A practical application of generative adversarial networks for RNA-seq analysis to predict the molecular progress of Alzheimer's disease.

Next-generation sequencing (NGS) technology has become a powerful tool for dissecting the molecular and pathological signatures of a variety of human diseases. However, the limited availability of biological samples from different disease stages is a major hurdle in studying disease progressions and identifying early pathological changes. Deep learning techniques have recently begun to be applied to analyze NGS data and thereby predict the progression of biological processes. In this study, we applied a deep learning technique called generative adversarial networks (GANs) to predict the molecular progress of Alzheimer's disease (AD). We successfully applied GANs to analyze RNA-seq data from a 5xFAD mouse model of AD, which recapitulates major AD features of massive amyloid-ß (Aß) accumulation in the brain. We examined how the generator is featured to have specific-sample generation and biological gene association. Based on the above observations, we suggested virtual disease progress by latent space interpolation to yield the transition curves of various genes with pathological changes from normal to AD state. By performing pathway analysis based on the transition curve patterns, we identified several pathological processes with progressive changes, such as inflammatory systems and synapse functions, which have previously been demonstrated to be involved in the pathogenesis of AD. Interestingly, our analysis indicates that alteration of cholesterol biosynthesis begins at a very early stage of AD, suggesting that it is the first effect to mediate the cholesterol metabolism of AD downstream of Aß accumulation. Here, we suggest that GANs are a useful tool to study disease progression, leading to the identification of early pathological signatures.

Uncovering multiloci-ordering by algebraic property of Laplacian matrix and its Fiedler vector.

MOTIVATION: The loci-ordering, based on two-point recombination fractions for a pair of loci, is the most important step in constructing a reliable and fine genetic map. RESULTS: Using the concept from complex graph theory, here we propose a Laplacian ordering approach which uncovers the loci-ordering of multiloci simultaneously. The algebraic property for a Fiedler vector of a Laplacian matrix, constructed from the recombination fraction of the loci-ordering for 26 loci of barley chromosome IV, 846 loci of Arabidopsis thaliana and 1903 loci of Malus domestica, together with the variable threshold uncovers their loci-orders. It offers an alternative yet robust approach for ordering multiloci. AVAILABILITY AND IMPLEMENTATION: Source code program with data set is available as supplementary data and also in a software category of the website (http://biophysics.dgist.ac.kr) CONTACT: crkim@pusan.ac.kr or iksoochang@dgist.ac.kr SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

Structural Conversion of Aß17-42 Peptides from Disordered Oligomers to U-Shape Protofilaments via Multiple Kinetic Pathways.

Discovering the mechanisms by which proteins aggregate into fibrils is an essential first step in understanding the molecular level processes underlying neurodegenerative diseases such as Alzheimer's and Parkinson's. The goal of this work is to provide insights into the structural changes that characterize the kinetic pathways by which amyloid-ß peptides convert from monomers to oligomers to fibrils. By applying discontinuous molecular dynamics simulations to PRIME20, a force field designed to capture the chemical and physical aspects of protein aggregation, we have been able to trace out the entire aggregation process for a system containing 8 Aß17-42 peptides. We uncovered two fibrillization mechanisms that govern the structural conversion of Aß17-42 peptides from disordered oligomers into protofilaments. The first mechanism is monomeric conversion templated by a U-shape oligomeric nucleus into U-shape protofilament. The second mechanism involves a long-lived and on-pathway metastable oligomer with S-shape chains, having a C-terminal turn, en route to the final U-shape protofilament. Oligomers with this C-terminal turn have been regarded in recent experiments as a major contributing element to cell toxicity in Alzheimer's disease. The internal structures of the U-shape protofilaments from our PRIME20/DMD simulation agree well with those from solid state NMR experiments. The approach presented here offers a simple molecular-level framework to describe protein aggregation in general and to visualize the kinetic evolution of a putative toxic element in Alzheimer's disease in particular.

Impact of sequence on the molecular assembly of short amyloid peptides.

The goal of this work is to understand how the sequence of a protein affects the likelihood that it will form an amyloid fibril and the kinetics along the fibrillization pathway. The focus is on very short fragments of amyloid proteins since these play a role in the fibrillization of the parent protein and can form fibrils themselves. Discontinuous molecular dynamics simulations using the PRIME20 force field were performed of the aggregation of 48-peptide systems containing SNQNNF (PrP (170-175)), SSTSAA (RNaseA(15-20)), MVGGVV (Aß(35-40)), GGVVIA (Aß(37-42)), and MVGGVVIA (Aß(35-42)). In our simulations SNQQNF, SSTTSAA, and MVGGVV form large numbers of fibrillar structures spontaneously (as in experiment). GGVVIA forms ß-sheets that do not stack into fibrils (unlike experiment). The combination sequence MVGGVVIA forms less fibrils than MVGGVV, hindered by the presence of the hydrophobic residues at the C-terminal. Analysis of the simulation kinetics and energetics reveals why MVGGVV forms fibrils and GGVVIA does not, and why adding I and A to MVGGVVIA reduces fibrillization and enhances amorphous aggregation into oligomeric structures. The latter helps explain why Aß(1-42) assembles into more complex oligomers than Aß(1-40), a consequence of which is that it is more strongly associated with Alzheimer's disease.

Complement receptor 4 mediates the clearance of extracellular tau fibrils by microglia.

Tauopathies exhibit a characteristic accumulation of misfolded tau aggregates in the brain. Tau pathology shows disease-specific spatiotemporal propagation through intercellular transmission, which is closely correlated with the progression of clinical manifestations. Therefore, identifying molecular mechanisms that prevent tau propagation is critical for developing therapeutic strategies for tauopathies. The various innate immune receptors, such as complement receptor 3 (CR3) and complement receptor 4 (CR4), have been reported to play a critical role in the clearance of various extracellular toxic molecules by microglia. However, their role in tau clearance has not been studied yet. In the present study, we investigated the role of CR3 and CR4 in regulating extracellular tau clearance. We found that CR4 selectively binds to tau fibrils but not to tau monomers, whereas CR3 does not bind to either of them. Inhibiting CR4, but not CR3, significantly reduces the uptake of tau fibrils by BV2 cells and primary microglia. By contrast, inhibiting CR4 has no effect on the uptake of tau monomers by BV2 cells. Furthermore, inhibiting CR4 suppresses the clearance of extracellular tau fibrils, leading to more seed-competent tau fibrils remaining in the extracellular space relative to control samples. We also provide evidence that the expression of CR4 is upregulated in the brains of human Alzheimer's disease patients and the PS19 mouse model of tauopathy. Taken together, our data strongly support that CR4 is a previously undescribed receptor for the clearance of tau fibrils in microglia and may represent a novel therapeutic target for tauopathy.

Investigation of early molecular alterations in tauopathy with generative adversarial networks.

The recent advances in deep learning-based approaches hold great promise for unravelling biological mechanisms, discovering biomarkers, and predicting gene function. Here, we deployed a deep generative model for simulating the molecular progression of tauopathy and dissecting its early features. We applied generative adversarial networks (GANs) for bulk RNA-seq analysis in a mouse model of tauopathy (TPR50-P301S). The union set of differentially expressed genes from four comparisons (two phenotypes with two time points) was used as input training data. We devised four-way transition curves for a virtual simulation of disease progression, clustered and grouped the curves by patterns, and identified eight distinct pattern groups showing different biological features from Gene Ontology enrichment analyses. Genes that were upregulated in early tauopathy were associated with vasculature development, and these changes preceded immune responses. We confirmed significant disease-associated differences in the public human data for the genes of the different pattern groups. Validation with weighted gene co-expression network analysis suggested that our GAN-based approach can be used to detect distinct patterns of early molecular changes during disease progression, which may be extremely difficult in in vivo experiments. The generative model is a valid systematic approach for exploring the sequential cascades of mechanisms and targeting early molecular events related to dementia.

Unveiling OASIS family as a key player in hypoxia-ischemia cases induced by cocaine using generative adversarial networks.

Repeated cocaine use poses many serious health risks to users. One of the risks is hypoxia and ischemia (HI). To restore the biological system against HI, complex biological mechanisms operate at the gene level. Despite the complexity of biological mechanisms, there are common denominator genes that play pivotal roles in various defense systems. Among these genes, the cAMP response element-binding (Creb) protein contributes not only to various aspects of drug-seeking behavior and drug reward, but also to protective mechanisms. However, it is still unclear which Creb members are key players in the protection of cocaine-induced HI conditions. Herein, using one of the state-of-the-art deep learning methods, the generative adversarial network, we revealed that the OASIS family, one of the Creb family, is a key player in various defense mechanisms such as angiogenesis and unfolded protein response against the HI state by unveiling hidden mRNA expression profiles. Furthermore, we identified mysterious kinases in the OASIS family and are able to explain why the prefrontal cortex and hippocampus are vulnerable to HI at the genetic level.

Spontaneous formation of twisted Aß(16-22) fibrils in large-scale molecular-dynamics simulations.

Protein aggregation is associated with fatal neurodegenerative diseases, including Alzheimer's and Parkinson's. Mapping out kinetics along the aggregation pathway could provide valuable insights into the mechanisms that drive oligomerization and fibrillization, but that is beyond the current scope of computational research. Here we trace out the full kinetics of the spontaneous formation of fibrils by 48 Aß(16-22) peptides, following the trajectories in molecular detail from an initial random configuration to a final configuration of twisted protofilaments with cross-ß-structure. We accomplish this by performing large-scale molecular-dynamics simulations based on an implicit-solvent, intermediate-resolution protein model, PRIME20. Structural details such as the intersheet distance, perfectly antiparallel ß-strands, and interdigitating side chains analogous to a steric zipper interface are explained by and in agreement with experiment. Two characteristic fibrillization mechanisms - nucleation/templated growth and oligomeric merging/structural rearrangement - emerge depending on the temperature.

Computer simulation study of amyloid fibril formation by palindromic sequences in prion peptides.

We simulate the aggregation of large systems containing palindromic peptides from the Syrian hamster prion protein SHaPrP 113-120 (AGAAAAGA) and the mouse prion protein MoPrP 111-120 (VAGAAAAGAV) and eight sequence variations: GAAAAAAG, (AG)(4) , A8, GAAAGAAA, A10, V10, GAVAAAAVAG, and VAVAAAAVAV The first two peptides are thought to act as the Velcro that holds the parent prion proteins together in amyloid structures and can form fibrils themselves. Kinetic events along the fibrillization pathway influence the types of structures that occur and variations in the sequence affect aggregation kinetics and fibrillar structure. Discontinuous molecular dynamics simulations using the PRIME20 force field are performed on systems containing 48 peptides starting from a random coil configuration. Depending on the sequence, fibrillar structures form spontaneously over a range of temperatures, below which amorphous aggregates form and above which no aggregation occurs. AGAAAAGA forms well organized fibrillar structures whereas VAGAAAAGAV forms less well organized structures that are partially fibrillar and partially amorphous. The degree of order in the fibrillar structure stems in part from the types of kinetic events leading up to its formation, with AGAAAAGA forming less amorphous structures early in the simulation than VAGAAAAGAV. The ability to form fibrils increases as the chain length and the length of the stretch of hydrophobic residues increase. However as the hydrophobicity of the sequence increases, the ability to form well-ordered structures decreases. Thus, longer hydrophobic sequences form slightly disordered aggregates that are partially fibrillar and partially amorphous. Subtle changes in sequence result in slightly different fibril structures.

A simple and exact Laplacian clustering of complex networking phenomena: application to gene expression profiles.

Unraveling of the unified networking characteristics of complex networking phenomena is of great interest yet a formidable task. There is currently no simple strategy with a rigorous framework. Using an analogy to the exact algebraic property for a transition matrix of a master equation in statistical physics, we propose a method based on a Laplacian matrix for the discovery and prediction of new classes in the unsupervised complex networking phenomena where the class of each sample is completely unknown. Using this proposed Laplacian approach, we can simultaneously discover different classes and determine the identity of each class. Through an illustrative test of the Laplacian approach applied to real datasets of gene expression profiles, leukemia data [Golub TR, et al. (1999) Science 286:531-537], and lymphoma data [Alizadeh AA, et al. (2000) Nature 403:503-511], we demonstrate that this approach is accurate and robust with a mathematical and physical realization. It offers a general framework for characterizing any kind of complex networking phenomenon in broad areas irrespective of whether they are supervised or unsupervised.

Cooperative folding kinetics of BBL protein and peripheral subunit-binding domain homologues.

Recent experiments claiming that Naf-BBL protein follows a global downhill folding raised an important controversy as to the folding mechanism of fast-folding proteins. Under the global downhill folding scenario, not only do proteins undergo a gradual folding, but folding events along the continuous folding pathway also could be mapped out from the equilibrium denaturation experiment. Based on the exact calculation using a free energy landscape, relaxation eigenmodes from a master equation, and Monte Carlo simulation of an extended Muñoz-Eaton model that incorporates multiscale-heterogeneous pairwise interactions between amino acids, here we show that the very nature of a two-state cooperative transition such as a bimodal distribution from an exact free energy landscape and biphasic relaxation kinetics manifest in the thermodynamics and folding-unfolding kinetics of BBL and peripheral subunit-binding domain homologues. Our results provide an unequivocal resolution to the fundamental controversy related to the global downhill folding scheme, whose applicability to other proteins should be critically reexamined.

Extending the PRIME model for protein aggregation to all 20 amino acids.

We extend PRIME, an intermediate-resolution protein model previously used in simulations of the aggregation of polyalanine and polyglutamine, to the description of the geometry and energetics of peptides containing all 20 amino acid residues. The 20 amino acid side chains are classified into 14 groups according to their hydrophobicity, polarity, size, charge, and potential for side chain hydrogen bonding. The parameters for extended PRIME, called PRIME 20, include hydrogen-bonding energies, side chain interaction range and energy, and excluded volume. The parameters are obtained by applying a perceptron-learning algorithm and a modified stochastic learning algorithm that optimizes the energy gap between 711 known native states from the PDB and decoy structures generated by gapless threading. The number of independent pair interaction parameters is chosen to be small enough to be physically meaningful yet large enough to give reasonably accurate results in discriminating decoys from native structures. The most physically meaningful results are obtained with 19 energy parameters.

The Role of HDAC6 in TDP-43-Induced Neurotoxicity and UPS Impairment.

Transactive response DNA-binding protein 43 (TDP-43)-induced neurotoxicity is currently well recognized as a contributor to the pathology of amyotrophic lateral sclerosis (ALS), and the deposition of TDP-43 has been linked to other neurodegenerative diseases, such as frontotemporal lobar degeneration (FTLD) and Alzheimer's disease (AD). Recent studies also suggest that TDP-43-induced neurotoxicity is associated with ubiquitin-proteasome system (UPS) impairment. Histone deacetylase 6 (HDAC6) is a well-known cytosolic deacetylase enzyme that suppresses the toxicity of UPS impairment. However, the role of HDAC6 in TDP-43-induced neurodegeneration is largely unknown. In this study, we found that HDAC6 overexpression decreased the levels of insoluble and cytosolic TDP-43 protein in TDP-43-overexpressing N2a cells. In addition, TDP-43 overexpression upregulated HDAC6 protein and mRNA levels, and knockdown of Hdac6 elevated the total protein level of TDP-43. We further found that HDAC6 modulates TDP-43-induced UPS impairment via the autophagy-lysosome pathway (ALP). We also showed that TDP-43 promoted a short lifespan in flies and that the accumulation of ubiquitin aggregates and climbing defects were significantly rescued by overexpression of HDAC6 in flies. Taken together, these findings suggest that HDAC6 overexpression can mitigate neuronal toxicity caused by TDP-43-induced UPS impairment, which may represent a novel therapeutic approach for ALS.

Calculation of the free energy barriers in the oligomerisation of Abeta peptide fragments.

Protein misfolding and aggregation are associated with a range of severe human neurodegenerative conditions. We use all-atom simulations to describe the process of assembly of the Abeta(16-22) and Abeta(25-35) fragments of Abeta, a peptide associated with Alzheimer's disease. Our results indicate that the pathways of aggregation of these two peptides depend predominantly on the relative strength of hydrophobic interactions and hydrogen bonding. In the Abeta(25-35) peptide, which is weakly hydrophobic, the tendency to form hydrogen bonds drives the crossing of a single major free energy barrier for the formation of a cross-beta structure. By contrast, in the more hydrophobic Abeta(16-22) peptide, the process of ordered assembly is preceded by an initial collapse into disordered oligomers. These results provide support for a recently proposed two-step mechanism of amyloid formation. We have also found that the barriers for reordering are lower for large oligomers than for small oligomers, a result that provides an explanation of the recent experimental observation that the efficiency of the seeding reaction depends on the size of the seeds themselves.

Structural reorganisation and potential toxicity of oligomeric species formed during the assembly of amyloid fibrils.

Increasing evidence indicates that oligomeric protein assemblies may represent the molecular species responsible for cytotoxicity in a range of neurological disorders including Alzheimer and Parkinson diseases. We use all-atom computer simulations to reveal that the process of oligomerization can be divided into two steps. The first is characterised by a hydrophobic coalescence resulting in the formation of molten oligomers in which hydrophobic residues are sequestered away from the solvent. In the second step, the oligomers undergo a process of reorganisation driven by interchain hydrogen bonding interactions that induce the formation of beta sheet rich assemblies in which hydrophobic groups can become exposed. Our results show that the process of aggregation into either ordered or amorphous species is largely determined by a competition between the hydrophobicity of the amino acid sequence and the tendency of polypeptide chains to form arrays of hydrogen bonds. We discuss how the increase in solvent-exposed hydrophobic surface resulting from such a competition offers an explanation for recent observations concerning the cytotoxicity of oligomeric species formed prior to mature amyloid fibrils.

Polymorphism of fibrillar structures depending on the size of assembled Aß17-42 peptides.

The size of assembled Aß17-42 peptides can determine polymorphism during oligomerization and fibrillization, but the mechanism of this effect is unknown. Starting from separate random monomers, various fibrillar oligomers with distinct structural characteristics were identified using discontinuous molecular dynamics simulations based on a coarse-grained protein model. From the structures observed in the simulations, two characteristic oligomer sizes emerged, trimer and paranuclei, which generated distinct structural patterns during fibrillization. A majority of the simulations for trimers and tetramers formed non-fibrillar oligomers, which primarily progress to off-pathway oligomers. Pentamers and hexamers were significantly converted into U-shape fibrillar structures, meaning that these oligomers, called paranuclei, might be potent on-pathway intermediates in fibril formation. Fibrillar oligomers larger than hexamers generated substantial polymorphism in which hybrid structures were readily formed and homogeneous fibrillar structures appeared infrequently.

Perceptron learning of pairwise contact energies for proteins incorporating the amino acid environment.

Although a coarse-grained description of proteins is a simple and convenient way to attack the protein folding problem, the construction of a global pairwise energy function which can simultaneously recognize the native folds of many proteins has resulted in partial success. We have sought the possibility of a systematic improvement of this pairwise-contact energy function as we extended the parameter space of amino acids, incorporating local environments of amino acids, beyond a 20 x 20 matrix. We have studied the pairwise contact energy functions of 20 x 20, 60 x 60, and 180 x 180 matrices depending on the extent of parameter space, and compared their effect on the learnability of energy parameters in the context of a gapless threading, bearing in mind that a 20 x 20 pairwise contact matrix has been shown to be too simple to recognize the native folds of many proteins. In this paper, we show that the construction of a global pairwise energy function was achieved using 1006 training proteins of a homology of less than 30%, which include all representatives of different protein classes. After parametrizing the local environments of the amino acids into nine categories depending on three secondary structures and three kinds of hydrophobicity (desolvation), the 16290 pairwise contact energies (scores) of the amino acids could be determined by perceptron learning and protein threading. These could simultaneously recognize all the native folds of the 1006 training proteins. When these energy parameters were tested on the 382 test proteins of a homology of less than 90%, 370 (96.9%) proteins could recognize their native folds. We set up a simple thermodynamic framework in the conformational space of decoys to calculate the unfolded fraction and the specific heat of real proteins. The different thermodynamic stabilities of E.coli ribonuclease H (RNase H) and its mutants were well described in our calculation, agreeing with the experiment.

Effects of macromolecular crowding on amyloid beta (16-22) aggregation using coarse-grained simulations.

To examine the effect of crowding on protein aggregation, discontinuous molecular dynamics (DMD) simulations combined with an intermediate resolution protein model, PRIME20, were applied to a peptide/crowder system. The systems contained 192 Aß(16-22) peptides and crowders of diameters 5, 20, and 40 Å, represented here by simple hard spheres, at crowder volume fractions of 0.00, 0.10, and 0.20. Results show that both crowder volume fraction and crowder diameter have a large impact on fibril and oligomer formation. The addition of crowders to a system of peptides increases the rate of oligomer formation, shifting from a slow ordered formation of oligomers in the absence of crowders, similar to nucleated polymerization, to a fast collapse of peptides and subsequent rearrangement characteristic of nucleated conformational conversion with a high maximum in the number of peptides in oligomers as the total crowder surface area increases. The rate of conversion from oligomers to fibrils also increases with increasing total crowder surface area, giving rise to an increased rate of fibril growth. In all cases, larger volume fractions and smaller crowders provide the greatest aggregation enhancement effects. We also show that the size of the crowders influences the formation of specific oligomer sizes. In our simulations, the 40 Å crowders enhance the number of dimers relative to the numbers of trimers, hexamers, pentamers, and hexamers, while the 5 Å crowders enhance the number of hexamers relative to the numbers of dimers, trimers, tetramers, and pentamers. These results are in qualitative agreement with previous experimental and theoretical work.

Influence of temperature on formation of perfect tau fragment fibrils using PRIME20/DMD simulations.

We investigate the fibrillization process for amyloid tau fragment peptides (VQIVYK) by applying the discontinuous molecular dynamics method to a system of 48 VQIVYK peptides modeled using a new protein model/force field, PRIME20. The aim of the article is to ascertain which factors are most important in determining whether or not a peptide system forms perfect coherent fibrillar structures. Two different directional criteria are used to determine when a hydrogen bond occurs: the original H-bond constraints and a parallel preference H-bond constraint that imparts a slight bias towards the formation of parallel versus antiparallel strands in a ß-sheet. Under the original H-bond constraints, the resulting fibrillar structures contain a mixture of parallel and antiparallel pairs of strands within each ß-sheet over the whole fibrillization temperature range. Under the parallel preference H-bond constraints, the ß-sheets within the fibrillar structures are more likely to be parallel and indeed become perfectly parallel, consistent with X-ray crystallography, at a high temperature slightly below the fibrillization temperature. The high temperature environment encourages the formation of perfect fibril structures by providing enough time and space for peptides to rearrange during the aggregation process. There are two different kinetic mechanisms, template assembly with monomer addition at high temperature and merging/rearrangement without monomer addition at low temperature, which lead to significant differences in the final fibrillar structure. This suggests that the diverse fibril morphologies generally observed in vitro depend more on environmental conditions than has heretofore been appreciated.

Fibrillization propensity for short designed hexapeptides predicted by computer simulation.

Assembly of normally soluble proteins into ordered aggregates, known as amyloid fibrils, is a cause or associated symptom of numerous human disorders, including Alzheimer's and the prion diseases. Here, we test the ability of discontinuous molecular dynamics (DMD) simulations based on PRIME20, a new intermediate-resolution protein force field, to predict which designed hexapeptide sequences will form fibrils, which will not, and how this depends on temperature and concentration. Simulations were performed on 48-peptide systems containing STVIIE, STVIFE, STVIVE, STAIIE, STVIAE, STVIGE, and STVIEE starting from random-coil configurations. By the end of the simulations, STVIIE and STVIFE (which form fibrils in vitro) form fibrils over a range of temperatures, STVIEE (which does not form fibrils in vitro) does not form fibrils, and STVIVE, STAIIE, STVIAE, and STVIGE (which do not form fibrils in vitro) form fibrils at lower temperatures but stop forming fibrils at higher temperatures. At the highest temperatures simulated, the results on the fibrillization propensity of the seven short de novo designed peptides all agree with the experiments of López de la Paz and Serrano. Our results suggest that the fibrillization temperature (temperature above which fibrils cease to form) is a measure of fibril stability and that by rank ordering the fibrillization temperatures of various sequences, PRIME20/DMD simulations could be used to ascertain their relative fibrillization propensities. A phase diagram showing regions in the temperature-concentration plane where fibrils are formed in our simulations is presented.