Maria Poptsova’s research while affiliated with National Research University Higher School of Economics and other places

The double-stranded RNA editing enzyme ADAR1 connects two forms of genetic programming, one based on codons and the other on flipons. ADAR1 recodes codons in pre-mRNA by deaminating adenosine to form inosine, which is translated as guanosine. ADAR1 also plays essential roles in the immune defense against viruses and cancers by recognizing left-handed Z-DNA and Z-RNA (collectively called ZNA). Here, we review various aspects of ADAR1 biology, starting with codons and progressing to flipons. ADAR1 has two major isoforms, with the p110 protein lacking the p150 Zα domain that binds ZNAs with high affinity. The p150 isoform is induced by interferon and targets ALU inverted repeats, a class of endogenous retroelement that promotes their transcription and retrotransposition by incorporating Z-flipons that encode ZNAs and G-flipons that form G-quadruplexes (GQ). Both p150 and p110 include the Zβ domain that is related to Zα but does not bind ZNAs. Here we report strong evidence that Zβ binds the GQ that are formed co-transcriptionally by ALU repeats and within R-loops. By binding GQ, ADAR1 suppresses ALU-mediated alternative splicing, generates most of the reported nonsynonymous edits and promotes R-loop resolution. The recognition of the various alternative nucleic acid conformations by ADAR1 connects genetic programming by flipons with the encoding of information by codons. The findings suggest that incorporating G-flipons into editmers might improve the therapeutic editing efficacy of ADAR1.

Background Polygenic risk score (PRS) prediction is widely used to assess the risk of diagnosis and progression of many diseases. Routinely, the weights of individual SNPs are estimated by the linear regression model that assumes independent and linear contribution of each SNP to the phenotype. However, for complex multifactorial diseases such as Alzheimer’s disease, diabetes, cardiovascular disease, cancer, and others, association between individual SNPs and disease could be non-linear due to epistatic interactions. The aim of the presented study is to explore the power of non-linear machine learning algorithms and deep learning models to predict the risk of multifactorial diseases with epistasis. Methods Simulated data with 2- and 3-loci interactions and tested three different models of epistasis: additive, multiplicative and threshold, were generated using the GAMETES. Penetrance tables were generated using PyTOXO package. For machine learning methods we used multilayer perceptron (MLP), convolutional neural network (CNN) and recurrent neural network (RNN), Lasso regression, random forest and gradient boosting models. Performance of machine learning models were assessed using accuracy, AUC-ROC, AUC-PR, recall, precision, and F1 score. Results First, we tested ensemble tree methods and deep learning neural networks against LASSO linear regression model on simulated data with different types and strength of epistasis. The results showed that with the increase of strength of epistasis effect, non-linear models significantly outperform linear. Then the higher performance of non-linear models over linear was confirmed on real genetic data for multifactorial phenotypes such as obesity, type 1 diabetes, and psoriasis. From non-linear models, gradient boosting appeared to be the best model in obesity and psoriasis while deep learning methods significantly outperform linear approaches in type 1 diabetes. Conclusion Overall, our study underscores the efficacy of non-linear models and deep learning approaches in more accurately accounting for the effects of epistasis in simulations with specific configurations and in the context of certain diseases.

Kolmogorov-Arnold Networks (KANs) emerged as a promising alternative for multilayer perceptrons in dense fully connected networks. Multiple attempts have been made to integrate KANs into various deep learning architectures in the domains of computer vision and natural language processing. Integrating KANs into deep learning models for genomic tasks has not been explored. Here, we tested linear KANs (LKANs) and convolutional KANs (CKANs) as replacement for MLP in baseline deep learning architectures for classification and generation of genomic sequences. We used three genomic benchmark datasets: Genomic Benchmarks, Genome Understanding Evaluation, and Flipon Benchmark. We demonstrated that LKANs outperformed both baseline and CKANs on almost all datasets. CKANs can achieve comparable results but struggle with scaling over large number of parameters. Ablation analysis demonstrated that the number of KAN layers correlates with the model performance. Overall, linear KANs show promising results in improving the performance of deep learning models with relatively small number of parameters. Unleashing KAN potential in different SOTA deep learning architectures currently used in genomics requires further research.

Background The development of prognostic models for the identification of high-risk myocardial infarction (MI) patients is a crucial step toward personalized medicine. Genetic factors are known to be associated with an increased risk of cardiovascular diseases; however, little is known about whether they can be used to predict major adverse cardiac events (MACEs) for MI patients. This study aimed to build a machine learning (ML) model to predict MACEs in MI patients based on clinical, imaging, laboratory, and genetic features and to assess the influence of genetics on the prognostic power of the model. Methods We analyzed the data from 218 MI patients admitted to the emergency department at the Surgut District Center for Diagnostics and Cardiovascular Surgery, Russia. Upon admission, standard clinical measurements and imaging data were collected for each patient. Additionally, patients were genotyped for VEGFR-2 variation rs2305948 (C/C, C/T, T/T genotypes with T being the minor risk allele). The study included a 9-year follow-up period during which major ischemic events were recorded. We trained and evaluated various ML models, including Gradient Boosting, Random Forest, Logistic Regression, and AutoML. For feature importance analysis, we applied the sequential feature selection (SFS) and Shapley’s scheme of additive explanation (SHAP) methods. Results The CatBoost algorithm, with features selected using the SFS method, showed the best performance on the test cohort, achieving a ROC AUC of 0.813. Feature importance analysis identified the dose of statins as the most important factor, with the VEGFR-2 genotype among the top 5. The other important features are coronary artery lesions (coronary artery stenoses ≥70%), left ventricular (LV) parameters such as lateral LV wall and LV mass, diabetes, type of revascularization (CABG or PCI), and age. We also showed that contributions are additive and that high risk can be determined by cumulative negative effects from different prognostic factors. Conclusion Our ML-based approach demonstrated that the VEGFR-2 genotype is associated with an increased risk of MACEs in MI patients. However, the risk can be significantly reduced by high-dose statins and positive factors such as the absence of coronary artery lesions, absence of diabetes, and younger age.

A long-standing question concerns the role of Z-DNA in transcription. Here we use a deep learning approach DeepZ that predicts Z-flipons based on DNA sequence, structural properties of nucleotides and omics data. We examined Z-flipons that are conserved between human and mouse genomes after generating whole-genome Z-flipon maps and then validated them by orthogonal approaches based on high resolution chemical mapping of Z-DNA and the transformer algorithm Z-DNABERT. For human and mouse, we revealed similar pattern of transcription factors, chromatin remodelers, and histone marks associated with conserved Z-flipons. We found significant enrichment of Z-flipons in alternative and bidirectional promoters associated with neurogenesis genes. We show that conserved Z-flipons are associated with increased experimentally determined transcription reinitiation rates compared to promoters without Z-flipons, but without affecting elongation or pausing. Our findings support a model where Z-flipons engage Transcription Factor E and impact phenotype by enabling the reset of preinitiation complexes when active, and the suppression of gene expression when engaged by repressive chromatin complexes.

Alternative DNA conformation formed by sequences called flipons potentially alter the readout of genetic information by directing the shape-specific assembly of complexes on DNA The biological roles of G-quadruplexes formed by motifs rich in guanosine repeats have been investigated experimentally using many different methodologies including G4-seq, G4 ChIP-seq, permanganate nuclease footprinting (KEx), KAS-seq, CUT&Tag with varying degrees of overlap between the results. Here we trained large language model DNABERT on existing data generated by KEx, a rapid chemical footprinting technique performed on live, intact cells using potassium permanganate. The snapshot of flipon state when combined with results from other in vitro methods that are performed on permeabilized cells, allows a high confidence mapping of G-flipons to proximal enhancer and promoter sequences. Using G4-DNABERT predictions,with data from ENdb, Zoonomia cCREs and single сell G4 CUT&Tag experiments, we found support for a model where G4-quadruplexes regulate gene expression through chromatin loop formation.

Motivation: Deep learning methods have been successfully applied to the tasks of predicting non-B DNA structures, however model performance depends on the availability of experimental data for training. Experimental technologies for non-B DNA structure detection are limited to the subsets that are active at the time of an experiment and cannot detect entire functional set of elements. Recently deep generative models demonstrated promising results in data augmentation approach improving classifier performance trained on augmented real and generated data. Here we aimed at testing performance of diffusion models in comparison to other generative models and explore the data augmentation approach for the task of non-B DNA structure prediction. Results: We tested denoising diffusion probabilistic and implicit models (DDPM and DDIM), Wasserstein generative adversarial network (WGAN) and vector quantised variational autoencoder (VQ-VAE) for the task of improving detection of Z-DNA, G-quadruplexes and H-DNA. We showed that data augmentation increased the quality of classifiers with diffusion models being the best for Z-DNA and H-DNA while WGAN worked better for G4s. Diffusion models are the best in diversity for all types of non-B DNA structures, WGAN produced the best novelty for G-quadruplexes and H- DNA. Since diffusion models require substantial resources, we showed that distillation technique can significantly enhance sampling in training diffusion models. When considering three criteria - quality of generated samples, sampling speed, and diversity, we conclude that trade-off is possible between generative diffusion model and other architectures such as WGAN and VQ-VAE. Availability: The code with conducted experiments is freely available at https://github.com/powidla/nonB-DNA- structures-generation.

Background Polygenic risk score (PRS) prediction is widely used to assess the risk of diagnosis and progression of many diseases. Routinely, the weights of individual SNPs are estimated by the linear regression model that assumes independent and linear contribution of each SNP to the phenotype. However, for complex multifactorial diseases such as Alzheimer's disease, diabetes, cardiovascular disease, cancer, and others, association between individual SNPs and disease could be non-linear due to epistatic interactions. The aim of the presented study is to explore the power of non-linear machine learning algorithms and deep learning models to predict the risk of multifactorial diseases with epistasis. Results First, we tested ensemble tree methods and deep learning neural networks against LASSO linear regression model on simulated data with different types and strength of epistasis. The results showed that with the increase of strength of epistasis effect, non-linear models significantly outperform linear. Then the higher performance of non-linear models over linear was confirmed on real genetic data for multifactorial phenotypes such as obesity, type 1 diabetes, and psoriasis. From non-linear models, gradient boosting appeared to be the best model in obesity and psoriasis while deep learning methods significantly outperform linear approaches in type 1 diabetes. Conclusions Overall, our study underscores the efficacy of non-linear models and deep learning approaches in more accurately accounting for the effects of epistasis in simulations with specific configurations and in the context of certain diseases.