonant-vowel syllables varying in voicing and place of articulation and control tones) elicited robust evoked potentials and high gamma activity on PLST. LFPs had greater across-subject variability, yet yielded higher classification accuracy, relative to high gamma power. Classification was enhanced by including temporal detail of LFPs and combining LFP and high gamma. We conclude that future studies should consider utilizing both LFP and high gamma when investigating the functional organization of human auditory cortex.