Controlling astrocyte-mediated synaptic pruning signals for schizophrenia drug repurposing with deep graph networks.

Schizophrenia is a debilitating psychiatric disorder, leading to both physical and social morbidity. Worldwide 1% of the population is struggling with the disease, with 100,000 new cases annually only in the United States. Despite its importance, the goal of finding effective treatments for schizophrenia remains a challenging task, and previous work conducted expensive large-scale phenotypic screens. This work investigates the benefits of Machine Learning for graphs to optimize drug phenotypic screens and predict compounds that mitigate abnormal brain reduction induced by excessive glial phagocytic activity in schizophrenia subjects. Given a compound and its concentration as input, we propose a method that predicts a score associated with three possible compound effects, i.e., reduce, increase, or not influence phagocytosis. We leverage a high-throughput screening to prove experimentally that our method achieves good generalization capabilities. The screening involves 2218 compounds at five different concentrations. Then, we analyze the usability of our approach in a practical setting, i.e., prioritizing the selection of compounds in the SWEETLEAD library. We provide a list of 64 compounds from the library that have the most potential clinical utility for glial phagocytosis mitigation. Lastly, we propose a novel approach to computationally validate their utility as possible therapies for schizophrenia.

A comparison of deterministic and stochastic approaches for sensitivity analysis in computational systems biology.

With the recent rising application of mathematical models in the field of computational systems biology, the interest in sensitivity analysis methods had increased. The stochastic approach, based on chemical master equations, and the deterministic approach, based on ordinary differential equations (ODEs), are the two main approaches for analyzing mathematical models of biochemical systems. In this work, the performance of these approaches to compute sensitivity coefficients is explored in situations where stochastic and deterministic simulation can potentially provide different results (systems with unstable steady states, oscillators with population extinction and bistable systems). We consider two methods in the deterministic approach, namely the direct differential method and the finite difference method, and five methods in the stochastic approach, namely the Girsanov transformation, the independent random number method, the common random number method, the coupled finite difference method and the rejection-based finite difference method. The reviewed methods are compared in terms of sensitivity values and computational time to identify differences in outcome that can highlight conditions in which one approach performs better than the other.

A robust computational pipeline for model-based and data-driven phenotype clustering.

MOTIVATION: Precision medicine is a promising field that proposes, in contrast to a one-size-fits-all approach, the tailoring of medical decisions, treatments or products. In this context, it is crucial to introduce innovative methods to stratify a population of patients on the basis of an accurate system-level knowledge of the disease. This is particularly important in very challenging conditions, where the use of standard statistical methods can be prevented by poor data availability or by the need of oversimplifying the processes regulating a complex disease. RESULTS: We define an innovative method for phenotype classification that combines experimental data and a mathematical description of the disease biology. The methodology exploits the mathematical model for inferring additional subject features relevant for the classification. Finally, the algorithm identifies the optimal number of clusters and classifies the samples on the basis of a subset of the features estimated during the model fit. We tested the algorithm in two test cases: an in silico case in the context of dyslipidemia, a complex disease for which a large population of patients has been generated, and a clinical test case, in the context of a lysosomal rare disorder, for which the amount of available data was limited. In both the scenarios, our methodology proved to be accurate and robust, and allowed the inference of an additional phenotype division that the experimental data did not show. AVAILABILITY AND IMPLEMENTATION: The code to reproduce the in silico results has been implemented in MATLAB v.2017b and it is available in the Supplementary Material. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

Multi-omics integration-a comparison of unsupervised clustering methodologies.

With the recent developments in the field of multi-omics integration, the interest in factors such as data preprocessing, choice of the integration method and the number of different omics considered had increased. In this work, the impact of these factors is explored when solving the problem of sample classification, by comparing the performances of five unsupervised algorithms: Multiple Canonical Correlation Analysis, Multiple Co-Inertia Analysis, Multiple Factor Analysis, Joint and Individual Variation Explained and Similarity Network Fusion. These methods were applied to three real data sets taken from literature and several ad hoc simulated scenarios to discuss classification performance in different conditions of noise and signal strength across the data types. The impact of experimental design, feature selection and parameter training has been also evaluated to unravel important conditions that can affect the accuracy of the result.

Incorporating extrinsic noise into the stochastic simulation of biochemical reactions: A comparison of approaches.

The stochastic simulation algorithm (SSA) has been widely used for simulating biochemical reaction networks. SSA is able to capture the inherently intrinsic noise of the biological system, which is due to the discreteness of species population and to the randomness of their reciprocal interactions. However, SSA does not consider other sources of heterogeneity in biochemical reaction systems, which are referred to as extrinsic noise. Here, we extend two simulation approaches, namely, the integration-based method and the rejection-based method, to take extrinsic noise into account by allowing the reaction propensities to vary in time and state dependent manner. For both methods, new efficient implementations are introduced and their efficiency and applicability to biological models are investigated. Our numerical results suggest that the rejection-based method performs better than the integration-based method when the extrinsic noise is considered.

Efficient stochastic simulation of biochemical reactions with noise and delays.

The stochastic simulation algorithm has been used to generate exact trajectories of biochemical reaction networks. For each simulation step, the simulation selects a reaction and its firing time according to a probability that is proportional to the reaction propensity. We investigate in this paper new efficient formulations of the stochastic simulation algorithm to improve its computational efficiency. We examine the selection of the next reaction firing and reduce its computational cost by reusing the computation in the previous step. For biochemical reactions with delays, we present a new method for computing the firing time of the next reaction. The principle for computing the firing time of our approach is based on recycling of random numbers. Our new approach for generating the firing time of the next reaction is not only computationally efficient but also easy to implement. We further analyze and reduce the number of propensity updates when a delayed reaction occurred. We demonstrate the applicability of our improvements by experimenting with concrete biological models.

SCUDO: a tool for signature-based clustering of expression profiles.

SCUDO (Signature-based ClUstering for DiagnOstic purposes) is an online tool for the analysis of gene expression profiles for diagnostic and classification purposes. The tool is based on a new method for the clustering of profiles based on a subject-specific, as opposed to disease-specific, signature. Our approach relies on construction of a reference map of transcriptional signatures, from both healthy and affected subjects, derived from their respective mRNA or miRNA profiles. A diagnosis for a new individual can then be performed by determining the position of the individual's transcriptional signature on the map. The diagnostic power of our method has been convincingly demonstrated in an open scientific competition (SBV Improver Diagnostic Signature Challenge), scoring second place overall and first place in one of the sub-challenges.

Qualitative and Quantitative Protein Complex Prediction Through Proteome-Wide Simulations.

Despite recent progress in proteomics most protein complexes are still unknown. Identification of these complexes will help us understand cellular regulatory mechanisms and support development of new drugs. Therefore it is really important to establish detailed information about the composition and the abundance of protein complexes but existing algorithms can only give qualitative predictions. Herein, we propose a new approach based on stochastic simulations of protein complex formation that integrates multi-source data--such as protein abundances, domain-domain interactions and functional annotations--to predict alternative forms of protein complexes together with their abundances. This method, called SiComPre (Simulation based Complex Prediction), achieves better qualitative prediction of yeast and human protein complexes than existing methods and is the first to predict protein complex abundances. Furthermore, we show that SiComPre can be used to predict complexome changes upon drug treatment with the example of bortezomib. SiComPre is the first method to produce quantitative predictions on the abundance of molecular complexes while performing the best qualitative predictions. With new data on tissue specific protein complexes becoming available SiComPre will be able to predict qualitative and quantitative differences in the complexome in various tissue types and under various conditions.

Accelerating rejection-based simulation of biochemical reactions with bounded acceptance probability.

Stochastic simulation of large biochemical reaction networks is often computationally expensive due to the disparate reaction rates and high variability of population of chemical species. An approach to accelerate the simulation is to allow multiple reaction firings before performing update by assuming that reaction propensities are changing of a negligible amount during a time interval. Species with small population in the firings of fast reactions significantly affect both performance and accuracy of this simulation approach. It is even worse when these small population species are involved in a large number of reactions. We present in this paper a new approximate algorithm to cope with this problem. It is based on bounding the acceptance probability of a reaction selected by the exact rejection-based simulation algorithm, which employs propensity bounds of reactions and the rejection-based mechanism to select next reaction firings. The reaction is ensured to be selected to fire with an acceptance rate greater than a predefined probability in which the selection becomes exact if the probability is set to one. Our new algorithm improves the computational cost for selecting the next reaction firing and reduces the updating the propensities of reactions.

Simulation of biochemical reactions with time-dependent rates by the rejection-based algorithm.

We address the problem of simulating biochemical reaction networks with time-dependent rates and propose a new algorithm based on our rejection-based stochastic simulation algorithm (RSSA) [Thanh et al., J. Chem. Phys. 141(13), 134116 (2014)]. The computation for selecting next reaction firings by our time-dependent RSSA (tRSSA) is computationally efficient. Furthermore, the generated trajectory is exact by exploiting the rejection-based mechanism. We benchmark tRSSA on different biological systems with varying forms of reaction rates to demonstrate its applicability and efficiency. We reveal that for nontrivial cases, the selection of reaction firings in existing algorithms introduces approximations because the integration of reaction rates is very computationally demanding and simplifying assumptions are introduced. The selection of the next reaction firing by our approach is easier while preserving the exactness.

On the rejection-based algorithm for simulation and analysis of large-scale reaction networks.

Stochastic simulation for in silico studies of large biochemical networks requires a great amount of computational time. We recently proposed a new exact simulation algorithm, called the rejection-based stochastic simulation algorithm (RSSA) [Thanh et al., J. Chem. Phys. 141(13), 134116 (2014)], to improve simulation performance by postponing and collapsing as much as possible the propensity updates. In this paper, we analyze the performance of this algorithm in detail, and improve it for simulating large-scale biochemical reaction networks. We also present a new algorithm, called simultaneous RSSA (SRSSA), which generates many independent trajectories simultaneously for the analysis of the biochemical behavior. SRSSA improves simulation performance by utilizing a single data structure across simulations to select reaction firings and forming trajectories. The memory requirement for building and storing the data structure is thus independent of the number of trajectories. The updating of the data structure when needed is performed collectively in a single operation across the simulations. The trajectories generated by SRSSA are exact and independent of each other by exploiting the rejection-based mechanism. We test our new improvement on real biological systems with a wide range of reaction networks to demonstrate its applicability and efficiency.

Efficient rejection-based simulation of biochemical reactions with stochastic noise and delays.

We propose a new exact stochastic rejection-based simulation algorithm for biochemical reactions and extend it to systems with delays. Our algorithm accelerates the simulation by pre-computing reaction propensity bounds to select the next reaction to perform. Exploiting such bounds, we are able to avoid recomputing propensities every time a (delayed) reaction is initiated or finished, as is typically necessary in standard approaches. Propensity updates in our approach are still performed, but only infrequently and limited for a small number of reactions, saving computation time and without sacrificing exactness. We evaluate the performance improvement of our algorithm by experimenting with concrete biological models.

Classification of Neisseria meningitidis genomes with a bag-of-words approach and machine learning.

Whole genome sequencing of bacteria is important to enable strain classification. Using entire genomes as an input to machine learning (ML) models would allow rapid classification of strains while using information from multiple genetic elements. We developed a "bag-of-words" approach to encode, using SentencePiece or k-mer tokenization, entire bacterial genomes and analyze these with ML. Initial model selection identified SentencePiece with 8,000 and 32,000 words as the best approach for genome tokenization. We then classified in Neisseria meningitidis genomes the capsule B group genotype with 99.6% accuracy and the multifactor invasive phenotype with 90.2% accuracy, in an independent test set. Subsequently, in silico knockouts of 2,808 genes confirmed that the ML model predictions aligned with our current understanding of the underlying biology. To our knowledge, this is the first ML method using entire bacterial genomes to classify strains and identify genes considered relevant by the classifier.

Computational approaches to understanding nutrient metabolism and metabolic disorders.

Computational methods are becoming more and more essential to elucidate biological systems. Many different approaches exist with pros and cons. This paper reviews the most useful technologies focusing on nutrient metabolism and metabolic disorders. Space limitation prevents from exploring the examples in details, but pointers to the relevant papers are reported.

Early outcome detection for COVID-19 patients.

With the outbreak of COVID-19 exerting a strong pressure on hospitals and health facilities, clinical decision support systems based on predictive models can help to effectively improve the management of the pandemic. We present a method for predicting mortality for COVID-19 patients. Starting from a large number of clinical variables, we select six of them with largest predictive power, using a feature selection method based on genetic algorithms and starting from a set of COVID-19 patients from the first wave. The algorithm is designed to reduce the impact of missing values in the set of variables measured, and consider only variables that show good accuracy on validation data. The final predictive model provides accuracy larger than 85% on test data, including a new patient cohort from the second COVID-19 wave, and on patients with imputed missing values. The selected clinical variables are confirmed to be relevant by recent literature on COVID-19.

Literature Mining and Mechanistic Graphical Modelling to Improve mRNA Vaccine Platforms.

RNA vaccines represent a milestone in the history of vaccinology. They provide several advantages over more traditional approaches to vaccine development, showing strong immunogenicity and an overall favorable safety profile. While preclinical testing has provided some key insights on how RNA vaccines interact with the innate immune system, their mechanism of action appears to be fragmented amid the literature, making it difficult to formulate new hypotheses to be tested in clinical settings and ultimately improve this technology platform. Here, we propose a systems biology approach, based on the combination of literature mining and mechanistic graphical modeling, to consolidate existing knowledge around mRNA vaccines mode of action and enhance the translatability of preclinical hypotheses into clinical evidence. A Natural Language Processing (NLP) pipeline for automated knowledge extraction retrieved key biological evidences that were joined into an interactive mechanistic graphical model representing the chain of immune events induced by mRNA vaccines administration. The achieved mechanistic graphical model will help the design of future experiments, foster the generation of new hypotheses and set the basis for the development of mathematical models capable of simulating and predicting the immune response to mRNA vaccines.

QSPcc reduces bottlenecks in computational model simulations.

Mathematical models have grown in size and complexity becoming often computationally intractable. In sensitivity analysis and optimization phases, critical for tuning, validation and qualification, these models may be run thousands of times. Scientific programming languages popular for prototyping, such as MATLAB and R, can be a bottleneck in terms of performance. Here we show a compiler-based approach, designed to be universal at handling engineering and life sciences modeling styles, that automatically translates models into fast C code. At first QSPcc is demonstrated to be crucial in enabling the research on otherwise intractable Quantitative Systems Pharmacology models, such as in rare Lysosomal Storage Disorders. To demonstrate the full value in seamlessly accelerating, or enabling, the R&D efforts in natural sciences, we then benchmark QSPcc against 8 solutions on 24 real-world projects from different scientific fields. With speed-ups of 22000x peak, and 1605x arithmetic mean, our results show consistent superior performances.

The Beta Workbench: a computational tool to study the dynamics of biological systems.

We introduce the Beta Workbench (BWB), a scalable tool built on top of the newly defined BlenX language to model, simulate and analyse biological systems. We show the features and the incremental modelling process supported by the BWB on a running example based on the mitogen-activated kinase pathway. Finally, we provide a comparison with related approaches and some hints for future extensions.

Calibration of dynamic models of biological systems with KInfer.

Methods for parameter estimation that are robust to experimental uncertainties and to stochastic and biological noise and that require a minimum of a priori input knowledge are of key importance in computational systems biology. The new method presented in this paper aims to ensure an inference model that deduces the rate constants of a system of biochemical reactions from experimentally measured time courses of reactants. This new method was applied to some challenging parameter estimation problems of nonlinear dynamic biological systems and was tested both on synthetic and real data. The synthetic case studies are the 12-state model of the SERCA pump and a model of a genetic network containing feedback loops of interaction between regulator and effector genes. The real case studies consist of a model of the reaction between the inhibitor kappaB kinase enzyme and its substrate in the signal transduction pathway of NF-kappaB, and a stiff model of the fermentation pathway of Lactococcus lactis.

DNA methylation during human adipogenesis and the impact of fructose.

BACKGROUND: Increased adipogenesis and altered adipocyte function contribute to the development of obesity and associated comorbidities. Fructose modified adipocyte metabolism compared to glucose, but the regulatory mechanisms and consequences for obesity are unknown. Genome-wide methylation and global transcriptomics in SGBS pre-adipocytes exposed to 0, 2.5, 5, and 10 mM fructose, added to a 5-mM glucose-containing medium, were analyzed at 0, 24, 48, 96, 192, and 384 h following the induction of adipogenesis. RESULTS: Time-dependent changes in DNA methylation compared to baseline (0 h) occurred during the final maturation of adipocytes, between 192 and 384 h. Larger percentages (0.1% at 192 h, 3.2% at 384 h) of differentially methylated regions (DMRs) were found in adipocytes differentiated in the glucose-containing control media compared to adipocytes differentiated in fructose-supplemented media (0.0006% for 10 mM, 0.001% for 5 mM, and 0.005% for 2.5 mM at 384 h). A total of 1437 DMRs were identified in 5237 differentially expressed genes at 384 h post-induction in glucose-containing (5 mM) control media. The majority of them inversely correlated with the gene expression, but 666 regions were positively correlated to the gene expression. CONCLUSIONS: Our studies demonstrate that DNA methylation regulates or marks the transformation of morphologically differentiating adipocytes (seen at 192 h), to the more mature and metabolically robust adipocytes (as seen at 384 h) in a genome-wide manner. Lower (2.5 mM) concentrations of fructose have the most robust effects on methylation compared to higher concentrations (5 and 10 mM), suggesting that fructose may be playing a signaling/regulatory role at lower concentrations of fructose and as a substrate at higher concentrations.