Mirnacle: machine learning with SMOTE and random forest for improving selectivity in pre-miRNA ab initio prediction

Improving mixed-integer temporal modeling by generating synthetic data using conditional generative adversarial networks: A case study of fluid overload prediction in the intensive care unit

OBJECTIVE

The challenge of mixed-integer temporal data, which is particularly prominent for medication use in the critically ill, limits the performance of predictive models. The purpose of this evaluation was to pilot test integrating synthetic data within an existing dataset of complex medication data to improve machine learning model prediction of fluid overload.

MATERIALS AND METHODS

This retrospective cohort study evaluated patients admitted to an ICU ≥ 72 h. Four machine learning algorithms to predict fluid overload after 48-72 h of ICU admission were developed using the original dataset. Then, two distinct synthetic data generation methodologies (synthetic minority over-sampling technique (SMOTE) and conditional tabular generative adversarial network (CTGAN)) were used to create synthetic data. Finally, a stacking ensemble technique designed to train a meta-learner was established. Models underwent training in three scenarios of varying qualities and quantities of datasets.

RESULTS

Training machine learning algorithms on the combined synthetic and original dataset overall increased the performance of the predictive models compared to training on the original dataset. The highest performing model was the meta-model trained on the combined dataset with 0.83 AUROC while it managed to significantly enhance the sensitivity across different training scenarios.

DISCUSSION

The integration of synthetically generated data is the first time such methods have been applied to ICU medication data and offers a promising solution to enhance the performance of machine learning models for fluid overload, which may be translated to other ICU outcomes. A meta-learner was able to make a trade-off between different performance metrics and improve the ability to identify the minority class.

Characterization of spleen and lymph node cell types via CITE-seq and machine learning methods

The spleen and lymph nodes are important functional organs for human immune system. The identification of cell types for spleen and lymph nodes is helpful for understanding the mechanism of immune system. However, the cell types of spleen and lymph are highly diverse in the human body. Therefore, in this study, we employed a series of machine learning algorithms to computationally analyze the cell types of spleen and lymph based on single-cell CITE-seq sequencing data. A total of 28,211 cell data (training vs. test = 14,435 vs. 13,776) involving 24 cell types were collected for this study. For the training dataset, it was analyzed by Boruta and minimum redundancy maximum relevance (mRMR) one by one, resulting in an mRMR feature list. This list was fed into the incremental feature selection (IFS) method, incorporating four classification algorithms (deep forest, random forest, K-nearest neighbor, and decision tree). Some essential features were discovered and the deep forest with its optimal features achieved the best performance. A group of related proteins (CD4, TCRb, CD103, CD43, and CD23) and genes (Nkg7 and Thy1) contributing to the classification of spleen and lymph nodes cell types were analyzed. Furthermore, the classification rules yielded by decision tree were also provided and analyzed. Above findings may provide helpful information for deepening our understanding on the diversity of cell types.

Identification of Type 2 Diabetes Biomarkers From Mixed Single-Cell Sequencing Data With Feature Selection Methods

Diabetes is the most common disease and a major threat to human health. Type 2 diabetes (T2D) makes up about 90% of all cases. With the development of high-throughput sequencing technologies, more and more fundamental pathogenesis of T2D at genetic and transcriptomic levels has been revealed. The recent single-cell sequencing can further reveal the cellular heterogenicity of complex diseases in an unprecedented way. With the expectation on the molecular essence of T2D across multiple cell types, we investigated the expression profiling of more than 1,600 single cells (949 cells from T2D patients and 651 cells from normal controls) and identified the differential expression profiling and characteristics at the transcriptomics level that can distinguish such two groups of cells at the single-cell level. The expression profile was analyzed by several machine learning algorithms, including Monte Carlo feature selection, support vector machine, and repeated incremental pruning to produce error reduction (RIPPER). On one hand, some T2D-associated genes (MTND4P24, MTND2P28, and LOC100128906) were discovered. On the other hand, we revealed novel potential pathogenic mechanisms in a rule manner. They are induced by newly recognized genes and neglected by traditional bulk sequencing techniques. Particularly, the newly identified T2D genes were shown to follow specific quantitative rules with diabetes prediction potentials, and such rules further indicated several potential functional crosstalks involved in T2D.

Computational Methods for Predicting Mature microRNAs

Tiny single-stranded noncoding RNAs with size 19-27 nucleotides serve as microRNAs (miRNAs), which have emerged as key gene regulators in the last two decades. miRNAs serve as one of the hallmarks in regulatory pathways with critical roles in human diseases. Ever since the discovery of miRNAs, researchers have focused on how mature miRNAs are produced from precursor mRNAs. Experimental methods are faced with notorious challenges in terms of experimental design, since it is time consuming and not cost-effective. Hence, different computational methods have been employed for the identification of miRNA sequences where most of them labeled as miRNA predictors are in fact pre-miRNA predictors and provide no information about the putative miRNA location within the pre-miRNA. This chapter provides an update and the current state of the art in this area covering various methods and 15 software suites used for prediction of mature miRNA.

iMPT-FDNPL: Identification of Membrane Protein Types with Functional Domains and a Natural Language Processing Approach

Membrane protein is an important kind of proteins. It plays essential roles in several cellular processes. Based on the intramolecular arrangements and positions in a cell, membrane proteins can be divided into several types. It is reported that the types of a membrane protein are highly related to its functions. Determination of membrane protein types is a hot topic in recent years. A plenty of computational methods have been proposed so far. Some of them used functional domain information to encode proteins. However, this procedure was still crude. In this study, we designed a novel feature extraction scheme to obtain informative features of proteins from their functional domain information. Such scheme termed domains as words and proteins, represented by its domains, as sentences. The natural language processing approach, word2vector, was applied to access the features of domains, which were further refined to protein features. Based on these features, RAndom k-labELsets with random forest as the base classifier was employed to build the multilabel classifier, namely, iMPT-FDNPL. The tenfold cross-validation results indicated the good performance of such classifier. Furthermore, such classifier was superior to other classifiers based on features derived from functional domains via one-hot scheme or derived from other properties of proteins, suggesting the effectiveness of protein features generated by the proposed scheme.

Recognizing Pattern and Rule of Mutation Signatures Corresponding to Cancer Types

Cancer has been generally defined as a cluster of systematic malignant pathogenesis involving abnormal cell growth. Genetic mutations derived from environmental factors and inherited genetics trigger the initiation and progression of cancers. Although several well-known factors affect cancer, mutation features and rules that affect cancers are relatively unknown due to limited related studies. In this study, a computational investigation on mutation profiles of cancer samples in 27 types was given. These profiles were first analyzed by the Monte Carlo Feature Selection (MCFS) method. A feature list was thus obtained. Then, the incremental feature selection (IFS) method adopted such list to extract essential mutation features related to 27 cancer types, find out 207 mutation rules and construct efficient classifiers. The top 37 mutation features corresponding to different cancer types were discussed. All the qualitatively analyzed gene mutation features contribute to the distinction of different types of cancers, and most of such mutation rules are supported by recent literature. Therefore, our computational investigation could identify potential biomarkers and prediction rules for cancers in the mutation signature level.

Machine Learning Assessment of Early Life Factors Predicting Suicide Attempt in Adolescence or Young Adulthood

IMPORTANCE

Although longitudinal studies have reported associations between early life factors (ie, in-utero/perinatal/infancy) and long-term suicidal behavior, they have concentrated on 1 or few selected factors, and established associations, but did not investigate if early-life factors predict suicidal behavior.

OBJECTIVE

To identify and evaluate the ability of early-life factors to predict suicide attempt in adolescents and young adults from the general population.

DESIGN, SETTING, AND PARTICIPANTS

This prognostic study used data from the Québec Longitudinal Study of Child Development, a population-based longitudinal study from Québec province, Canada. Participants were followed-up from birth to age 20 years. Random forest classification algorithms were developed to predict suicide attempt. To avoid overfitting, prediction performance indices were assessed across 50 randomly split subsamples, and then the mean was calculated. Data were analyzed from November 2019 to June 2020.

EXPOSURES

Factors considered in the analysis included 150 variables, spanning virtually all early life domains, including pregnancy and birth information; child, parents, and neighborhood characteristics; parenting and family functioning; parents' mental health; and child temperament, as assessed by mothers, fathers, and hospital birth records.

MAIN OUTCOMES AND MEASURES

The main outcome was self-reported suicide attempt by age 20 years.

RESULTS

Among 1623 included youths aged 20 years, 845 (52.1%) were female and 778 (47.9%) were male. Models show moderate prediction performance. The areas under the curve for the prediction of suicide attempt were 0.72 (95% CI, 0.71-0.73) for females and 0.62 (95% CI, 0.60-0.62) for males. The models showed low sensitivity (females, 0.50; males, 0.32), moderate positive predictive values (females, 0.60; males, 0.62), and good specificity (females, 0.76; males, 0.82) and negative predicted values (females, 0.75; males, 0.71). The most important factors contributing to the prediction included socioeconomic and demographic characteristics of the family (eg, mother and father education and age, socioeconomic status, neighborhood characteristics), parents' psychological state (specifically parents' antisocial behaviors) and parenting practices. Birth-related variables also contributed to the prediction of suicidal behavior (eg, prematurity). Sex differences were also identified, with family-related socioeconomic and demographic characteristics being the top factors for females and parents' antisocial behavior being the top factor for males.

CONCLUSIONS AND RELEVANCE

These findings suggest that early life factors contributed modestly to the prediction of suicidal behavior in adolescence and young adulthood. Although these factors may inform the understanding of the etiological processes of suicide, their utility in the long-term prediction of suicide attempt was limited.

Identifying Discriminative Biological Function Features and Rules for Cancer-Related Long Non-coding RNAs

Cancer has been a major public health problem worldwide for many centuries. Cancer is a complex disease associated with accumulative genetic mutations, epigenetic aberrations, chromosomal instability, and expression alteration. Increasing lines of evidence suggest that many non-coding transcripts, which are termed as non-coding RNAs, have important regulatory roles in cancer. In particular, long non-coding RNAs (lncRNAs) play crucial roles in tumorigenesis. Cancer-related lncRNAs serve as oncogenic factors or tumor suppressors. Although many lncRNAs are identified as potential regulators in tumorigenesis by using traditional experimental methods, they are time consuming and expensive considering the tremendous amount of lncRNAs needed. Thus, effective and fast approaches to recognize tumor-related lncRNAs should be developed. The proposed approach should help us understand not only the mechanisms of lncRNAs that participate in tumorigenesis but also their satisfactory performance in distinguishing cancer-related lncRNAs. In this study, we utilized a decision tree (DT), a type of rule learning algorithm, to investigate cancer-related lncRNAs with functional annotation contents [gene ontology (GO) terms and KEGG pathways] of their co-expressed genes. Cancer-related and other lncRNAs encoded by the key enrichment features of GO and KEGG filtered by feature selection methods were used to build an informative DT, which further induced several decision rules. The rules provided not only a new tool for identifying cancer-related lncRNAs but also connected the lncRNAs and cancers with the combinations of GO terms. Results provided new directions for understanding cancer-related lncRNAs.

Identification of Latent Oncogenes with a Network Embedding Method and Random Forest

Oncogene is a special type of genes, which can promote the tumor initiation. Good study on oncogenes is helpful for understanding the cause of cancers. Experimental techniques in early time are quite popular in detecting oncogenes. However, their defects become more and more evident in recent years, such as high cost and long time. The newly proposed computational methods provide an alternative way to study oncogenes, which can provide useful clues for further investigations on candidate genes. Considering the limitations of some previous computational methods, such as lack of learning procedures and terming genes as individual subjects, a novel computational method was proposed in this study. The method adopted the features derived from multiple protein networks, viewing proteins in a system level. A classic machine learning algorithm, random forest, was applied on these features to capture the essential characteristic of oncogenes, thereby building the prediction model. All genes except validated oncogenes were ranked with a measurement yielded by the prediction model. Top genes were quite different from potential oncogenes discovered by previous methods, and they can be confirmed to become novel oncogenes. It was indicated that the newly identified genes can be essential supplements for previous results.

Classification of Widely and Rarely Expressed Genes with Recurrent Neural Network

A tissue-specific gene expression shapes the formation of tissues, while gene expression changes reflect the immune response of the human body to environmental stimulations or pressure, particularly in disease conditions, such as cancers. A few genes are commonly expressed across tissues or various cancers, while others are not. To investigate the functional differences between widely and rarely expressed genes, we defined the genes that were expressed in 32 normal tissues/cancers (i.e., called widely expressed genes; FPKM >1 in all samples) and those that were not detected (i.e., called rarely expressed genes; FPKM <1 in all samples) based on the large gene expression data set provided by Uhlen et al. Each gene was encoded using the gene ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment scores. Minimum redundancy maximum relevance (mRMR) was used to measure and rank these features on the mRMR feature list. Thereafter, we applied the incremental feature selection method with a supervised classifier recurrent neural network (RNN) to select the discriminate features for classifying widely expressed genes from rarely expressed genes and construct an optimum RNN classifier. The Youden's indexes generated by the optimum RNN classifier and evaluated using a 10-fold cross validation were 0.739 for normal tissues and 0.639 for cancers. Furthermore, the underlying mechanisms of the key discriminate GO and KEGG features were analyzed. Results can facilitate the identification of the expression landscape of genes and elucidation of how gene expression shapes tissues and the microenvironment of cancers.

•

Some genes are widely expressed across tissues or various cancers.

•

A number of genes are rarely expressed across tissues or various cancers.

•

The functional differences between widely and rarely expressed genes were studied.

•

Several GO terms and KEGG pathways were extracted and analyzed.

Comparing miRNA structure of mirtrons and non-mirtrons

BACKGROUND

MicroRNAs proceeds through the different canonical and non-canonical pathways; the most frequent of the non-canonical ones is the splicing-dependent biogenesis of mirtrons. We compare the mirtrons and non-mirtrons of human and mouse to explore how their maturation appears in the precursor structure around the miRNA.

RESULTS

We found the coherence of the overhang lengths what indicates the dependence between the cleavage sites. To explain this dependence we suggest the 2-lever model of the Dicer structure that couples the imprecisions in Drosha and Dicer. Considering the secondary structure of all animal pre-miRNAs we confirmed that single-stranded nucleotides tend to be located near the miRNA boundaries and in its center and are characterized by a higher mutation rate. The 5' end of the canonical 5' miRNA approaches the nearest single-stranded nucleotides what suggests the extension of the loop-counting rule from the Dicer to the Drosha cleavage site. A typical structure of the annotated mirtron pre-miRNAs differs from the canonical pre-miRNA structure and possesses the 1- and 2 nt hanging ends at the hairpin base. Together with the excessive variability of the mirtron Dicer cleavage site (that could be partially explained by guanine at its ends inherited from splicing) this is one more evidence for the 2-lever model. In contrast with the canonical miRNAs the mirtrons have higher snp densities and their pre-miRNAs are inversely associated with diseases. Therefore we supported the view that mirtrons are under positive selection while canonical miRNAs are under negative one and we suggested that mirtrons are an intrinsic source of silencing variability which produces the disease-promoting variants. Finally, we considered the interference of the pre-miRNA structure and the U2snRNA:pre-mRNA basepairing. We analyzed the location of the branchpoints and found that mirtron structure tends to expose the branchpoint site what suggests that the mirtrons can readily evolve from occasional hairpins in the immediate neighbourhood of the 3' splice site.

CONCLUSION

The miRNA biogenesis manifests itself in the footprints of the secondary structure. Close inspection of these structural properties can help to uncover new pathways of miRNA biogenesis and to refine the known miRNA data, in particular, new non-canonical miRNAs may be predicted or the known miRNAs can be re-classified.