Optimizing hybrid ensemble feature selection strategies for transcriptomic biomarker discovery in complex diseases.

Biomedical research takes advantage of omic data, such as transcriptomics, to unravel the complexity of diseases. A conventional strategy identifies transcriptomic biomarkers characterized by expression patterns associated with a phenotype by relying on feature selection approaches. Hybrid ensemble feature selection (HEFS) has become increasingly popular as it ensures robustness of the selected features by performing data and functional perturbations. However, it remains difficult to make the best suited choices at each step when designing such approaches. We conducted an extensive analysis of four possible HEFS scenarios for the identification of Stage IV colorectal, Stage I kidney and lung and Stage III endometrial cancer biomarkers from transcriptomic data. These scenarios investigate the use of two types of feature reduction by filters (differentially expressed genes and variance) conjointly with two types of resampling strategies (repeated holdout by distribution-balanced stratified and random stratified) for downstream feature selection through an aggregation of thousands of wrapped machine learning models. Based on our results, we emphasize the advantages of using HEFS approaches to identify complex disease biomarkers, given their ability to produce generalizable and stable results to both data and functional perturbations. Finally, we highlight critical issues that need to be considered in the design of such strategies.

Target repositioning using multi-layer networks and machine learning: The case of prostate cancer.

The discovery of novel therapeutic targets, defined as proteins which drugs can interact with to induce therapeutic benefits, typically represent the first and most important step of drug discovery. One solution for target discovery is target repositioning, a strategy which relies on the repurposing of known targets for new diseases, leading to new treatments, less side effects and potential drug synergies. Biological networks have emerged as powerful tools for integrating heterogeneous data and facilitating the prediction of biological or therapeutic properties. Consequently, they are widely employed to predict new therapeutic targets by characterizing potential candidates, often based on their interactions within a Protein-Protein Interaction (PPI) network, and their proximity to genes associated with the disease. However, over-reliance on PPI networks and the assumption that potential targets are necessarily near known genes can introduce biases that may limit the effectiveness of these methods. This study addresses these limitations in two ways. First, by exploiting a multi-layer network which incorporates additional information such as gene regulation, metabolite interactions, metabolic pathways, and several disease signatures such as Differentially Expressed Genes, mutated genes, Copy Number Alteration, and structural variants. Second, by extracting relevant features from the network using several approaches including proximity to disease-associated genes, but also unbiased approaches such as propagation-based methods, topological metrics, and module detection algorithms. Using prostate cancer as a case study, the best features were identified and utilized to train machine learning algorithms to predict 5 novel promising therapeutic targets for prostate cancer: IGF2R, C5AR, RAB7, SETD2 and NPBWR1.

3D microscopic reconstruction of pearls using combined optical microscopy and photogrammetry.

In this study, we introduce an affordable and accessible method that combines optical microscopy and photogrammetry to reconstruct 3D models of Tahitian pearls. We present a novel device designed for acquiring microscopic images around a sphere using translational displacement stages and outline our method for reconstructing these images. We successfully created 3D models of two individual pearl rings, each representing 6.3% of the pearl's surface. Additionally, we generated a combined model representing 10.3% of the pearl's surface. This showcases the potential for reconstructing entire pearls with appropriate instrumentation. We emphasize that our approach extends beyond pearls and spherical objects and can be adapted for various object types using appropriate acquisition devices. We provide a proof of concept demonstrating the feasibility of 3D photogrammetry using optical microscopy. Consequently, our method offers a practical and cost-effective alternative for generating 3D models at a microscopic scale, particularly when detailed internal structure information is unnecessary.

A Machine Learning Approach to Identify Key Residues Involved in Protein-Protein Interactions Exemplified with SARS-CoV-2 Variants.

Human infection with the coronavirus disease 2019 (COVID-19) is mediated by the binding of the spike protein of the severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) to the human angiotensin-converting enzyme 2 (ACE2). The frequent mutations in the receptor-binding domain (RBD) of the spike protein induced the emergence of variants with increased contagion and can hinder vaccine efficiency. Hence, it is crucial to better understand the binding mechanisms of variant RBDs to human ACE2 and develop efficient methods to characterize this interaction. In this work, we present an approach that uses machine learning to analyze the molecular dynamics simulations of RBD variant trajectories bound to ACE2. Along with the binding free energy calculation, this method was used to characterize the major differences in ACE2-binding capacity of three SARS-CoV-2 RBD variants-namely the original Wuhan strain, Omicron BA.1, and the more recent Omicron BA.5 sublineages. Our analyses assessed the differences in binding free energy and shed light on how it affects the infectious rates of different variants. Furthermore, this approach successfully characterized key binding interactions and could be deployed as an efficient tool to predict different binding inhibitors to pave the way for new preventive and therapeutic strategies.

BERNN: Enhancing classification of Liquid Chromatography Mass Spectrometry data with batch effect removal neural networks.

Liquid Chromatography Mass Spectrometry (LC-MS) is a powerful method for profiling complex biological samples. However, batch effects typically arise from differences in sample processing protocols, experimental conditions, and data acquisition techniques, significantly impacting the interpretability of results. Correcting batch effects is crucial for the reproducibility of omics research, but current methods are not optimal for the removal of batch effects without compressing the genuine biological variation under study. We propose a suite of Batch Effect Removal Neural Networks (BERNN) to remove batch effects in large LC-MS experiments, with the goal of maximizing sample classification performance between conditions. More importantly, these models must efficiently generalize in batches not seen during training. A comparison of batch effect correction methods across five diverse datasets demonstrated that BERNN models consistently showed the strongest sample classification performance. However, the model producing the greatest classification improvements did not always perform best in terms of batch effect removal. Finally, we show that the overcorrection of batch effects resulted in the loss of some essential biological variability. These findings highlight the importance of balancing batch effect removal while preserving valuable biological diversity in large-scale LC-MS experiments.

Combined transcriptomics and proteomics unveil the impact of vitamin C in modulating specific protein abundance in the mouse liver.

BACKGROUND: Vitamin C (ascorbate) is a water-soluble antioxidant and an important cofactor for various biosynthetic and regulatory enzymes. Mice can synthesize vitamin C thanks to the key enzyme gulonolactone oxidase (Gulo) unlike humans. In the current investigation, we used Gulo-/- mice, which cannot synthesize their own ascorbate to determine the impact of this vitamin on both the transcriptomics and proteomics profiles in the whole liver. The study included Gulo-/- mouse groups treated with either sub-optimal or optimal ascorbate concentrations in drinking water. Liver tissues of females and males were collected at the age of four months and divided for transcriptomics and proteomics analysis. Immunoblotting, quantitative RT-PCR, and polysome profiling experiments were also conducted to complement our combined omics studies. RESULTS: Principal component analyses revealed distinctive differences in the mRNA and protein profiles as a function of sex between all the mouse cohorts. Despite such sexual dimorphism, Spearman analyses of transcriptomics data from females and males revealed correlations of hepatic ascorbate levels with transcripts encoding a wide array of biological processes involved in glucose and lipid metabolisms as well as in the acute-phase immune response. Moreover, integration of the proteomics data showed that ascorbate modulates the abundance of various enzymes involved in lipid, xenobiotic, organic acid, acetyl-CoA, and steroid metabolism mainly at the transcriptional level, especially in females. However, several proteins of the mitochondrial complex III significantly correlated with ascorbate concentrations in both males and females unlike their corresponding transcripts. Finally, poly(ribo)some profiling did not reveal significant enrichment difference for these mitochondrial complex III mRNAs between Gulo-/- mice treated with sub-optimal and optimal ascorbate levels. CONCLUSIONS: Thus, the abundance of several subunits of the mitochondrial complex III are regulated by ascorbate at the post-transcriptional levels. Our extensive omics analyses provide a novel resource of altered gene expression patterns at the transcriptional and post-transcriptional levels under ascorbate deficiency.

SARS-CoV-2 infection modifies the transcriptome of the megakaryocytes in the bone marrow.

ABSTRACT: Megakaryocytes (MKs), integral to platelet production, predominantly reside in the bone marrow (BM) and undergo regulated fragmentation within sinusoid vessels to release platelets into the bloodstream. Inflammatory states and infections influence MK transcription, potentially affecting platelet functionality. Notably, COVID-19 has been associated with altered platelet transcriptomes. In this study, we investigated the hypothesis that severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) infection could affect the transcriptome of BM MKs. Using spatial transcriptomics to discriminate subpopulations of MKs based on proximity to BM sinusoids, we identified ∼19 000 genes in MKs. Machine learning techniques revealed that the transcriptome of healthy murine BM MKs exhibited minimal differences based on proximity to sinusoid vessels. Furthermore, at peak SARS-CoV-2 viremia, when the disease primarily affected the lungs, MKs were not significantly different from those from healthy mice. Conversely, a significant divergence in the MK transcriptome was observed during systemic inflammation, although SARS-CoV-2 RNA was never detected in the BM, and it was no longer detectable in the lungs. Under these conditions, the MK transcriptional landscape was enriched in pathways associated with histone modifications, MK differentiation, NETosis, and autoimmunity, which could not be explained by cell proximity to sinusoid vessels. Notably, the type I interferon signature and calprotectin (S100A8/A9) were not induced in MKs under any condition. However, inflammatory cytokines induced in the blood and lungs of COVID-19 mice were different from those found in the BM, suggesting a discriminating impact of inflammation on this specific subset of cells. Collectively, our data indicate that a new population of BM MKs may emerge through COVID-19-related pathogenesis.

Predictive model for vitamin C levels in hyperinsulinemic individuals based on age, sex, waist circumference, low-density lipoprotein, and immune-associated serum proteins.

Vitamin C (ascorbic acid) is an important water-soluble antioxidant associated with decreased oxidative stress in type 2 diabetes (T2D) patients. A previous targeted plasma proteomic study has indicated that ascorbic acid is associated with markers of the immune system in healthy subjects. However, the association between the levels of ascorbic acid and blood biomarkers in subjects at risk of developing T2D is still unknown. Serum ascorbic acid was measured by ultra-performance liquid chromatography and serum proteins were quantified by untargeted liquid-chromatography mass spectrometry in 25 hyperinsulinemia subjects that were randomly assigned a high dairy intake diet or an adequate dairy intake diet for 6 weeks, then crossed-over after a 6-week washout period. Spearman correlation followed by gene ontology analyses were performed to identify biological pathways associated with ascorbic acid. Finally, machine learning analysis was performed to obtain a specific serum protein signature that could predict ascorbic acid levels. After adjustments for waist circumference, LDL, HDL, fasting insulin, fasting blood glucose, age, gender, and dairy intake; serum ascorbic acid correlated positively with different aspects of the immune system. Machine learning analysis indicated that a signature composed of 21 features that included 17 proteins (mainly from the immune system), age, sex, waist circumference, and LDL could predict serum ascorbic acid levels in hyperinsulinemia subjects. In conclusion, the result reveals a correlation as well as modulation between serum ascorbic acid levels and proteins that play vital roles in regulating different aspects of the immune response in individuals at risk of T2D. The development of a predictive signature for ascorbic acid will further help the assessment of ascorbic acid status in clinical settings.

BioDiscViz: A visualization support and consensus signature selector for BioDiscML results.

Machine learning (ML) algorithms are powerful tools to find complex patterns and biomarker signatures when conventional statistical methods fail to identify them. While the ML field made significant progress, state of the art methodologies to build efficient and non-overfitting models are not always applied in the literature. To this purpose, automatic programs, such as BioDiscML, were designed to identify biomarker signatures and correlated features while escaping overfitting using multiple evaluation strategies, such as cross validation, bootstrapping and repeated holdout. To further improve BioDiscML and reach a broader audience, better visualization support and flexibility in choosing the best models and signatures are needed. Thus, to provide researchers with an easily accessible and usable tool for in depth investigation of the results from BioDiscML outputs, we developed a visual interaction tool called BioDiscViz. This tool provides summaries, tables and graphics, in the form of Principal Component Analysis (PCA) plots, UMAP, t-SNE, heatmaps and boxplots for the best model and the correlated features. Furthermore, this tool also provides visual support to extract a consensus signature from BioDiscML models using a combination of filters. BioDiscViz will be a great visual support for research using ML, hence new opportunities in this field by opening it to a broader community.

Developing a cluster-based approach for deciphering complexity in individuals with neurodevelopmental differences.

Objective: Individuals with neurodevelopmental disorders such as global developmental delay (GDD) present both genotypic and phenotypic heterogeneity. This diversity has hampered developing of targeted interventions given the relative rarity of each individual genetic etiology. Novel approaches to clinical trials where distinct, but related diseases can be treated by a common drug, known as basket trials, which have shown benefits in oncology but have yet to be used in GDD. Nonetheless, it remains unclear how individuals with GDD could be clustered. Here, we assess two different approaches: agglomerative and divisive clustering. Methods: Using the largest cohort of individuals with GDD, which is the Deciphering Developmental Disorders (DDD), characterized using a systematic approach, we extracted genotypic and phenotypic information from 6,588 individuals with GDD. We then used a k-means clustering (divisive) and hierarchical agglomerative clustering (HAC) to identify subgroups of individuals. Next, we extracted gene network and molecular function information with regard to the clusters identified by each approach. Results: HAC based on phenotypes identified in individuals with GDD revealed 16 clusters, each presenting with one dominant phenotype displayed by most individuals in the cluster, along with other minor phenotypes. Among the most common phenotypes reported were delayed speech, absent speech, and seizure. Interestingly, each phenotypic cluster molecularly included several (3-12) gene sub-networks of more closely related genes with diverse molecular function. k-means clustering also segregated individuals harboring those phenotypes, but the genetic pathways identified were different from the ones identified from HAC. Conclusion: Our study illustrates how divisive (k-means) and agglomerative clustering can be used in order to group individuals with GDD for future basket trials. Moreover, the result of our analysis suggests that phenotypic clusters should be subdivided into molecular sub-networks for an increased likelihood of successful treatment. Finally, a combination of both agglomerative and divisive clustering may be required for developing of a comprehensive treatment.

Sex difference contributes to phenotypic diversity in individuals with neurodevelopmental disorders.

Objective: Gain a better understanding of sex-specific differences in individuals with global developmental delay (GDD), with a focus on phenotypes and genotypes. Methods: Using the Deciphering Developmental Disorders (DDD) dataset, we extracted phenotypic information from 6,588 individuals with GDD and then identified statistically significant variations in phenotypes and genotypes based on sex. We compared genes with pathogenic variants between sex and then performed gene network and molecular function enrichment analysis and gene expression profiling between sex. Finally, we contrasted individuals with autism as an associated condition. Results: We identified significantly differentially expressed phenotypes in males vs. females individuals with GDD. Autism and macrocephaly were significantly more common in males whereas microcephaly and stereotypies were more common in females. Importantly, 66% of GDD genes with pathogenic variants overlapped between both sexes. In the cohort, males presented with only slightly increased X-linked genes (9% vs. 8%, respectively). Individuals from both sexes harbored a similar number of pathogenic variants overall (3) but females presented with a significantly higher load for GDD genes with high intolerance to loss of function. Sex difference in gene expression correlated with genes identified in a sex specific manner. While we identified sex-specific GDD gene mutations, their pathways overlapped. Interestingly, individuals with GDD but also co-morbid autism phenotypes, we observed distinct mutation load, pathways and phenotypic presentation. Conclusion: Our study shows for the first time that males and females with GDD present with significantly different phenotypes. Moreover, while most GDD genes overlapped, some genes were found uniquely in each sex. Surprisingly they shared similar molecular functions. Sorting genes by predicted tolerance to loss of function (pLI) led to identifying an increased mutation load in females with GDD, suggesting potentially a tolerance to GDD genes of higher pLI compared to overall GDD genes. Finally, we show that considering associated conditions (for instance autism) may influence the genomic underpinning found in individuals with GDD and highlight the importance of comprehensive phenotyping.

Pearl shape classification using deep convolutional neural networks from Tahitian pearl rotation in Pinctada margaritifera.

Tahitian pearls, artificially cultivated from the black-lipped pearl oyster Pinctada margaritifera, are renowned for their unique color and large size, making the pearl industry vital for the French Polynesian economy. Understanding the mechanisms of pearl formation is essential for enabling quality and sustainable production. In this paper, we explore the process of pearl formation by studying pearl rotation. Here we show, using a deep convolutional neural network, a direct link between the rotation of the pearl during its formation in the oyster and its final shape. We propose a new method for non-invasive pearl monitoring and a model for predicting the final shape of the pearl from rotation data with 81.9% accuracy. These novel resources provide a fresh perspective to study and enhance our comprehension of the overall mechanism of pearl formation, with potential long-term applications for improving pearl production and quality control in the industry.

Design of experiment and machine learning inform on the 3D printing of hydrogels for biomedical applications.

In the biomedical field, 3D printing has the potential to deliver on some of the promises of personalized therapy, notably by enabling point-of-care fabrication of medical devices, dosage forms and bioimplants. To achieve this full potential, a better understanding of the 3D printing processes is necessary, and non-destructive characterization methods must be developed. This study proposes methodologies to optimize the 3D printing parameters for soft material extrusion. We hypothesize that combining image processing with design of experiment (DoE) analyses and machine learning could help obtaining useful information from a quality-by-design perspective. Herein, we investigated the impact of three critical process parameters (printing speed, printing pressure and infill percentage) on three critical quality attributes (gel weight, total surface area and heterogeneity) monitored with a non-destructive methodology. DoE and machine learning were combined to obtain information on the process. This work paves the way for a rational approach to optimize 3D printing parameters in the biomedical field.

Long-term Memory Testing in Children With Typical Development and Neurodevelopmental Disorders: Remote Web-based Image Task Feasibility Study.

BACKGROUND: Neurodevelopmental disorders (NDD) cause individuals to have difficulty in learning facts, procedures, or social skills. NDD has been linked to several genes, and several animal models have been used to identify potential therapeutic candidates based on specific learning paradigms for long-term and associative memory. In individuals with NDD, however, such testing has not been used so far, resulting in a gap in translating preclinical results to clinical practice. OBJECTIVE: We aim to assess if individuals with NDD could be tested for paired association learning and long-term memory deficit, as shown in previous animal models. METHODS: We developed an image-based paired association task, which can be performed at different time points using remote web-based testing, and evaluated its feasibility in children with typical development (TD), as well as NDD. We included 2 tasks: object recognition as a simpler task and paired association. Learning was tested immediately after training and also the next day for long-term memory. RESULTS: We found that children aged 5-14 years with TD (n=128) and with NDD of different types (n=57) could complete testing using the Memory Game. Children with NDD showed deficits in both recognition and paired association tasks on the first day of learning, in both 5-9-year old (P<.001 and P=.01, respectively) and 10-14-year old groups (P=.001 and P<.001, respectively). The reaction times to stimuli showed no significant difference between individuals with TD or NDD. Children with NDD exhibited a faster 24-hour memory decay for the recognition task than those with TD in the 5-9-year old group. This trend is reversed for the paired association task. Interestingly, we found that children with NDD had their retention for recognition improved and matched with typically developing individuals by 10-14 years of age. The NDD group also showed improved retention deficits in the paired association task at 10-14 years of age compared to the TD group. CONCLUSIONS: We showed that web-based learning testing using simple picture association is feasible for children with TD, as well as with NDD. We showed how web-based testing allows us to train children to learn the association between pictures, as shown in immediate test results and those completed 1 day after. This is important as many models for learning deficits in NDD target both short- and long-term memory for therapeutic intervention. We also demonstrated that despite potential confounding factors, such as self-reported diagnosis bias, technical issues, and varied participation, the Memory Game shows significant differences between typically developing children and those with NDD. Future experiments will leverage this potential of web-based testing for larger cohorts and cross-validation with other clinical or preclinical cognitive tasks.

Industrial and Ruminant Trans-Fatty Acids-Enriched Diets Differentially Modulate the Microbiome and Fecal Metabolites in C57BL/6 Mice.

Industrially originated trans-fatty acids (I-tFAs), such as elaidic acid (EA), and ruminant trans-fatty acids (R-tFAs), such as trans-palmitoleic acid (TPA), may have opposite effects on metabolic health. The objective was to compare the effects of consuming 2-3% I-tFA or R-tFA on the gut microbiome and fecal metabolite profile in mice after 7 and 28 days. Forty C57BL/6 mice were assigned to one of the four prepared formulations: lecithin nanovesicles, lecithin nanovesicles with EA or TPA, or water. Fecal samples and animals' weights were collected on days 0, 7, and 28. Fecal samples were used to determine gut microbiome profiles by 16S rRNA sequencing and metabolite concentrations by GC/MS. At 28 days, TPA intake decreased the abundance of Staphylococcus sp55 but increased Staphylococcus sp119. EA intake also increased the abundance of Staphylococcus sp119 but decreased Ruminococcaceae UCG-014, Lachnospiraceae, and Clostridium sensu stricto 1 at 28 days. Fecal short-chain fatty acids were increased after TPA while decreased after EA after 7 and 28 days. This study shows that TPA and EA modify the abundance of specific microbial taxa and fecal metabolite profiles in distinct ways.

Overview of methods for characterization and visualization of a protein-protein interaction network in a multi-omics integration context.

At the heart of the cellular machinery through the regulation of cellular functions, protein-protein interactions (PPIs) have a significant role. PPIs can be analyzed with network approaches. Construction of a PPI network requires prediction of the interactions. All PPIs form a network. Different biases such as lack of data, recurrence of information, and false interactions make the network unstable. Integrated strategies allow solving these different challenges. These approaches have shown encouraging results for the understanding of molecular mechanisms, drug action mechanisms, and identification of target genes. In order to give more importance to an interaction, it is evaluated by different confidence scores. These scores allow the filtration of the network and thus facilitate the representation of the network, essential steps to the identification and understanding of molecular mechanisms. In this review, we will discuss the main computational methods for predicting PPI, including ones confirming an interaction as well as the integration of PPIs into a network, and we will discuss visualization of these complex data.

DIA Proteomics and Machine Learning for the Fast Identification of Bacterial Species in Biological Samples.

Identification of bacterial species in biological samples is essential in many applications. However, the standard methods usually use a time-consuming bacterial culture (24-48 h) and sometimes lack in specificity. To overcome these limitations, we developed a new protocol, combining LC-MS/MS analysis in Data Independent Acquisition mode and machine learning algorithms, enabling the accurate identification of the bacterial species contaminating a sample in a few hours without bacterial culture. In this chapter, we describe the three steps of the protocol (spectral libraries generation, training step, identification step) to generate customized peptide signatures and use them for bacterial identification in biological samples through targeted proteomics analyses and prediction models.

Machine Learning and Deep Learning Applications in Metagenomic Taxonomy and Functional Annotation.

Shotgun sequencing of environmental DNA (i.e., metagenomics) has revolutionized the field of environmental microbiology, allowing the characterization of all microorganisms in a sequencing experiment. To identify the microbes in terms of taxonomy and biological activity, the sequenced reads must necessarily be aligned on known microbial genomes/genes. However, current alignment methods are limited in terms of speed and can produce a significant number of false positives when detecting bacterial species or false negatives in specific cases (virus, plasmids, and gene detection). Moreover, recent advances in metagenomics have enabled the reconstruction of new genomes using de novo binning strategies, but these genomes, not yet fully characterized, are not used in classic approaches, whereas machine and deep learning methods can use them as models. In this article, we attempted to review the different methods and their efficiency to improve the annotation of metagenomic sequences. Deep learning models have reached the performance of the widely used k-mer alignment-based tools, with better accuracy in certain cases; however, they still must demonstrate their robustness across the variety of environmental samples and across the rapid expansion of accessible genomes in databases.

DIA proteomics data from a UPS1-spiked E.coli protein mixture processed with six software tools.

In this article, we provide a proteomic reference dataset that has been initially generated for a benchmarking of software tools for Data-Independent Acquisition (DIA) analysis. This large dataset includes 96 DIA .raw files acquired from a complex proteomic standard composed of an E.coli protein background spiked-in with 8 different concentrations of 48 human proteins (UPS1 Sigma). These 8 samples were analyzed in triplicates on an Orbitrap mass spectrometer with 4 different DIA window schemes. We also provide the spectral libraries and FASTA file used for their analysis and the software outputs of the six tools used in this study: DIA-NN, Spectronaut, ScaffoldDIA, DIA-Umpire, Skyline and OpenSWATH. This dataset also contains post-processed quantification tables where the peptides and proteins have been validated, their intensities normalized and the missing values imputed with a noise value. All the files are available on ProteomeXchange. Altogether, these files represent the most comprehensive DIA reference dataset acquired on an Orbitrap instrument ever published. It will be a very useful resource to the proteomic scientists in order to assess the performance of DIA software tools or to test their processing pipelines, to the software developers to improve their tools or develop new ones and to the students for their training on proteomics data analysis.

T cell apoptosis characterizes severe Covid-19 disease.

Severe SARS-CoV-2 infections are characterized by lymphopenia, but the mechanisms involved are still elusive. Based on our knowledge of HIV pathophysiology, we hypothesized that SARS-CoV-2 infection-mediated lymphopenia could also be related to T cell apoptosis. By comparing intensive care unit (ICU) and non-ICU COVID-19 patients with age-matched healthy donors, we found a strong positive correlation between plasma levels of soluble FasL (sFasL) and T cell surface expression of Fas/CD95 with the propensity of T cells to die and CD4 T cell counts. Plasma levels of sFasL and T cell death are correlated with CXCL10 which is part of the signature of 4 biomarkers of disease severity (ROC, 0.98). We also found that members of the Bcl-2 family had modulated in the T cells of COVID-19 patients. More importantly, we demonstrated that the pan-caspase inhibitor, Q-VD, prevents T cell death by apoptosis and enhances Th1 transcripts. Altogether, our results are compatible with a model in which T-cell apoptosis accounts for T lymphopenia in individuals with severe COVID-19. Therefore, a strategy aimed at blocking caspase activation could be beneficial for preventing immunodeficiency in COVID-19 patients.