The use of diverse modalities, such as omics, medical images, and clinical data can not only improve the performance of prognostic models but also deepen an understanding of disease mechanisms and facilitate the development of novel treatment approaches. However, medical data are complex, often incomplete, and contains missing modalities, making effective handling its crucial for training multimodal models. We introduce impuTMAE, a novel transformer-based end-to-end approach with an efficient multimodal pre-training strategy. It learns inter- and intra-modal interactions while simultaneously imputing missing modalities by reconstructing masked patches. Our model is pre-trained on heterogeneous, incomplete data and fine-tuned for glioma survival prediction using TCGA-GBM/LGG and BraTS datasets, integrating five modalities: genetic (DNAm, RNA-seq), imaging (MRI, WSI), and clinical data. By addressing missing data during pre-training and enabling efficient resource utilization, impuTMAE surpasses prior multimodal approaches, achieving state-of-the-art performance in glioma patient survival prediction. Our code is available at https://github.com/maryjis/mtcp

Kidney cancer has a high metastatic potential with up to 30% of patients developing distant metastasis after surgery. We assessed the value of AI models in predicting the metastatic potential of clear cell renal cell carcinoma (ccRCC), based on the genetic data. Tissue samples from patients with both metastatic and non-metastatic squamous cell carcinoma were analyzed, focusing on the expression and methylation levels of specific protein-coding (PC) and microRNA (miRNA) genes. Using quantitative PCR and data classification techniques, we found a correlation between metastasis and reduced expression of PC-genes CA9, NDUFA4L2, EGLN3, and BHLHE41, as well as increased methylation in miRNA genes MIR125B-1, MIR137, MIR375, MIR193A, and MIR34B. AI models were built for predicting distant metastases based on the expression values and methylation status of selected genes. One model is based on solving a regression problem and is non-interpretable, while another one is based on proposed decision rules and is interpretable. The quality of the models was assessed using sensitivity and specificity metrics, and cross-validation technology was used to ensure the reliability of the results.

Enterotypes of the human gut microbiome have been proposed to be a powerful prognostic tool to evaluate the correlation between lifestyle, nutrition, and disease. However, the number of enterotypes suggested in the literature ranged from two to four. The growth of available metagenome data and the use of exact, non-linear methods of data analysis challenges the very concept of clusters in the multidimensional space of bacterial microbiomes. Using several published human gut microbiome datasets of variable 16S rRNA regions, we demonstrate the presence of a lower-dimensional structure in the microbiome space, with high-dimensional data concentrated near a low-dimensional non-linear submanifold, but the absence of distinct and stable clusters that could represent enterotypes. This observation is robust with regard to diverse combinations of dimensionality reduction techniques and clustering algorithms.

The present stage of development of artificial intelligence (AI) is based on advanced technologies and modern methods and algorithms of machine learning (ML) including deep machine learning, intellectual data analysis and other fundamental scientific areas, which in turn play an important role in delivering applied solutions in almost all areas of digital economy. However, the corresponding expansion of the AI application range and complication of the class of problems, as well as of the spectrum and the volume of data used for creation of applied AI models and AI-based intelligent system call for a significant expansion of the theoretical and algorithmic basis of AI including development of ML methods that involve mathematical and physical models of objects and phenomena of domains, methods of fusion of multimodal data, methods for constructing geometric and topological components of deep neural networks, methods for modeling 3D objects, etc. In 2021, in order to cope well with these challenges, Skoltech Research Center for Applied Artificial Intelligence was created with support from the Federal Program “Artificial Intelligence.” The aims of the Center are as follows: creation of a scientific and technological basis for solution of a wide range of relevant applied tasks for purposes of sustainable development of the economy of the Russian Federation including optimization problems of managerial decisions aimed at reduction of the carbon footprint and other urgent ESG tasks; environmental monitoring tasks for detection of anomalies and extreme situations forecast; assessment of economic and social risks and their dynamics caused by climatic changes; predictive analytic problems, etc. In the present paper, we describe new AI technologies, models, methods and algorithms developed in the Center, describe the main applied areas of research of the Center, and provide a briefly survey of the already developed scientific and applied results.

Electroencephalogram (EEG) is a set of time series each of which can be represented as a 2D image (spectrogram), so that EEG recording can be mapped to the C-dimensional image (where C denotes the number of channels in the image and equals to the number of electrodes in EEG montage). In this paper, a novel approach for automated feature extraction from spectrogram representation is proposed. The method involves the usage of autoencoder models based on 3-dimensional convolution layers and 2-dimensional deformable convolution layers. Features, extracted by autoencoders, can be used to classify patients with Major Depressive Disorder (MDD) from healthy controls based on resting-state EEG. The proposed approach outperforms baseline ML models trained on spectral features extracted manually.

Mild cognitive impairment (MCI) is an important public health problem that has enormous consequences for patients, their families, the health care system, and the economy. MCI is highly likely to progress to dementia. Early premedical intervention for MCI could prevent or delay dementia's onset, which in turn would improve patients' quality of life. Thus, the goal of our research is to develop an MCI detection system using MRI-based measures. The dataset for the study consists of 62 healthy volunteers and 145 patients with MCI obtained from the «Memory Clinic» (Moscow, Russia). The morphometric data were extracted from T1-weighted structural MR images. We propose a machine learning method, which achieved an accuracy of 73%, and extract key features in brain structures that can distinguish the healthy control (HC) group from the MCI group. This study confirms the previously identified changes in the structural regions of the brain in MCI patients. Compared to the healthy control group, in the MCI group, volume increases in the Lateral Ventricle, Inferior Lateral Ventricle, and Cerebrospinal Fluid as well as volume reductions in the Amygdala, Hippocampus, and Accumbens have been found.

Disorders of systemic immunity and immune processes in the brain have now been shown to play an essential role in the development and progression of schizophrenia. Nevertheless, only a few works were devoted to the study of some immune parameters to objectify the diagnosis by means of machine learning. At the same time, machine learning methods have not yet been applied to a set of data fully reflecting systemic characteristics of the immune status (parameters of adaptive immunity, the level of inflammatory markers, the content of major cytokines). Considering a complex nature of immune system disorders in schizophrenia, incorporation of a broad panel of immunological data into machine learning models is promising for improving classification accuracy and identifying the parameters reflecting the immune disorders typical for the majority of patients. The aim of the study is to assess the possibility of using immunological parameters to objectify the diagnosis of schizophrenia applying machine learning models. Materials and Methods. We have analyzed 17 immunological parameters in 63 schizophrenia patients and 36 healthy volunteers. The parameters of humoral immunity, systemic level of the key cytokines of adaptive immunity, anti-inflammatory and pro-inflammatory cytokines, and other inflammatory markers were determined by enzyme immunoassay. Applied methods of machine learning covered the main group of approaches to supervised learning such as linear models (logistic regression), quadratic discriminant analysis (QDA), support vector machine (linear SVM, RBF SVM), k-nearest neighbors algorithm, Gaussian processes, naive Bayes classifier, decision trees, and ensemble models (AdaBoost, random forest, XGBoost). The importance of features for prediction from the best fold has been analyzed for the machine learning methods, which demonstrated the best quality. The most significant features were selected using 70% quantile threshold. Results. The AdaBoost ensemble model with ROC AUC of 0.71 +/- 0.15 and average accuracy (ACC) of 0.78 +/- 0.11 has demonstrated the best quality on a 10-fold cross validation test sample. Within the frameworks of the present investigation, the AdaBoost model has shown a good quality of classification between the patients with schizophrenia and healthy volunteers (ROC AUC over 0.70) at a high stability of the results (sigma less than 0.2). The most important immunological parameters have been established for differentiation between the patients and healthy volunteers: the level of some systemic inflammatory markers, activation of humoral immunity, pro-inflammatory cytokines, immunoregulatory cytokines and proteins, Th1 and Th2 immunity cytokines. It was for the first time that the possibility of differentiating schizophrenia patients from healthy volunteers was shown with the accuracy of more than 70% with the help of machine learning using only immune parameters. The results of this investigation confirm a high importance of the immune system in the pathogenesis of schizophrenia.

Major depressive disorder (MDD) is a common mental disorder and is amongst the most prevalent psychiatric disorders. MDD remains challenging to diagnose and predict its onset due to its heterogeneous phenotype and complex etiology. Hence, early detection using diagnostic biomarkers is critical for rapid intervention. In this study, a mixture of AI and bioinformatics were used to mine transcriptomic data from publicly available datasets including 170 MDD patients and 121 healthy controls. Bioinformatics analysis using gene set enrichment analysis (GSEA) and machine learning (ML) algorithms were applied. The GSEA revealed that differentially expressed genes in MDD patients are mainly enriched in pathways related to immune response, inflammatory response, neurodegeneration pathways and cerebellar atrophy pathways. Feature selection methods and ML provided predicted models based on MDD-altered genes with ≥75% of accuracy. The integrative analysis between the bioinformatics and ML approaches identified ten key MDD-related biomarkers including NRG1, CEACAM8, CLEC12B, DEFA4, HP, LCN2, OLFM4, SERPING1, TCN1 and THBS1. Among them, NRG1, active in synaptic plasticity and neurotransmission, was the most robust and reliable to distinguish between MDD patients and healthy controls amongst independent external datasets consisting of a mixture of populations. Further evaluation using saliva samples from an independent cohort of MDD and healthy individuals confirmed the upregulation of NRG1 in patients with MDD compared to healthy controls. Functional mapping to the human brain regions showed NRG1 to have high expression in the main subcortical limbic brain regions implicated in depression. In conclusion, integrative bioinformatics and ML approaches identified putative non-invasive diagnostic MDD-related biomarkers panel for the onset of depression.

Astrocytes respond to injury and disease in the central nervous system with reactive changes that influence the outcome of the disorder 1 – 4 . These changes include differentially expressed genes (DEGs) whose contextual diversity and regulation are poorly understood. Here we combined biological and informatic analyses, including RNA sequencing, protein detection, assay for transposase-accessible chromatin with high-throughput sequencing (ATAC-seq) and conditional gene deletion, to predict transcriptional regulators that differentially control more than 12,000 DEGs that are potentially associated with astrocyte reactivity across diverse central nervous system disorders in mice and humans. DEGs associated with astrocyte reactivity exhibited pronounced heterogeneity across disorders. Transcriptional regulators also exhibited disorder-specific differences, but a core group of 61 transcriptional regulators was identified as common across multiple disorders in both species. We show experimentally that DEG diversity is determined by combinatorial, context-specific interactions between transcriptional regulators. Notably, the same reactivity transcriptional regulators can regulate markedly different DEG cohorts in different disorders; changes in the access of transcriptional regulators to DNA-binding motifs differ markedly across disorders; and DEG changes can crucially require multiple reactivity transcriptional regulators. We show that, by modulating reactivity, transcriptional regulators can substantially alter disorder outcome, implicating them as therapeutic targets. We provide searchable resources of disorder-related reactive astrocyte DEGs and their predicted transcriptional regulators. Our findings show that transcriptional changes associated with astrocyte reactivity are highly heterogeneous and are customized from vast numbers of potential DEGs through context-specific combinatorial transcriptional-regulator interactions.

The inability of neurons to regenerate long axons within the CNS is a major impediment to improving outcome after spinal cord injury, stroke, and other CNS insults. Recent advances have uncovered an intrinsic program that involves coordinate regulation by multiple transcription factors that can be manipulated to enhance growth in the peripheral nervous system. Here, we use a systems genomics approach to characterize regulatory relationships of regeneration-associated transcription factors, identifying RE1-Silencing Transcription Factor (REST; Neuron-Restrictive Silencer Factor, NRSF) as a predicted upstream suppressor of a pro-regenerative gene program associated with axon regeneration in the CNS. We validate our predictions using multiple paradigms, showing that mature mice bearing cell type-specific deletions of REST or expressing dominant-negative mutant REST show improved regeneration of the corticospinal tract and optic nerve after spinal cord injury and optic nerve crush, which is accompanied by upregulation of regeneration-associated genes in cortical motor neurons and retinal ganglion cells, respectively. These analyses identify a role for REST as an upstream suppressor of the intrinsic regenerative program in the CNS and demonstrate the utility of a systems biology approach involving integrative genomics and bio-informatics to prioritize hypotheses relevant to CNS repair.

A constant blood supply to the brain is required for mental function. Research with Doppler ultrasonography has important clinical value and burgeoning potential with machine learning applications in studies predicting gestational age and vascular aging. Critically, studies on ultrasound metrics in school-age children are sparse and no machine learning study to date has used color duplex ultrasonography to predict age and classify age-group. The purpose of our study is two-fold: first to document cerebrovascular hemodynamics considering age, gender, and hemisphere in three arteries; and second to construct machine learning models that can predict and classify the age and age-group of a participant using ultrasonography metrics. We record peak systolic, end-diastolic, and time-averaged maximum velocities bilaterally in internal carotid, vertebral, and middle cerebral arteries from 821 participants. Results confirm that ultrasonography values decrease with age and reveal that gender and hemispheres show more similarities than differences, which depend on age, artery, and metric. Machine learning algorithms predict age and classifier models distinguish cerebrovascular hemodynamics between children and adults. Blood velocities, rather than blood vessel diameters, are more important for classifier models, and common and distinct variables contribute to age classification models for males and females.

Most approaches to the estimation of brain functional connectivity from the functional magnetic resonance imaging (fMRI) data rely on computing some measure of statistical dependence, or more generally, a distance between univariate representative time series of regions of interest (ROIs) consisting of multiple voxels. However, summarizing a ROI's multiple time series with its mean or the first principal component (1PC) may result to the loss of information as, for example, 1PC explains only a small fraction of variance of the multivariate signal of the neuronal activity. We propose to compare ROIs directly, without the use of representative time series, defining a new measure of multivariate connectivity between ROIs, not necessarily consisting of the same number of voxels, based on the Wasserstein distance. We assess the proposed Wasserstein functional connectivity measure on the autism screening task, demonstrating its superiority over commonly used univariate and multivariate functional connectivity measures.

Schizophrenia is a socially significant mental disorder resulting frequently in severe forms of disability. Diagnosis, choice of treatment tactics, and rehabilitation in clinical psychiatry are mainly based on the assessment of behavioral patterns, socio-demographic data, and other investigations such as clinical observations and neuropsychological testing including examination of patients by the psychiatrist, self-reports, and questionnaires. In many respects, these data are subjective and therefore a large number of works have appeared in recent years devoted to the search for objective characteristics (indices, biomarkers) of the processes going on in the human body and reflected in the behavioral and psychoneurological patterns of patients. Such biomarkers are based on the results of instrumental and laboratory studies (neuroimaging, electro-physiological, biochemical, immunological, genetic, and others) and are successfully being used in neurosciences for understanding the mechanisms of the emergence and development of nervous system pathologies.Presently, with the advent of new effective neuroimaging, laboratory, and other methods of investigation and also with the development of modern methods of data analysis, machine learning, and artificial intelligence, a great number of scientific and clinical studies is being conducted devoted to the search for the markers which have diagnostic and prognostic value and may be used in clinical practice to objectivize the processes of establishing and clarifying the diagnosis, choosing and optimizing treatment and rehabilitation tactics, predicting the course and outcome of the disease.This review presents the analysis of the works which describe the correlates between the diagnosis of schizophrenia, established by health professionals, various manifestations of the psychiatric disorder (its subtype, variant of the course, severity degree, observed symptoms, etc.), and objectively measured characteristics/quantitative indicators (anatomical, functional, immunological, genetic, and others) obtained during instrumental and laboratory examinations of patients.A considerable part of these works has been devoted to correlates/biomarkers of schizophrenia based on the data of structural and functional (at rest and under cognitive load) MRI, EEG, tractography, and immunological data. The found correlates/biomarkers reflect anatomic disorders in the specific brain regions, impairment of functional activity of brain regions and their interconnections, specific microstructure of the brain white matter and the levels of connectivity between the tracts of various structures, alterations of electrical activity in various parts of the brain in different EEG spectral ranges, as well as changes in the innate and adaptive links of immunity.Current methods of data analysis and machine learning to search for schizophrenia biomarkers using the data of diverse modalities and their application during building and interpretation of predictive diagnostic models of schizophrenia have been considered in the present review.

ABIDE is the largest open-source autism spectrum disorder database with both fMRI data and full phenotype description. These data were extensively studied based on functional connectivity analysis as well as with deep learning on raw data, with top models accuracy close to 75% for separate scanning sites. Yet there is still a problem of models transferability between different scanning sites within ABIDE. In the current paper, we for the first time perform domain adaptation for brain pathology classification problem on raw neuroimaging data. We use 3D convolutional autoencoders to build the domain irrelevant latent space image representation and demonstrate this method to outperform existing approaches on ABIDE data.

Deep learning shows high potential for many medical image analysis tasks. Neural networks can work with full-size data without extensive preprocessing and feature generation and, thus, information loss. Recent work has shown that the morphological difference in specific brain regions can be found on MRI with the means of Convolution Neural Networks (CNN). However, interpretation of the existing models is based on a region of interest and can not be extended to voxel-wise image interpretation on a whole image. In the current work, we consider the classification task on a large-scale open-source dataset of young healthy subjects—an exploration of brain differences between men and women. In this paper, we extend the previous findings in gender differences from diffusion-tensor imaging on T1 brain MRI scans. We provide the voxel-wise 3D CNN interpretation comparing the results of three interpretation methods: Meaningful Perturbations, Grad CAM and Guided Backpropagation, and contribute with the open-source library.

Focal cortical dysplasia (FCD) is one of the most common epileptogenic lesions associated with cortical development malformations. However, the accurate detection of the FCD relies on the radiologist professionalism, and in many cases, the lesion could be missed. In this work, we solve the problem of automatic identification of FCD on magnetic resonance images (MRI). For this task, we improve recent methods of Deep Learning-based FCD detection and apply it for a dataset of 15 labeled FCD patients. The model results in the successful detection of FCD on 11 out of 15 subjects.

The impressive capabilities of recent generative models to create texts that are challenging to distinguish from the human-written ones can be misused for generating fake news, product reviews, and even abusive content. Despite the prominent performance of existing methods for artificial text detection, they still lack interpretability and robustness towards unseen models. To this end, we propose three novel types of interpretable topological features for this task based on Topological Data Analysis (TDA) which is currently understudied in the field of NLP. We empirically show that the features derived from the BERT model outperform count- and neural-based baselines up to 10% on three common datasets, and tend to be the most robust towards unseen GPT-style generation models as opposed to existing methods. The probing analysis of the features reveals their sensitivity to the surface and syntactic properties. The results demonstrate that TDA is a promising line with respect to NLP tasks, specifically the ones that incorporate surface and structural information.

Astrocytic contributions to neuroinflammation are widely implicated in disease, but they remain incompletely explored. We assess medial prefrontal cortex (PFC) and visual cortex (VCX) astrocyte and whole-tissue gene expression changes in mice following peripherally induced neuroinflammation triggered by a systemic bacterial endotoxin, lipopolysaccharide, which produces sickness-related behaviors, including anhedonia. Neuroinflammation-mediated behavioral changes and astrocyte-specific gene expression alterations peak when anhedonia is greatest and then reverse to normal. Notably, region-specific molecular identities of PFC and VCX astrocytes are largely maintained during reactivity changes. Gene pathway analyses reveal alterations of diverse cell signaling pathways, including changes in cell-cell interactions of multiple cell types that may underlie the central effects of neuroinflammation. Certain astrocyte molecular signatures accompanying neuroinflammation are shared with changes reported in Alzheimer's disease and mouse models. However, we find no evidence of altered neuronal survival or function in the PFC even when neuroinflammation-induced astrocyte reactivity and behavioral changes are significant.

Machine learning and computer vision methods are showing good performance in medical imagery analysis. Yet only a few applications are now in clinical use and one of the reasons for that is poor transferability of the models to data from different sources or acquisition domains. Development of new methods and algorithms for the transfer of training and adaptation of the domain in multi-modal medical imaging data is crucial for the development of accurate models and their use in clinics. In present work, we overview methods used to tackle the domain shift problem in machine learning and computer vision. The algorithms discussed in this survey include advanced data processing, model architecture enhancing and featured training, as well as predicting in domain invariant latent space. The application of the autoencoding neural networks and their domain-invariant variations are heavily discussed in a survey. We observe the latest methods applied to the magnetic resonance imaging (MRI) data analysis and conclude on their performance as well as propose directions for further research.

Deep learning shows high potential for many medical image analysis tasks. Neural networks work with full-size data without extensive preprocessing and feature generation and, thus, information loss. Recent work has shown that morphological difference between specific brain regions can be found on MRI with deep learning techniques. We consider the pattern recognition task based on a large open-access dataset of healthy subjects - an exploration of brain differences between men and women. However, interpretation of the lately proposed models is based on a region of interest and can not be extended to pixel or voxel-wise image interpretation, which is considered to be more informative. In this paper, we confirm the previous findings in sex differences from diffusion-tensor imaging on T1 weighted brain MRI scans. We compare the results of three voxel-based 3D CNN interpretation methods: Meaningful Perturbations, GradCam and Guided Backpropagation and provide the open-source code.