Editorial: Artificial Intelligence and Bioinformatics Applications for Omics and Multi-Omics Studies

Angelo Facchiano^1* Dominik Heider² Margherita Mutarelli³

• ¹Institute of Food Sciences, National Research Council (CNR), Italy
• ²Department of Mathematics & Computer Science, University of Marburg, Marburg, Germany, Germany
• ³Institute of Applied Sciences and Intelligent Systems, Department of Physical Sciences and Technologies of Matter, National Research Council (CNR), Italy

models for understanding the connections between omics, thereby aiding in systematic generation of mechanistic hypotheses in cancer biology.Liu et al. (https://www.frontiersin.org/articles/10.3389/fgene.2022.1090394/full) introduced a novel model for classifying gastric cancer subtypes. It utilizes a residual graph convolutional network, combining multi-omics data and patient similarity networks. The study demonstrates that the new approach significantly outperforms traditional methods in predictive performance. This approach offers potential advancements in understanding gastric cancer subtypes and could assist in developing more targeted treatments.Luo et al (https://www.frontiersin.org/articles/10.3389/fgene.2023.1109269/full) presented SupCAM, a method that improves the identification of chromosome clusters for karyotyping. The new approach involves pre-training the backbone network with supervised contrastive learning on ChrCluster, incorporating variable image composition by category, and introducing self-marginal loss. Fine-tuning the network results in a final model, with SupCAM achieving a 94.99% accuracy on the ChrCluster dataset, surpassing previous methods.Zheng et al (https://www.frontiersin.org/articles/10.3389/fgene.2023.1133775/full) proposed a machine learning method for predicting if pairs of enhancers and promoters physically interact. They built a model called HARD from the names of the four (epi)genomic signals included: the histone modification H3K27ac and ATAC-seq to represent chromatin accessibility, RAD21 subunit of cohesin that is important in loop formation and the Distance between the promoter and the enhancer and classify them using a random forest algorithm. The method was tested on enhancer-promoter interaction benchmarks from the BENGI database (Moore, 2020) and compared with two existing methods outperforming them in the majority of measures, thus proving to be a useful new approach to this important although complex task.
(https://www.frontiersin.org/articles/10.3389/fgene.2023.1215232/full) presented article focused on gene expression analysis, using p-values to identify significant genes, gene ontology terms and similarity scores to understand biological pathways, regulation, and gene networks, and machine learning for gene prioritization. The study proposes using deep neural network algorithms for gene clustering based on regulatory pathways. The work validates findings through the detection of genetic interactions. Specific tissues with normalized gene expression and occurrence frequencies are considered, particularly in the context of glomerular diseases. The results highlight the relevance of genes like EGR1, IL33, BMP2, and SLAMF8 in glomerular diseases. described the development of an AI-based R package called MBMethPred to classify childhood medulloblastoma (MB) subtypes from DNA methylation and gene expression data. The two data types were combined using a similarity network fusion approach and feature selection was performed with random forests. The authors then applied six different machine-learning algorithms for subtype predictions, all scoring very good with a variety of performance measures and the selected biomarkers were challenged for biological and clinical relevance using survival and network analysis. The study represents a useful advancement towards the goal of accurate classification of molecular subgroups in MB patients that are vital to choose the best therapeutical plans for them.
Klinkhammer et al. https://www.frontiersin.org/articles/10.3389/fgene.2022.1076440/full) presented an article describing the development of a boosting algorithm, called snpboost, for creating polygenic risk scores (PRS) directly from genetic data, with the aim of improving predictive accuracy in clinical risk stratification. The approach efficiently addresses the highdimensional nature of genotype data and outperforms other methods in terms of predictive performance and computational time.Klau et al. (https://www.frontiersin.org/articles/10.3389/fgene.2023.1217860/full) focused on improving disease risk prediction using polygenic risk scores (PRS). They investigated whether incorporating multiple PRS from different diseases and applying machine learning models can enhance predictive accuracy compared to traditional single-PRS models using regression. Their results show that multi-PRS models, especially when combined with deep learning techniques, significantly outperform single-PRS models in predicting risks for diseases like cancer, diabetes, and cardiovascular diseases. This advancement could lead to more effective disease prediction and personalized medicine approaches.
Waechter et al. (https://www.frontiersin.org/articles/10.3389/fgene.2023.1213829/full) investigated the effectiveness of two different 16S rRNA primer sets for sequencing human fecal microbiomes using the Nanopore platform. They compared the conventional 27F primer included in the 16S Barcoding Kit by Oxford Nanopore Technologies and a more degenerate 27F primer. The study reveals significant differences in the detection of taxonomic diversity and relative abundance of various taxa between these primer sets. The more degenerate primer set appears to provide a more accurate and diverse representation of the fecal microbiome compared to the conventional primer set.