Hardware implementation of backpropagation using progressive gradient descent for in situ training of multilayer neural networks.

Neural network training can be slow and energy-expensive due to the frequent transfer of weight data between digital memory and processing units. Neuromorphic systems can accelerate neural networks by performing multiply-accumulate operations in parallel using nonvolatile analog memory. However, executing the widely used backpropagation training algorithm in multilayer neural networks requires information-and therefore storage-of the partial derivatives of the weight values preventing suitable and scalable implementation in hardware. Here, we propose a hardware implementation of the backpropagation algorithm that progressively updates each layer using in situ stochastic gradient descent, avoiding this storage requirement. We experimentally demonstrate the in situ error calculation and the proposed progressive backpropagation method in a multilayer hardware-implemented neural network. We confirm identical learning characteristics and classification performance compared to conventional backpropagation in software. We show that our approach can be scaled to large and deep neural networks, enabling highly efficient training of advanced artificial intelligence computing systems.

Cargo selective vesicle tethering: The structural basis for binding of specific cargo proteins by the Golgi tether component TBC1D23.

The Golgi-localized golgins golgin-97 and golgin-245 capture transport vesicles arriving from endosomes via the protein TBC1D23. The amino-terminal domain of TBC1D23 binds to the golgins, and the carboxyl-terminal domain of TBC1D23 captures the vesicles, but how it recognizes specific vesicles was unclear. A search for binding partners of the carboxyl-terminal domain unexpectedly revealed direct binding to carboxypeptidase D and syntaxin-16, known cargo proteins of the captured vesicles. Binding is via a threonine-leucine-tyrosine (TLY) sequence present in both proteins next to an acidic cluster. A crystal structure reveals how this acidic TLY motif binds to TBC1D23. An acidic TLY motif is also present in the tails of other endosome-to-Golgi cargo, and these also bind TBC1D23. Structure-guided mutations in the carboxyl-terminal domain that disrupt motif binding in vitro also block vesicle capture in vivo. Thus, TBC1D23 attached to golgin-97 and golgin-245 captures vesicles by a previously undescribed mechanism: the recognition of a motif shared by cargo proteins carried by the vesicle.

Enhancer-promoter interactions are reconfigured through the formation of long-range multiway hubs as mouse ES cells exit pluripotency.

Enhancers bind transcription factors, chromatin regulators, and non-coding transcripts to modulate the expression of target genes. Here, we report 3D genome structures of single mouse ES cells as they are induced to exit pluripotency and transition through a formative stage prior to undergoing neuroectodermal differentiation. We find that there is a remarkable reorganization of 3D genome structure where inter-chromosomal intermingling increases dramatically in the formative state. This intermingling is associated with the formation of a large number of multiway hubs that bring together enhancers and promoters with similar chromatin states from typically 5-8 distant chromosomal sites that are often separated by many Mb from each other. In the formative state, genes important for pluripotency exit establish contacts with emerging enhancers within these multiway hubs, suggesting that the structural changes we have observed may play an important role in modulating transcription and establishing new cell identities.

Reconstructing diploid 3D chromatin structures from single cell Hi-C data with a polymer-based approach.

Detailed understanding of the 3D structure of chromatin is a key ingredient to investigate a variety of processes inside the cell. Since direct methods to experimentally ascertain these structures lack the desired spatial fidelity, computational inference methods based on single cell Hi-C data have gained significant interest. Here, we develop a progressive simulation protocol to iteratively improve the resolution of predicted interphase structures by maximum-likelihood association of ambiguous Hi-C contacts using lower-resolution predictions. Compared to state-of-the-art methods, our procedure is not limited to haploid cell data and allows us to reach a resolution of up to 5,000 base pairs per bead. High resolution chromatin models grant access to a multitude of structural phenomena. Exemplarily, we verify the formation of chromosome territories and holes near aggregated chromocenters as well as the inversion of the CpG content for rod photoreceptor cells.

Functional unknomics: Systematic screening of conserved genes of unknown function.

The human genome encodes approximately 20,000 proteins, many still uncharacterised. It has become clear that scientific research tends to focus on well-studied proteins, leading to a concern that poorly understood genes are unjustifiably neglected. To address this, we have developed a publicly available and customisable "Unknome database" that ranks proteins based on how little is known about them. We applied RNA interference (RNAi) in Drosophila to 260 unknown genes that are conserved between flies and humans. Knockdown of some genes resulted in loss of viability, and functional screening of the rest revealed hits for fertility, development, locomotion, protein quality control, and resilience to stress. CRISPR/Cas9 gene disruption validated a component of Notch signalling and 2 genes contributing to male fertility. Our work illustrates the importance of poorly understood genes, provides a resource to accelerate future research, and highlights a need to support database curation to ensure that misannotation does not erode our awareness of our own ignorance.

De novo design of modular peptide-binding proteins by superhelical matching.

General approaches for designing sequence-specific peptide-binding proteins would have wide utility in proteomics and synthetic biology. However, designing peptide-binding proteins is challenging, as most peptides do not have defined structures in isolation, and hydrogen bonds must be made to the buried polar groups in the peptide backbone1-3. Here, inspired by natural and re-engineered protein-peptide systems4-11, we set out to design proteins made out of repeating units that bind peptides with repeating sequences, with a one-to-one correspondence between the repeat units of the protein and those of the peptide. We use geometric hashing to identify protein backbones and peptide-docking arrangements that are compatible with bidentate hydrogen bonds between the side chains of the protein and the peptide backbone12. The remainder of the protein sequence is then optimized for folding and peptide binding. We design repeat proteins to bind to six different tripeptide-repeat sequences in polyproline II conformations. The proteins are hyperstable and bind to four to six tandem repeats of their tripeptide targets with nanomolar to picomolar affinities in vitro and in living cells. Crystal structures reveal repeating interactions between protein and peptide interactions as designed, including ladders of hydrogen bonds from protein side chains to peptide backbones. By redesigning the binding interfaces of individual repeat units, specificity can be achieved for non-repeating peptide sequences and for disordered regions of native proteins.

Zasp52 strengthens whole embryo tissue integrity through supracellular actomyosin networks.

During morphogenesis, large-scale changes of tissue primordia are coordinated across an embryo. In Drosophila, several tissue primordia and embryonic regions are bordered or encircled by supracellular actomyosin cables, junctional actomyosin enrichments networked between many neighbouring cells. We show that the single Drosophila Alp/Enigma-family protein Zasp52, which is most prominently found in Z-discs of muscles, is a component of many supracellular actomyosin structures during embryogenesis, including the ventral midline and the boundary of the salivary gland placode. We reveal that Zasp52 contains within its central coiled-coil region a type of actin-binding motif usually found in CapZbeta proteins, and this domain displays actin-binding activity. Using endogenously-tagged lines, we identify that Zasp52 interacts with junctional components, including APC2, Polychaetoid and Sidekick, and actomyosin regulators. Analysis of zasp52 mutant embryos reveals that the severity of the embryonic defects observed scales inversely with the amount of functional protein left. Large tissue deformations occur where actomyosin cables are found during embryogenesis, and in vivo and in silico analyses suggest a model whereby supracellular Zasp52-containing cables aid to insulate morphogenetic changes from one another.

Adaptive Biosensing and Neuromorphic Classification Based on an Ambipolar Organic Mixed Ionic-Electronic Conductor.

Organic mixed ionic-electronic conductors (OMIECs) are central to bioelectronic applications such as biosensors, health-monitoring devices, and neural interfaces, and have facilitated efficient next-generation brain-inspired computing and biohybrid systems. Despite these examples, smart and adaptive circuits that can locally process and optimize biosignals have not yet been realized. Here, a tunable sensing circuit is shown that can locally modulate biologically relevant signals like electromyograms (EMGs) and electrocardiograms (ECGs), that is based on a complementary logic inverter combined with a neuromorphic memory element, and that is constructed from a single polymer mixed conductor. It is demonstrated that a small neuromorphic array based on this material effects high classification accuracy in heartbeat anomaly detection. This high-performance material allows for straightforward monolithic integration, which reduces fabrication complexity while also achieving high on/off ratios with excellent ambient p- and n-type stability in transistor performance. This material opens a route toward simple and straightforward fabrication and integration of more sophisticated adaptive circuits for future smart bioelectronics.

GOLPH3 and GOLPH3L are broad-spectrum COPI adaptors for sorting into intra-Golgi transport vesicles.

The fidelity of Golgi glycosylation is, in part, ensured by compartmentalization of enzymes within the stack. The COPI adaptor GOLPH3 has been shown to interact with the cytoplasmic tails of a subset of Golgi enzymes and direct their retention. However, other mechanisms of retention, and other roles for GOLPH3, have been proposed, and a comprehensive characterization of the clientele of GOLPH3 and its paralogue GOLPH3L is lacking. GOLPH3's role is of particular interest as it is frequently amplified in several solid tumor types. Here, we apply two orthogonal proteomic methods to identify GOLPH3+3L clients and find that they act in diverse glycosylation pathways or have other roles in the Golgi. Binding studies, bioinformatics, and a Golgi retention assay show that GOLPH3+3L bind the cytoplasmic tails of their clients through membrane-proximal positively charged residues. Furthermore, deletion of GOLPH3+3L causes multiple defects in glycosylation. Thus, GOLPH3+3L are major COPI adaptors that impinge on most, if not all, of the glycosylation pathways of the Golgi.

A Comprehensive Subcellular Atlas of the Toxoplasma Proteome via hyperLOPIT Provides Spatial Context for Protein Functions.

Apicomplexan parasites cause major human disease and food insecurity. They owe their considerable success to highly specialized cell compartments and structures. These adaptations drive their recognition, nondestructive penetration, and elaborate reengineering of the host's cells to promote their growth, dissemination, and the countering of host defenses. The evolution of unique apicomplexan cellular compartments is concomitant with vast proteomic novelty. Consequently, half of apicomplexan proteins are unique and uncharacterized. Here, we determine the steady-state subcellular location of thousands of proteins simultaneously within the globally prevalent apicomplexan parasite Toxoplasma gondii. This provides unprecedented comprehensive molecular definition of these unicellular eukaryotes and their specialized compartments, and these data reveal the spatial organizations of protein expression and function, adaptation to hosts, and the underlying evolutionary trajectories of these pathogens.

Rho-Kinase Planar Polarization at Tissue Boundaries Depends on Phospho-regulation of Membrane Residence Time.

The myosin II activator Rho-kinase (Rok) is planar polarized at the tissue boundary of the Drosophila embryonic salivary gland placode through a negative regulation by the apical polarity protein Crumbs that is anisotropically localized at the boundary. However, in inner cells of the placode, both Crumbs and Rok are isotropically enriched at junctions. We propose that modulation of Rok membrane residence time by Crumbs' downstream effectors can reconcile both behaviors. Using FRAP combined with in silico simulations, we find that the lower membrane dissociation rate (koff) of Rok at the tissue boundary with low Crumbs explains this boundary-specific effect. The S/T kinase Pak1, recruited by Crumbs and Cdc42, negatively affects Rok membrane association in vivo and in vitro can phosphorylate Rok near the pleckstrin homology (PH) domain that mediates membrane association. These data reveal an important mechanism of the modulation of Rok membrane residence time via affecting the koff that may be widely employed during tissue morphogenesis.

Proteome Mapping of a Cyanobacterium Reveals Distinct Compartment Organization and Cell-Dispersed Metabolism.

Cyanobacteria are complex prokaryotes, incorporating a Gram-negative cell wall and internal thylakoid membranes (TMs). However, localization of proteins within cyanobacterial cells is poorly understood. Using subcellular fractionation and quantitative proteomics, we produced an extensive subcellular proteome map of an entire cyanobacterial cell, identifying ∼67% of proteins in Synechocystis sp. PCC 6803, ∼1000 more than previous studies. Assigned to six specific subcellular regions were 1,712 proteins. Proteins involved in energy conversion localized to TMs. The majority of transporters, with the exception of a TM-localized copper importer, resided in the plasma membrane (PM). Most metabolic enzymes were soluble, although numerous pathways terminated in the TM (notably those involved in peptidoglycan monomer, NADP+, heme, lipid, and carotenoid biosynthesis) or PM (specifically, those catalyzing lipopolysaccharide, molybdopterin, FAD, and phylloquinol biosynthesis). We also identified the proteins involved in the TM and PM electron transport chains. The majority of ribosomal proteins and enzymes synthesizing the storage compound polyhydroxybuyrate formed distinct clusters within the data, suggesting similar subcellular distributions to one another, as expected for proteins operating within multicomponent structures. Moreover, heterogeneity within membrane regions was observed, indicating further cellular complexity. Cyanobacterial TM protein localization was conserved in Arabidopsis (Arabidopsis thaliana) chloroplasts, suggesting similar proteome organization in more developed photosynthetic organisms. Successful application of this technique in Synechocystis suggests it could be applied to mapping the proteomes of other cyanobacteria and single-celled organisms. The organization of the cyanobacterial cell revealed here substantially aids our understanding of these environmentally and biotechnologically important organisms.

Separating Golgi Proteins from Cis to Trans Reveals Underlying Properties of Cisternal Localization.

The order of enzymatic activity across Golgi cisternae is essential for complex molecule biosynthesis. However, an inability to separate Golgi cisternae has meant that the cisternal distribution of most resident proteins, and their underlying localization mechanisms, are unknown. Here, we exploit differences in surface charge of intact cisternae to perform separation of early to late Golgi subcompartments. We determine protein and glycan abundance profiles across the Golgi; over 390 resident proteins are identified, including 136 new additions, with over 180 cisternal assignments. These assignments provide a means to better understand the functional roles of Golgi proteins and how they operate sequentially. Protein and glycan distributions are validated in vivo using high-resolution microscopy. Results reveal distinct functional compartmentalization among resident Golgi proteins. Analysis of transmembrane proteins shows several sequence-based characteristics relating to pI, hydrophobicity, Ser abundance, and Phe bilayer asymmetry that change across the Golgi. Overall, our results suggest that a continuum of transmembrane features, rather than discrete rules, guide proteins to earlier or later locations within the Golgi stack.

Rate of Cognitive Decline in Alzheimer's Disease Stratified by Age.

BACKGROUND: There is only limited information available about the effect of age on course of cognitive decline in patients with onset of Alzheimer's disease (AD) over the age of 64 years. OBJECTIVE: We compared the rate of, and factors affecting, cognitive decline in patients with AD aged < 65 years (young-onset AD), 65-74 years (middle-onset AD), and ≥75 years (late-onset AD). METHOD: The study used longitudinal data from the Essex Memory Clinic which included a total of 305 participants; 56 had YOAD, 73 had MOAD, and 176 had LOAD. The rate of cognitive decline was measured using scores from the Mini-Mental State Examination (MMSE), and the data were examined using multilevel models analysis. RESULTS: There was evidence of a difference in cognitive decline across the age groups with the YOAD group declining 2.8 MMSE points per year, those with MOAD declined 2.0 MMSE points per year, and the LOAD group declined 1.4 MMSE points per year. CONCLUSIONS: Patients with LOAD have a better prognosis than YOAD and MOAD. However, even between the MOAD and LOAD groups, age is a significant predictor of cognitive decline, with older patients having a more benign course.

Triparental inheritance in Dictyostelium.

Sex promotes the recombination and reassortment of genetic material and is prevalent across eukaryotes, although our knowledge of the molecular details of sexual inheritance is scant in several major lineages. In social amoebae, sex involves a promiscuous mixing of cytoplasm before zygotes consume the majority of cells, but for technical reasons, sexual progeny have been difficult to obtain and study. We report here genome-wide characterization of meiotic progeny in Dictyostelium discoideum We find that recombination occurs at high frequency in pairwise crosses between all three mating types, despite the absence of the Spo11 enzyme that is normally required to initiate crossover formation. Fusions of more than two gametes to form transient syncytia lead to frequent triparental inheritance, with haploid meiotic progeny bearing recombined nuclear haplotypes from two parents and the mitochondrial genome from a third. Cells that do not contribute genetically to the Dictyostelium zygote nucleus thereby have a stake in the next haploid generation. D. discoideum mitochondrial genomes are polymorphic, and our findings raise the possibility that some of this variation might be a result of sexual selection on genes that can promote the spread of individual organelle genomes during sex. This kind of self-interested mitochondrial behavior may have had important consequences during eukaryogenesis and the initial evolution of sex.

Canonical and single-cell Hi-C reveal distinct chromatin interaction sub-networks of mammalian transcription factors.

BACKGROUND: Transcription factor (TF) binding to regulatory DNA sites is a key determinant of cell identity within multi-cellular organisms and has been studied extensively in relation to site affinity and chromatin modifications. There has been a strong focus on the inference of TF-gene regulatory networks and TF-TF physical interaction networks. Here, we present a third type of TF network, the spatial network of co-localized TF binding sites within the three-dimensional genome. RESULTS: Using published canonical Hi-C data and single-cell genome structures, we assess the spatial proximity of a genome-wide array of potential TF-TF co-localizations in human and mouse cell lines. For individual TFs, the abundance of occupied binding sites shows a positive correspondence with their clustering in three dimensions, and this is especially apparent for weak TF binding sites and at enhancer regions. An analysis between different TF proteins identifies significantly proximal pairs, which are enriched in reported physical interactions. Furthermore, clustering of different TFs based on proximity enrichment identifies two partially segregated co-localization sub-networks, involving different TFs in different cell types. Using data from both human lymphoblastoid cells and mouse embryonic stem cells, we find that these sub-networks are enriched within, but not exclusive to, different chromosome sub-compartments that have been identified previously in Hi-C data. CONCLUSIONS: This suggests that the association of TFs within spatial networks is closely coupled to gene regulatory networks. This applies to both differentiated and undifferentiated cells and is a potential causal link between lineage-specific TF binding and chromosome sub-compartment segregation.

Combining fluorescence imaging with Hi-C to study 3D genome architecture of the same single cell.

Fluorescence imaging and chromosome conformation capture assays such as Hi-C are key tools for studying genome organization. However, traditionally, they have been carried out independently, making integration of the two types of data difficult to perform. By trapping individual cell nuclei inside a well of a 384-well glass-bottom plate with an agarose pad, we have established a protocol that allows both fluorescence imaging and Hi-C processing to be carried out on the same single cell. The protocol identifies 30,000-100,000 chromosome contacts per single haploid genome in parallel with fluorescence images. Contacts can be used to calculate intact genome structures to better than 100-kb resolution, which can then be directly compared with the images. Preparation of 20 single-cell Hi-C libraries using this protocol takes 5 d of bench work by researchers experienced in molecular biology techniques. Image acquisition and analysis require basic understanding of fluorescence microscopy, and some bioinformatics knowledge is required to run the sequence-processing tools described here.

Prion-like protein aggregates exploit the RHO GTPase to cofilin-1 signaling pathway to enter cells.

Protein aggregation is a hallmark of diverse neurodegenerative diseases. Multiple lines of evidence have revealed that protein aggregates can penetrate inside cells and spread like prions. How such aggregates enter cells remains elusive. Through a focused siRNA screen targeting genes involved in membrane trafficking, we discovered that mutant SOD1 aggregates, like viruses, exploit cofilin-1 to remodel cortical actin and enter cells. Upstream of cofilin-1, signalling from the RHO GTPase and the ROCK1 and LIMK1 kinases controls cofilin-1 activity to remodel actin and modulate aggregate entry. In the spinal cord of symptomatic SOD1G93A transgenic mice, cofilin-1 phosphorylation is increased and actin dynamics altered. Importantly, the RHO to cofilin-1 signalling pathway also modulates entry of tau and α-synuclein aggregates. Our results identify a common host cell signalling pathway that diverse protein aggregates exploit to remodel actin and enter cells.

Calculation of 3D genome structures for comparison of chromosome conformation capture experiments with microscopy: An evaluation of single-cell Hi-C protocols.

Single-cell chromosome conformation capture approaches are revealing the extent of cell-to-cell variability in the organization and packaging of genomes. These single-cell methods, unlike their multi-cell counterparts, allow straightforward computation of realistic chromosome conformations that may be compared and combined with other, independent, techniques to study 3D structure. Here we discuss how single-cell Hi-C and subsequent 3D genome structure determination allows comparison with data from microscopy. We then carry out a systematic evaluation of recently published single-cell Hi-C datasets to establish a computational approach for the evaluation of single-cell Hi-C protocols. We show that the calculation of genome structures provides a useful tool for assessing the quality of single-cell Hi-C data because it requires a self-consistent network of interactions, relating to the underlying 3D conformation, with few errors, as well as sufficient longer-range cis- and trans-chromosomal contacts.

Multiple marker abundance profiling: combining selected reaction monitoring and data-dependent acquisition for rapid estimation of organelle abundance in subcellular samples.

Measuring changes in protein or organelle abundance in the cell is an essential, but challenging aspect of cell biology. Frequently-used methods for determining organelle abundance typically rely on detection of a very few marker proteins, so are unsatisfactory. In silico estimates of protein abundances from publicly available protein spectra can provide useful standard abundance values but contain only data from tissue proteomes, and are not coupled to organelle localization data. A new protein abundance score, the normalized protein abundance scale (NPAS), expands on the number of scored proteins and the scoring accuracy of lower-abundance proteins in Arabidopsis. NPAS was combined with subcellular protein localization data, facilitating quantitative estimations of organelle abundance during routine experimental procedures. A suite of targeted proteomics markers for subcellular compartment markers was developed, enabling independent verification of in silico estimates for relative organelle abundance. Estimation of relative organelle abundance was found to be reproducible and consistent over a range of tissues and growth conditions. In silico abundance estimations and localization data have been combined into an online tool, multiple marker abundance profiling, available in the SUBA4 toolbox (http://suba.live).