Stochastic Neighbor Embedding Research Articles

A nondestructive method for determination of green tea quality by hyperspectral imaging

This study utilizes hyperspectral imaging technology to capture spectral images of three grades of green tea. Reflectance spectra of the three quality grades of tea, namely Grade A, Grade B, and Grade C, were collected between 370 nm and 1040 nm. Three methods, namely Savitzky-Golay (SG) smoothing, standard normal variable (SNV), and first derivative (FD), were employed to preprocess the raw spectra. Dimensionality reduction and visual display were accomplished using t-distributed stochastic neighbor embedding, and the preprocessed spectra with the most optimal visualization effect were selected to extract the key wavelengths using principal components analysis (PCA). The extracted key bands could aid in achieving the detection of tea quality grades, as tea polyphenols are more sensitive between 650 and 800 nm, where 664 nm and 765 nm represent the key wavelengths of catechins. K-nearest neighbor (KNN) and support vector machine (SVM) discriminative models were deployed to model the key wavelengths. The model built by FD-PCA-SVM exhibited the best discriminant effect, with an accuracy of 93.8 % in the training set and 98.2 % in the test set. FD-PCA-SVM discriminant model was used to identify and visualize the tea quality grades with good results. Hyperspectral imaging technology is well-suited to identifying the quality of green tea.

Rapid Classification of Sarcomas Using Methylation Fingerprint: A Pilot Study.

Sarcoma classification is challenging and can lead to treatment delays. Previous studies used DNA aberrations and machine-learning classifiers based on methylation profiles for diagnosis. We aimed to classify sarcomas by analyzing methylation signatures obtained from low-coverage whole-genome sequencing, which also identifies copy-number alterations. DNA was extracted from 23 suspected sarcoma samples and sequenced on an Oxford Nanopore sequencer. The methylation-based classifier, applied in the nanoDx pipeline, was customized using a reference set based on processed Illumina-based methylation data. Classification analysis utilized the Random Forest algorithm and t-distributed stochastic neighbor embedding, while copy-number alterations were detected using a designated R package. Out of the 23 samples encompassing a restricted range of sarcoma types, 20 were successfully sequenced, but two did not contain tumor tissue, according to the pathologist. Among the 18 tumor samples, 14 were classified as reported in the pathology results. Four classifications were discordant with the pathological report, with one compatible and three showing discrepancies. Improving tissue handling, DNA extraction methods, and detecting point mutations and translocations could enhance accuracy. We envision that rapid, accurate, point-of-care sarcoma classification using nanopore sequencing could be achieved through additional validation in a diverse tumor cohort and the integration of methylation-based classification and other DNA aberrations.

An Unsupervised Machine Learning Approach for Monitoring Data Fusion and Health Indicator Construction.

The prediction of system degradation is very important as it serves as an important basis for the formulation of condition-based maintenance strategies. An effective health indicator (HI) plays a key role in the prediction of system degradation as it enables vital information for critical tasks ranging from fault diagnosis to remaining useful life prediction. To address this issue, a method for monitoring data fusion and health indicator construction based on an autoencoder (AE) and a long short-term memory (LSTM) network is proposed in this study to improve the predictability and effectiveness of health indicators. Firstly, an unsupervised method and overall framework for HI construction is built based on a deep autoencoder and an LSTM neural network. The neural network is trained fully based on the normal operating monitoring data and then the construction error of the AE model is adopted as the health indicator of the system. Secondly, we propose related machine learning techniques for monitoring data processing to overcome the issue of data fusion, such as mutual information for sensor selection and t-distributed stochastic neighbor embedding (T-SNE) for operating condition identification. Thirdly, in order to verify the performance of the proposed method, experiments are conducted based on the CMAPSS dataset and results are compared with algorithms of principal component analysis (PCA) and a vanilla autoencoder model. Result shows that the LSTM-AE model outperforms the PCA and Vanilla-AE model in the metrics of monotonicity, trendability, prognosability, and fitness. Fourthly, in order to analyze the impact of the time step of the LSMT-AE model on HI construction, we construct and analyze the system HI curve under different time steps of 5, 10, 15, 20, and 25 cycles. Finally, the results demonstrate that the proposed method for HI construction can effectively characterize the health state of a system, which is helpful for the development of further failure prognostics and converting the scheduled maintenance into condition-based maintenance.

Ultra-Short-Term Load Forecasting for Customer-Level Integrated Energy Systems Based on Composite VTDS Models

A method is proposed to address the challenging issue of load prediction in user-level integrated energy systems (IESs) using a composite VTDS model. Firstly, an IES multi-dimensional load time series is decomposed into multiple intrinsic mode functions (IMFs) using variational mode decomposition (VMD). Then, each IMF, along with other influential features, is subjected to data dimensionality reduction and clustering denoising using t-distributed stochastic neighbor embedding (t-SNE) and fast density-based spatial clustering of applications with noise (FDBSCAN) to perform major feature selection. Subsequently, the reduced and denoised data are reconstructed, and a time-aware long short-term memory (T-LSTM) artificial neural network is employed to fill in missing data by incorporating time interval information. Finally, the selected multi-factor load time series is used as input into a support vector regression (SVR) model optimized using the quantum particle swarm optimization (QPSO) algorithm for load prediction. Using measured load data from a specific user-level IES at the Tempe campus of Arizona State University, USA, as a case study, a comparative analysis between the VTDS method and other approaches is conducted. The results demonstrate that the method proposed in this study achieved higher accuracy in short-term forecasting of the IES’s multiple loads.

CO and NOx emissions prediction in gas turbine using a novel modeling pipeline based on the combination of deep forest regressor and feature engineering

The main objective of this study is to estimate carbon oxides (CO) and nitrogen oxides (NOx) emissions from a gas turbine using the predictive emission monitoring systems dataset. First, four methods were developed for feature generation: Principal Component Analysis, t-Distributed Stochastic Neighbor Embedding, Uniform Manifold Approximation and Projection, and Potential of Heat-diffusion for Affinity-based Trajectory Embedding. Then Feature Relevance-based Unsupervised Feature Selection was evaluated for ranking features. With all features generated, the regression models Ridge Regression, Least Absolute Shrinkage and Selection Operator, k-Nearest Neighbor, Cubist Regression, Random Forest, Light Gradient Boosting Machine, Categorical Boosting, and Deep Forest Regression (DFR) were evaluated. The hyperparameters were tuned with Randomized Search Cross Validation using a 5-fold cross-validation (CV) procedure. In this paper, an innovative modeling pipeline based on the mentioned regressors, feature engineering, and effective hyperparameters tuning was proposed. The developed regression models based on feature engineering and hyperparameters tuning were compared using performance criteria, including mean 5-fold CV values of the coefficient of determination (R2) and root mean square error (RMSE). The obtained best results based on mean CV values using the DFR presented R2 values equal to 0.9647 and 0.5355 and RMSE values equal to 0.4245 and 1.4508 for CO emissions, for the training and validation datasets, respectively. On the other hand, DFR presented R2 equal to 0.9866 and 0.4737 and RMSE equal to 1.3876 and 7.6528 for NOx emissions, for the training and validation datasets, respectively.The results show that the proposed DFR model has higher potential for CO and NOx emissions prediction and that feature engineering and hyperparameter tuning has demonstrated impacts on the predictive capacity of the evaluated regression models.

Biased Deep Learning Methods in Detection of COVID-19 Using CT Images: A Challenge Mounted by Subject-Wise-Split ISFCT Dataset.

Accurate detection of respiratory system damage including COVID-19 is considered one of the crucial applications of deep learning (DL) models using CT images. However, the main shortcoming of the published works has been unreliable reported accuracy and the lack of repeatability with new datasets, mainly due to slice-wise splits of the data, creating dependency between training and test sets due to shared data across the sets. We introduce a new dataset of CT images (ISFCT Dataset) with labels indicating the subject-wise split to train and test our DL algorithms in an unbiased manner. We also use this dataset to validate the real performance of the published works in a subject-wise data split. Another key feature provides more specific labels (eight characteristic lung features) rather than being limited to COVID-19 and healthy labels. We show that the reported high accuracy of the existing models on current slice-wise splits is not repeatable for subject-wise splits, and distribution differences between data splits are demonstrated using t-distribution stochastic neighbor embedding. We indicate that, by examining subject-wise data splitting, less complicated models show competitive results compared to the exiting complicated models, demonstrating that complex models do not necessarily generate accurate and repeatable results.

Investigating the impact of structured reporting on the linguistic standardization of radiology reports through natural language processing over a 10-year period

ObjectivesTo investigate how a transition from free text to structured reporting affects reporting language with regard to standardization and distinguishability.MethodsA total of 747,393 radiology reports dictated between January 2011 and June 2020 were retrospectively analyzed. The body and cardiothoracic imaging divisions introduced a reporting concept using standardized language and structured reporting templates in January 2016.Reports were segmented by a natural language processing algorithm and converted into a 20-dimension document vector. For analysis, dimensionality was reduced to a 2D visualization with t-distributed stochastic neighbor embedding and matched with metadata. Linguistic standardization was assessed by comparing distinct report types’ vector spreads (e.g., run-off MR angiography) between reporting standards. Changes in report type distinguishability (e.g., CT abdomen/pelvis vs. MR abdomen) were measured by comparing the distance between their centroids.ResultsStructured reports showed lower document vector spread (thus higher linguistic similarity) compared with free-text reports overall (21.9 [free-text] vs. 15.9 [structured]; − 27.4%; p < 0.001) and for most report types, e.g., run-off MR angiography (15.2 vs. 1.8; − 88.2%; p < 0.001) or double-rule-out CT (26.8 vs. 10.0; − 62.7%; p < 0.001). No changes were observed for reports continued to be written in free text, e.g., CT head reports (33.2 vs. 33.1; − 0.3%; p = 1).Distances between the report types’ centroids increased with structured reporting (thus better linguistic distinguishability) overall (27.3 vs. 54.4; + 99.3 ± 98.4%) and for specific report types, e.g., CT abdomen/pelvis vs. MR abdomen (13.7 vs. 37.2; + 171.5%).ConclusionStructured reporting and the use of factual language yield more homogenous and standardized radiology reports on a linguistic level, tailored to specific reporting scenarios and imaging studies.Clinical relevanceInformation transmission to referring physicians, as well as automated report assessment and content extraction in big data analyses, may benefit from standardized reporting, due to consistent report organization and terminology used for pathologies and normal findings.Key Points• Natural language processing and t-distributed stochastic neighbor embedding can transform radiology reports into numeric vectors, allowing the quantification of their linguistic standardization.• Structured reporting substantially increases reports’ linguistic standardization (mean: − 27.4% in vector spread) and distinguishability (mean: + 99.3 ± 98.4% increase in vector distance) compared with free-text reports.• Higher standardization and homogeneity outline potential benefits of structured reporting for information transmission and big data analyses.

Multiscale Time-Frequency Sparse Transformer Based on Partly Interpretable Method for Bearing Fault Diagnosis

Transformer model is being gradually studied and applied in bearing fault diagnosis tasks, which can overcome the feature extraction defects caused by long-term dependencies in convolution neural network (CNN) and recurrent neural network (RNN). To optimize the structure of existing transformer-like methods and improve the diagnostic accuracy, we proposed a novel method based on the multiscale time-frequency sparse transformer (MTFST) in this paper. First, a novel tokenizer based on shot-time Fourier transform (STFT) is designed, which processes the 1D format raw signals into 2D format discrete time-frequency sequences in the embedding space. Second, a sparse self-attention mechanism is designed to eliminate the feature mapping defect in naive self-attention mechanism. Then, the novel encoder-decoder structure is presented, the multiple encoders are employed to extract the hidden feature of different time-frequency sequences obtained by STFT with different window widths, and the decoder is used to remap the deep information and connect to the classifier for discriminating fault types. The proposed method is tested in the XJTU-SY bearing dataset and self-made experiment rig dataset, and the following work is conducted. The influences of hyperparameters on diagnosis accuracy and number of parameters are analysed in detail. The weights of the attention mechanism (AM) are visualized and analysed to study the interpretability, which explains the partly working pattern of the network. In the comparison test with other existing CNN, RNN, and transformer models, the diagnosis accuracy of different methods is statistically analysed, feature vectors are presented via the t-distributed stochastic neighbor embedding (t-SNE) method, and the proposed MTFST obtains the best accuracy and feature distribution form. The results demonstrate the effectiveness and superiority of the proposed method in bearing fault diagnosis.

Novel SERS biosensor for rapid detection of breast cancer based on Ag2O-Ag-PSi nanochips

Ag2O-Ag-PSi (porous silicon) surface-enhanced Raman scattering (SERS) chip was successfully synthesized by electrochemical corrosion, in situ reduction and heat treatment technology. The influence of different heat treatment temperature on SERS performance of the chip is studied. The results show that the chip treated at 300 °C has the best SERS performance. The chip was composed of Ag2O-Ag nano core shell with a diameter of 40–60 nm and porous silicon substrate. Then, the optimized chip was used to perform SERS test on serum samples from 30 healthy volunteers and 30 early breast cancer patients, and the baseline was corrected by LabSpec6 software. Finally, the data were analyzed by principal component analysis combined with t-distributed Stochastic Neighbor Embedding (PCA-t-SNE). The results showed that the accuracy of the improved substrate combined with multivariate statistical method was 98%. The shelf life of the chips exceeded six months due to the presence of the Ag2O shell. This study provides a basis for developing a low-cost rapid and sensitive early screening technology for breast cancer.

Galactic ArchaeoLogIcaL ExcavatiOns (GALILEO)

Based on high-quality Apache Point Observatory Galactic Evolution Experiment (APOGEE) DR17 and Gaia DR3 data for 1742 red giants stars within 5 kpc of the Sun and not rotating with the Galactic disk (Vϕ &lt; 100 km s−1), we used the nonlinear technique of unsupervised analysis t-Distributed Stochastic Neighbor Embedding (t-SNE) to detect coherent structures in the space of ten chemical-abundance ratios: [Fe/H], [O/Fe], [Mg/Fe], [Si/Fe], [Ca/Fe], [C/Fe], [N/Fe], [Al/Fe], [Mn/Fe], and [Ni/Fe]. Additionally, we obtained orbital parameters for each star using the nonaxisymmetric gravitational potential GravPot16. Seven structures are detected, including Splash, Gaia-Sausage-Enceladus (GSE), the high-α heated-disk population, N-C-O peculiar stars, and inner disk-like stars, plus two other groups that did not match anything previously reported in the literature, here named Galileo 5 and Galileo 6 (G5 and G6). These two groups overlap with Splash in [Fe/H], with G5 having a lower metallicity than G6, and they are both between GSE and Splash in the [Mg/Mn] versus [Al/Fe] plane, with G5 being in the α-rich in situ locus and G6 on the border of the α-poor in situ one. Nonetheless, their low [Ni/Fe] hints at a possible ex situ origin. Their orbital energy distributions are between Splash and GSE, with G5 being slightly more energetic than G6. We verified the robustness of all the obtained groups by exploring a large range of t-SNE parameters, applying it to various subsets of data, and also measuring the effect of abundance errors through Monte Carlo tests.

Machine Learning of the Whole Genome Sequence of Mycobacterium tuberculosis: A Scoping PRISMA-Based Review.

Tuberculosis (TB) remains one of the most significant global health problems, posing a significant challenge to public health systems worldwide. However, diagnosing drug-resistant tuberculosis (DR-TB) has become increasingly challenging due to the rising number of multidrug-resistant (MDR-TB) cases, despite the development of new TB diagnostic tools. Even the World Health Organization-recommended methods such as Xpert MTB/XDR or Truenat are unable to detect all the Mycobacterium tuberculosis genome mutations associated with drug resistance. While Whole Genome Sequencing offers a more precise DR profile, the lack of user-friendly bioinformatics analysis applications hinders its widespread use. This review focuses on exploring various artificial intelligence models for predicting DR-TB profiles, analyzing relevant English-language articles using the PRISMA methodology through the Covidence platform. Our findings indicate that an Artificial Neural Network is the most commonly employed method, with non-statistical dimensionality reduction techniques preferred over traditional statistical approaches such as Principal Component Analysis or t-distributed Stochastic Neighbor Embedding.

Adulteration quantification of cheap honey in high-quality Manuka honey by two-dimensional correlation spectroscopy combined with deep learning

The adulteration of cheap honey in high value honey has a negative impact on consumer health, and its detection is also a challenging task. A fast, non-invasive and cost-effective spectroscopic strategy was proposed to quantify the honey fraud based on the two-dimensional correlation spectroscopy (2D-COS) of Raman spectra combined with multiple deep learning techniques. 700 Raman spectra of Manuka, acacia and multi-floral honeys were collected, and the corresponding synchronous, asynchronous and integrative correlation spectra were obtained. The t-distributed stochastic neighbor embedding (t-SNE) and partial least square regression (PLSR) were used to analyze one-dimensional spectra and 2D-COS image datasets of the same sample, demonstrating that 2D-COS can highlight the complex fingerprint features of samples and thus contribute to the spectral characterization. The combination of synchronous 2D-COS and deep residual shrinkage networks (DRSN) achieved the best performance compared to the other models, with the root mean square errors of prediction (RMSEP) of 3.1166 for Manuka honey and 2.3188 for acacia honey, respectively. The impressive performance of the proposed spectral method is sufficient to quantify the adulteration of cheap honey. This study provides a novel alternative for relevant detection tasks and facilitates to ascertain the authenticity of high-value honey products.

Unsupervised Feature Representation of Sleep EEG Data with Transient Deep Boltzmann Machine.

The supervised sleep staging methods are challenged by their strict requirements of a labelled and large dataset. This study considers an unsupervised dimensionality reduction method, the Deep Boltzmann Machine (DBM), trained to a transient state for binary classification of sleep stages. First, the joint time-frequency domain features from the polysomnographic recordings are extracted. Second, the extracted features are smoothed using 2 min rolling window to include contextual temporal information, and finally, they serve as an input for unsupervised training of DBM_transient. The results show that our method effectively separates the sleep stages in two-dimensional feature space with a large Fisher's discriminant value. The classification performance by the DBM_transient achieves a 96.1% F1 score, which is higher than DBM converged to an equilibrium state (95.2%), Principal Component Analysis (92.5%), Isometric Feature Mapping (95.9%), t-distributed Stochastic Neighbor Embedding (94.9%), and Uniform Manifold Approximation (95.0%) on the widely used sleep-EDF database. Additionally, Fisher's discriminant function demonstrates the superiority of the DBM_transient. The significance of the DBM transient lies in its ease of interpretability in two-dimensional space, and future multi-class implementation of the method may facilitate its usage in clinical applications.

Disruption prediction on EAST with different wall conditions based on a multi-scale deep hybrid neural network

Plasma disruption is a very dangerous event for future tokamaks and fusion reactors. Therefore, predicting disruption is crucial for ensuring the safety and performance of reactors. In this study, the features of two deep learning algorithms are integrated to establish a multi-scale hybrid network disruption predictor. Firstly, 43 diagnostic signals are extracted by a convolutional neural network (CNN), and the time information is learned by a long short-term memory network. The predictor is trained and tested on a database containing non-all-metal wall discharges. Its area under the receiver operator characteristic curve (AUC), which is a common performance metric for deep learning algorithms, reaches 0.97, and the true positive rate is , while the false positive rate is . Since EAST was upgraded in 2020, the wall condition has been upgraded from non-all-metal to all-metal. To examine the robustness of the predictor, the EAST disruption predictor is migrated to the all-metal wall experiment for the first time. It is again tested under the all-metal wall experimental data, and its warning performance decreases significantly, with an AUC of only 0.79. To improve the robustness and sensitivity of the predictor against disruptions, the convolutional attention mechanism is introduced into the CNN. After training and testing with the same data set, the warning performance for the all-metal wall data is improved, with the AUC value increasing to 0.84. To further improve the robustness of the predictor, the t-distributed stochastic neighbor embedding algorithm is employed to explore the difference of the sample distribution before and after EAST upgradation, and the transfer learning algorithm is used to reduce this difference. By applying transfer learning to a small amount of discharges from the all-metal wall data, the warning performance in the all-metal wall experiment is further improved, and the AUC value increases to 0.93. Overall, the proposed data-driven predictor is promising for applications in future devices such as ITER.

An efficient ensemble-based Machine Learning for breast cancer detection

Breast cancer is a very severe type of cancer that often develops in breast cells. Attempting to develop an effective predictive model for breast cancer prognosis prediction is urgently needed despite substantial advancements in the management of symptomatic breast cancer over the past ten years. The precise prediction will offer numerous advantages, including the ability to diagnose cancer at an early stage and protect patients from needless medical care and related costs. In the medical field, recall is just as important as model accuracy. Even more crucially in the medical area, a model is not very good if its accuracy is high but its recall is low. To boost accuracy while still assigning equal weight to recall, we proposed a model that ensembles Feature Selection (FS), Feature Extraction (FE), and 5 Machine Learning (ML) models. There are three steps in our proposed model. The Correlation Coefficient (CC) and Anova (Anv) feature selection methodologies to choose the features in the first stage. Applying Uniform Manifold Approximation and Projection (UMAP), t-distributed Stochastic Neighbour Embedding (t-SNE), and Principal Component Analysis (PCA) to extract the features in the second stage without compromising the crucial information. With 5 ML models and ensemble models such as Voting Classifier (VC) and Stacking Classifier (SC) after selecting and extracting features from the dataset to predict the disease will be the last stage. The results show that the proposed model CC-Anv with PCA using a SC outperformed all the existing methodologies with 100% accuracy, precision, recall, and f1-score.

Topological Comparison of Some Dimension Reduction Methods Using Persistent Homology on EEG Data

In this paper, we explore how to use topological tools to compare dimension reduction methods. We first make a brief overview of some of the methods often used in dimension reduction such as isometric feature mapping, Laplacian Eigenmaps, fast independent component analysis, kernel ridge regression, and t-distributed stochastic neighbor embedding. We then give a brief overview of some of the topological notions used in topological data analysis, such as barcodes, persistent homology, and Wasserstein distance. Theoretically, when these methods are applied on a data set, they can be interpreted differently. From EEG data embedded into a manifold of high dimension, we discuss these methods and we compare them across persistent homologies of dimensions 0, 1, and 2, that is, across connected components, tunnels and holes, shells around voids, or cavities. We find that from three dimension clouds of points, it is not clear how distinct from each other the methods are, but Wasserstein and Bottleneck distances, topological tests of hypothesis, and various methods show that the methods qualitatively and significantly differ across homologies. We can infer from this analysis that topological persistent homologies do change dramatically at seizure, a finding already obtained in previous analyses. This suggests that looking at changes in homology landscapes could be a predictor of seizure.

Identification of Robust Antibiotic Subgroups by Integrating Multi-Species Drug-Drug Interactions.

Previous studies have shown that antibiotics can be divided into groups, and drug-drug interactions (DDI) depend on their groups. However, these studies focused on a specific bacteria strain (i.e., Escherichia coli BW25113). Existing datasets often contain noise. Noisy labeled data may have a bad effect on the clustering results. To address this problem, we developed a multi-source information fusion method for integrating DDI information from multiple bacterial strains. Specifically, we calculated drug similarities based on the DDI network of each bacterial strain and then fused these drug similarity matrices to obtain a new fused similarity matrix. The fused similarity matrix was combined with the T-distributed stochastic neighbor embedding algorithm, and hierarchical clustering algorithm can effectively identify antibiotic subgroups. These antibiotic subgroups are strongly correlated with known antibiotic classifications, and group-group interactions are almost monochromatic. In summary, our method provides a promising framework for understanding the mechanism of action of antibiotics and exploring multi-species group-group interactions.

A Preprocessing Manifold Learning Strategy Based on t-Distributed Stochastic Neighbor Embedding.

In machine learning and data analysis, dimensionality reduction and high-dimensional data visualization can be accomplished by manifold learning using a t-Distributed Stochastic Neighbor Embedding (t-SNE) algorithm. We significantly improve this manifold learning scheme by introducing a preprocessing strategy for the t-SNE algorithm. In our preprocessing, we exploit Laplacian eigenmaps to reduce the high-dimensional data first, which can aggregate each data cluster and reduce the Kullback-Leibler divergence (KLD) remarkably. Moreover, the k-nearest-neighbor (KNN) algorithm is also involved in our preprocessing to enhance the visualization performance and reduce the computation and space complexity. We compare the performance of our strategy with that of the standard t-SNE on the MNIST dataset. The experiment results show that our strategy exhibits a stronger ability to separate different clusters as well as keep data of the same kind much closer to each other. Moreover, the KLD can be reduced by about 30% at the cost of increasing the complexity in terms of runtime by only 1-2%.

PromGER: Promoter Prediction Based on Graph Embedding and Ensemble Learning for Eukaryotic Sequence.

Promoters are DNA non-coding regions around the transcription start site and are responsible for regulating the gene transcription process. Due to their key role in gene function and transcriptional activity, the prediction of promoter sequences and their core elements accurately is a crucial research area in bioinformatics. At present, models based on machine learning and deep learning have been developed for promoter prediction. However, these models cannot mine the deeper biological information of promoter sequences and consider the complex relationship among promoter sequences. In this work, we propose a novel prediction model called PromGER to predict eukaryotic promoter sequences. For a promoter sequence, firstly, PromGER utilizes four types of feature-encoding methods to extract local information within promoter sequences. Secondly, according to the potential relationships among promoter sequences, the whole promoter sequences are constructed as a graph. Furthermore, three different scales of graph-embedding methods are applied for obtaining the global feature information more comprehensively in the graph. Finally, combining local features with global features of sequences, PromGER analyzes and predicts promoter sequences through a tree-based ensemble-learning framework. Compared with seven existing methods, PromGER improved the average specificity of 13%, accuracy of 10%, Matthew's correlation coefficient of 16%, precision of 4%, F1 score of 6%, and AUC of 9%. Specifically, this study interpreted the PromGER by the t-distributed stochastic neighbor embedding (t-SNE) method and SHAPley Additive exPlanations (SHAP) value analysis, which demonstrates the interpretability of the model.

Manifold learning for fMRI time-varying functional connectivity.

Whole-brain functional connectivity (FC) measured with functional MRI (fMRI) evolves over time in meaningful ways at temporal scales going from years (e.g., development) to seconds [e.g., within-scan time-varying FC (tvFC)]. Yet, our ability to explore tvFC is severely constrained by its large dimensionality (several thousands). To overcome this difficulty, researchers often seek to generate low dimensional representations (e.g., 2D and 3D scatter plots) hoping those will retain important aspects of the data (e.g., relationships to behavior and disease progression). Limited prior empirical work suggests that manifold learning techniques (MLTs)-namely those seeking to infer a low dimensional non-linear surface (i.e., the manifold) where most of the data lies-are good candidates for accomplishing this task. Here we explore this possibility in detail. First, we discuss why one should expect tvFC data to lie on a low dimensional manifold. Second, we estimate what is the intrinsic dimension (ID; i.e., minimum number of latent dimensions) of tvFC data manifolds. Third, we describe the inner workings of three state-of-the-art MLTs: Laplacian Eigenmaps (LEs), T-distributed Stochastic Neighbor Embedding (T-SNE), and Uniform Manifold Approximation and Projection (UMAP). For each method, we empirically evaluate its ability to generate neuro-biologically meaningful representations of tvFC data, as well as their robustness against hyper-parameter selection. Our results show that tvFC data has an ID that ranges between 4 and 26, and that ID varies significantly between rest and task states. We also show how all three methods can effectively capture subject identity and task being performed: UMAP and T-SNE can capture these two levels of detail concurrently, but LE could only capture one at a time. We observed substantial variability in embedding quality across MLTs, and within-MLT as a function of hyper-parameter selection. To help alleviate this issue, we provide heuristics that can inform future studies. Finally, we also demonstrate the importance of feature normalization when combining data across subjects and the role that temporal autocorrelation plays in the application of MLTs to tvFC data. Overall, we conclude that while MLTs can be useful to generate summary views of labeled tvFC data, their application to unlabeled data such as resting-state remains challenging.