Application of supervised machine learning algorithms for the classification of regulatory RNA riboswitches

Characterizing the landscape of cervical squamous cell carcinoma immune microenvironment by integrating the single-cell transcriptomics and RNA-Seq

Background

Cervical squamous cell carcinoma (CSCC), caused by the infection of high‐risk human papillomavirus, is one of the most common malignancies in women worldwide.

Methods

RNA expression data, including those from the Cancer Genome Atlas, Gene Expression Omnibus, and Genotype‐Tissue Expression databases, were used to identify the expression of RNAs in normal and tumor tissue. Correlation analysis was performed to identify the immune‐related long noncoding RNAs (IRLs) and hypoxia‐related genes (IRHs) that can influence the activity of the immune system. Prognosis models of immune‐related RNAs (IRRs) were used to construct a coexpression network of the immune system. We identified the role of IRRs in immunotherapy by correlation analysis with immune checkpoint genes (ICGs). We then validated the expression data by integrating two single‐cell sequencing data sets of CSCC to identify the key immune features.

Results

In total, six immune‐related gene (IRG), four IRL, and five IRH signatures that can significantly influence the characteristics of the tumor immune microenvironment (TIME) were selected using machine learning methods. The expression level of ICGs was significantly upregulated in GZMB⁺CD8⁺ T‐cells and tumor‐associated macrophages (TAMs) in tumor tissues. TGFBI⁺ TAMs are a kind of blood‐derived monocyte‐derived M0‐like TAM linked to hypoxia and a poor prognosis. IFI30⁺ M1‐like TAMs participate in the process of immune‐regulation and showed a role in the promotion of CD8⁺ T‐cells and Type 1 T helper (Th1)/Th2 cells in the coexpression network, together with several IRLs, IRGs, and ICGs.

Conclusions

CD16⁺ monocyte‐derived IFI30⁺ TAMs participated in our coexpression network to regulate the TIME, showing the potential to be a novel immunotherapy target. The enrichment of M0‐like TAMs was associated with a worse prognosis in the high‐risk score group with IRH signatures. Remarkably, M0‐like TAMs in tumor tissues overexpressed TGFBI and were associated with several well‐known tumor‐proliferation pathways.

Colon cancer-specific diagnostic and prognostic biomarkers based on genome-wide abnormal DNA methylation

Abnormal DNA methylation is a major early contributor to colon cancer (COAD) development. We conducted a cohort-based systematic investigation of genome-wide DNA methylation using 299 COAD and 38 normal tissue samples from TCGA. Through conditional screening and machine learning with a training cohort, we identified one hypomethylated and nine hypermethylated differentially methylated CpG sites as potential diagnostic biomarkers, and used them to construct a COAD-specific diagnostic model. Unlike previous models, our model precisely distinguished COAD from nine other cancer types (e.g., breast cancer and liver cancer; error rate ≤ 0.05) and from normal tissues in the training cohort (AUC = 1). The diagnostic model was verified using a validation cohort from The Cancer Genome Atlas (AUC = 1) and five independent cohorts from the Gene Expression Omnibus (AUC ≥ 0.951). Using Cox regression analyses, we established a prognostic model based on six CpG sites in the training cohort, and verified the model in the validation cohort. The prognostic model sensitively predicted patients’ survival (p ≤ 0.00011, AUC ≥ 0.792) independently of important clinicopathological characteristics of COAD (e.g., gender and age). Thus, our DNA methylation analysis provided precise biomarkers and models for the early diagnosis and prognostic evaluation of COAD.

Sequence-to-function deep learning frameworks for engineered riboregulators

While synthetic biology has revolutionized our approaches to medicine, agriculture, and energy, the design of completely novel biological circuit components beyond naturally-derived templates remains challenging due to poorly understood design rules. Toehold switches, which are programmable nucleic acid sensors, face an analogous design bottleneck; our limited understanding of how sequence impacts functionality often necessitates expensive, time-consuming screens to identify effective switches. Here, we introduce Sequence-based Toehold Optimization and Redesign Model (STORM) and Nucleic-Acid Speech (NuSpeak), two orthogonal and synergistic deep learning architectures to characterize and optimize toeholds. Applying techniques from computer vision and natural language processing, we 'un-box' our models using convolutional filters, attention maps, and in silico mutagenesis. Through transfer-learning, we redesign sub-optimal toehold sensors, even with sparse training data, experimentally validating their improved performance. This work provides sequence-to-function deep learning frameworks for toehold selection and design, augmenting our ability to construct potent biological circuit components and precision diagnostics.

Riboflow: Using Deep Learning to Classify Riboswitches With ∼99% Accuracy

Riboswitches are cis-regulatory genetic elements that use an aptamer to control gene expression. Specificity to cognate ligand and diversity of such ligands have expanded the functional repertoire of riboswitches to mediate mounting apt responses to sudden metabolic demands and signal changes in environmental conditions. Given their critical role in microbial life, riboswitch characterisation remains a challenging computational problem. Here we have addressed the issue with advanced deep learning frameworks, namely convolutional neural networks (CNN), and bidirectional recurrent neural networks (RNN) with Long Short-Term Memory (LSTM). Using a comprehensive dataset of 32 ligand classes and a stratified train-validate-test approach, we demonstrated the accurate performance of both the deep learning models (CNN and RNN) relative to conventional hyperparameter-optimized machine learning classifiers on all key performance metrics, including the ROC curve analysis. In particular, the bidirectional LSTM RNN emerged as the best-performing learning method for identifying the ligand-specificity of riboswitches with an accuracy >0.99 and macro-averaged F-score of 0.96. An additional attraction is that the deep learning models do not require prior feature engineering. A dynamic update functionality is built into the models to factor for the constant discovery of new riboswitches, and extend the predictive modeling to new classes. Our work would enable the design of genetic circuits with custom-tuned riboswitch aptamers that would effect precise translational control in synthetic biology. The associated software is available as an open-source Python package and standalone resource for use in genome annotation, synthetic biology, and biotechnology workflows.

A novel riboswitch classification based on imbalanced sequences achieved by machine learning

Riboswitch, a part of regulatory mRNA (50-250nt in length), has two main classes: aptamer and expression platform. One of the main challenges raised during the classification of riboswitch is imbalanced data. That is a circumstance in which the records of a sequences of one group are very small compared to the others. Such circumstances lead classifier to ignore minority group and emphasize on majority ones, which results in a skewed classification. We considered sixteen riboswitch families, to be in accord with recent riboswitch classification work, that contain imbalanced sequences. The sequences were split into training and test set using a newly developed pipeline. From 5460 k-mers (k value 1 to 6) produced, 156 features were calculated based on CfsSubsetEval and BestFirst function found in WEKA 3.8. Statistically tested result was significantly difference between balanced and imbalanced sequences (p < 0.05). Besides, each algorithm also showed a significant difference in sensitivity, specificity, accuracy, and macro F-score when used in both groups (p < 0.05). Several k-mers clustered from heat map were discovered to have biological functions and motifs at the different positions like interior loops, terminal loops and helices. They were validated to have a biological function and some are riboswitch motifs. The analysis has discovered the importance of solving the challenges of majority bias analysis and overfitting. Presented results were generalized evaluation of both balanced and imbalanced models, which implies their ability of classifying, to classify novel riboswitches. The Python source code is available at https://github.com/Seasonsling/riboswitch.

Classification of seed members of five riboswitch families as short sequences based on the features extracted by Block Location-Based Feature Extraction (BLBFE) method

Introduction: Riboswitches are short regulatory elements generally found in the untranslated regions of prokaryotes' mRNAs and classified into several families. Due to the binding possibility between riboswitches and antibiotics, their usage as engineered regulatory elements and also their evolutionary contribution, the need for bioinformatics tools of riboswitch detection is increasing. We have previously introduced an alignment independent algorithm for the identification of frequent sequential blocks in the families of riboswitches. Herein, we report the application of block location-based feature extraction strategy (BLBFE), which uses the locations of detected blocks on riboswitch sequences as features for classification of seed sequences. Besides, mono- and dinucleotide frequencies, k-mer, DAC, DCC, DACC, PC-PseDNC-General and SC-PseDNC-General methods as some feature extraction strategies were investigated. Methods: The classifiers of the Decision tree, KNN, LDA, and Naïve Bayes, as well as k-fold cross-validation, were employed for all methods of feature extraction to compare their performances based on the criteria of accuracy, sensitivity, specificity, and f-score performance measures. Results: The outcome of the study showed that the BLBFE strategy classified the riboswitches indicating 87.65% average correct classification rate (CCR). Moreover, the performance of the proposed feature extraction method was confirmed with average values of 94.31%, 85.01%, 95.45% and 85.38% for accuracy, sensitivity, specificity, and f-score, respectively. Conclusion: Our result approved the performance of the BLBFE strategy in the classification and discrimination of the riboswitch groups showing remarkable higher values of CCR, accuracy, sensitivity, specificity and f-score relative to previously studied feature extraction methods.

Development of a new oligonucleotide block location-based feature extraction (BLBFE) method for the classification of riboswitches

As knowledge of genetics and genome elements increases, the demand for the development of bioinformatics tools for analyzing these data is raised. Riboswitches are genetic components, usually located in the untranslated regions of mRNAs, that regulate gene expression. Additionally, their interaction with antibiotics has been recently suggested, implying a role in antibiotic effects and resistance. Following a previously published sequential block finding algorithm, herein, we report the development of a new block location-based feature extraction strategy (BLBFE). This procedure utilizes the locations of family-specific sequential blocks on riboswitch sequences as features. Furthermore, the performance of other feature extraction strategies, including mono- and dinucleotide frequencies, k-mer, DAC, DCC, DACC, PC-PseDNC-General and SC-PseDNC-General methods, was investigated. KNN, LDA, naïve Bayes, PNN and decision tree classifiers accompanied by V-fold cross-validation were applied for all methods of feature extraction, and their performances based on the defined feature extraction strategies were compared. Performance measures of accuracy, sensitivity, specificity and F-score for each method of feature extraction were studied. The proposed feature extraction strategy resulted in classification of riboswitches with an average correct classification rate (CCR) of 90.8%. Furthermore, the obtained data confirmed the performance of the developed feature extraction method with an average accuracy of 96.1%, an average sensitivity of 90.8%, an average specificity of 97.52% and an average F-score of 90.69%. Our results implied that the proposed feature extraction (BLBFE) method can classify and discriminate riboswitch families with high CCR, accuracy, sensitivity, specificity and F-score values.

Classification of Riboswitch Families Using Block Location-Based Feature Extraction (BLBFE) Method

Purpose: Riboswitches are special non-coding sequences usually located in mRNAs' un-translated regions and regulate gene expression and consequently cellular function. Furthermore, their interaction with antibiotics has been recently implicated. This raises more interest in development of bioinformatics tools for riboswitch studies. Herein, we describe the development and employment of novel block location-based feature extraction (BLBFE) method for classification of riboswitches. Methods: We have already developed and reported a sequential block finding (SBF) algorithm which, without operating alignment methods, identifies family specific sequential blocks for riboswitch families. Herein, we employed this algorithm for 7 riboswitch families including lysine, cobalamin, glycine, SAM-alpha, SAM-IV, cyclic-di-GMP-I and SAH. Then the study was extended toward implementation of BLBFE method for feature extraction. The outcome features were applied in various classifiers including linear discriminant analysis (LDA), probabilistic neural network (PNN), decision tree and k-nearest neighbors (KNN) classifiers for classification of the riboswitch families. The performance of the classifiers was investigated according to performance measures such as correct classification rate (CCR), accuracy, sensitivity, specificity and f-score. Results: As a result, average CCR for classification of riboswitches was 87.87%. Furthermore, application of BLBFE method in 4 classifiers displayed average accuracies of 93.98% to 96.1%, average sensitivities of 76.76% to 83.61%, average specificities of 96.53% to 97.69% and average f-scores of 74.9% to 81.91%. Conclusion: Our results approved that the proposed method of feature extraction; i.e. BLBFE method; can be successfully used for classification and discrimination of the riboswitch families with high CCR, accuracy, sensitivity, specificity and f-score values.

Tuning the Performance of Synthetic Riboswitches using Machine Learning

Riboswitch development for clinical, technological, and synthetic biology applications constantly seeks to optimize regulatory behavior. Here, we present a machine learning approach to improve the regulation of a tetracycline (tc)-dependent riboswitch device composed of two individual tc aptamers. We developed a bioinformatics model that combines random forest analysis with a convolutional neural network to predict the switching behavior of such tandem riboswitches. We found that both biophysical parameters and the hydrogen bond pattern influence regulation. Our new design pipeline led to significant improvement of the tc riboswitch device with a dynamic range extension from 8.5 to 40-fold. We are confident that our novel method not only results in an excellent tc-dependent riboswitch device but further holds great promise and potential for the optimization of other riboswitches.