Search

Proteins get used in a wide variety of applications. For some applications, a specific protein property needs to be improved. Deep mutational scanning data can then be used to train a neural network, which can predict the effect of variants that have not been experimentally determined. Many approaches use the protein’s sequence in one form or another as data representation for their prediction method. Since proteins are not one-dimensional strings but feature a complex 3D structure containing interacting residues, using the protein’s structure as input brings the advantage of already encoding the interactions rather than having to learn them. Here we present a new representation for describing protein structures using stacked 2D contact maps. The contact maps represent residue interactions, their evolutionary conservation, and changes in interactions that occur due to mutations. Additionally, we provide different methods to improve the performance when training neural networks on small (50 - 2000 data points big) deep mutational scanning data sets. To test this approach, we trained three neural network architectures on three deep mutational scanning datasets and assessed their performance compared to a neural network trained on protein sequences alone. We demonstrate the utility of this representation for training neural networks to predict the outcome of deep mutational scanning experiments. The use of structural representation, as well as data augmentation and pre-training, substantially reduced the training data requirements and led to improved performance on smaller data sets and comparable performance when bigger data sets are available compared to the state-of-the-art sequence convolutional neural network, which demonstrates the power of our approach. This work could pave the way to expanded usage of deep mutational scanning by substantially reducing experimental requirements and opens the door for smaller facilities that might not have access to high throughput screening.Author summaryWe introduce a novel approach to improve predictions of protein properties based on deep mutational scanning (DMS) data. Our method utilizes a unique data representation for protein structures through stacked 2D contact maps, which capture residue interactions and alterations in these interactions due to amino acid substitutions, as well as evolutionary conservation. We also tested several techniques to enhance the performance of neural networks when trained on small datasets. By training three neural network architectures on three public DMS datasets, we demonstrate the value of integrating protein structure information in training neural networks to predict the effects of amino acid substitutions. Utilizing structural information along with data augmentation and pre-training significantly improves performance on smaller datasets compared to state-of-the-art sequence convolutional neural networks while achieving similar performance on larger datasets. Thus, our approach could substantially reduce the experimental input needed and provide valuable predictions even for smaller DMS datasets when high-throughput screening is not feasible.

Many membraneless organelles are thought to be biomolecular condensates formed by phase separation of proteins and other biopolymers. Post-translational modifications (PTMs) can impact protein phase separation behavior, although for many PTMs this aspect of their function is unknown. O-linked β-D-N-acetylglucosaminylation (O-GlcNAcylation) is an abundant form of intracellular glycosylation whose roles in regulating biomolecular condensate assembly and dynamics have not been delineated. Using an in vitro approach, we found that O-GlcNAcylation reduces the phase separation propensity of the EWS N-terminal low complexity region (LCRN) under different conditions, including in the presence of the arginine- and glycine-rich RNA-binding domains (RBD). O-GlcNAcylation enhances fluorescence recovery after photobleaching (FRAP) within EWS LCRN condensates and causes the droplets to exhibit more liquid-like relaxation following fusion. Following extended incubation times, EWS LCRN+RBD condensates exhibit diminished FRAP, indicating a loss of fluidity, while condensates containing the O-GlcNAcylated LCRN do not. In HeLa cells, EWS is less O-GlcNAcylated following OGT knockdown, which correlates with its increased accumulation in a filter retardation assay. Relative to the human proteome, O-GlcNAcylated proteins are enriched with regions that are predicted to phase separate, suggesting a general role of O-GlcNAcylation in regulation of biomolecular condensates.

The AlphaFold Protein Structure Database contains predicted structures for millions of proteins. For the majority of human proteins that contain intrinsically disordered regions (IDRs), which do not adopt a stable structure, it is generally assumed these regions have low AlphaFold2 confidence scores that reflect low-confidence structural predictions. Here, we show that AlphaFold2 assigns confident structures to nearly 15% of human IDRs. By comparison to experimental NMR data for a subset of IDRs that are known to conditionally fold (i.e., upon binding or under other specific conditions), we find that AlphaFold2 often predicts the structure of the conditionally folded state. Based on databases of IDRs that are known to conditionally fold, we estimate that AlphaFold2 can identify conditionally folding IDRs at a precision as high as 88% at a 10% false positive rate, which is remarkable considering that conditionally folded IDR structures were minimally represented in its training data. We find that human disease mutations are nearly 5-fold enriched in conditionally folded IDRs over IDRs in general, and that up to 80% of IDRs in prokaryotes are predicted to conditionally fold, compared to less than 20% of eukaryotic IDRs. These results indicate that a large majority of IDRs in the proteomes of human and other eukaryotes function in the absence of conditional folding, but the regions that do acquire folds are more sensitive to mutations. We emphasize that the AlphaFold2 predictions do not reveal functionally relevant structural plasticity within IDRs and cannot offer realistic ensemble representations of conditionally folded IDRs.Significance StatementAlphaFold2 and other machine learning-based methods can accurately predict the structures of most proteins. However, nearly two-thirds of human proteins contain segments that are highly flexible and do not autonomously fold, otherwise known as intrinsically disordered regions (IDRs). In general, IDRs interconvert rapidly between a large number of different conformations, posing a significant problem for protein structure prediction methods that define one or a small number of stable conformations. Here, we found that AlphaFold2 can readily identify structures for a subset of IDRs that fold under certain conditions (conditional folding). We leverage AlphaFold2’s predictions of conditionally folded IDRs to quantify the extent of conditional folding across the tree of life, and to rationalize disease-causing mutations in IDRs.Classifications: Biological Sciences; Biophysics and Computational Biology

Autism Spectrum Disorder (ASD) is genetically complex with ~100 copy number variants and genes involved. To try to establish more definitive genotype and phenotype correlations in ASD, we searched genome sequence data, and the literature, for recurrent predicted damaging sequence-level variants affecting single genes. We identified 18 individuals from 16 unrelated families carrying a heterozygous guanine duplication (c.3679dup; p.Ala1227Glyfs*69) occurring within a string of 8 guanines (genomic location [hg38]g.50,721,512dup) affecting SHANK3, a prototypical ASD gene (0.08% of ASD-affected individuals carried the predicted p.Ala1227Glyfs*69 frameshift variant). Most probands carried de novo mutations, but five individuals in three families inherited it through somatic mosaicism. We scrutinized the phenotype of p.Ala1227Glyfs*69 carriers, and while everyone (17/17) formally tested for ASD carried a diagnosis, there was the variable expression of core ASD features both within and between families. Defining such recurrent mutational mechanisms underlying an ASD outcome is important for genetic counseling and early intervention.

The precise regulation of gene expression is fundamental to neurodevelopment, plasticity, and cognitive function. While several studies have deeply profiled mRNA dynamics in the developing human brain, there is a fundamental gap in our understanding of accompanying translational regulation. We perform ribosome profiling from more than 70 human prenatal and adult cortex samples across ontogeny and into adulthood, mapping translation events at nucleotide resolution. In addition to characterizing the translational regulation of annotated open reading frames (ORFs), we identify thousands of previously unknown translation events, including small open reading frames (sORFs) that give rise to human- and/or brain-specific microproteins, many of which we independently verify using size-selected proteomics. Ribosome profiling in stem cell-derived human neuronal cultures further corroborates these findings and shows that several neuronal activity-induced long non-coding RNAs (lncRNAs), including LINC00473, a primate-specific lncRNA implicated in depression, encode previously undescribed microproteins. Physicochemical analysis of these brain microproteinss identifies a large class harboring arginine-glycine-glycine (RGG) repeats as strong candidates for regulating RNA metabolism. Moreover, we find that, collectively, these previously unknown human brain sORFs are enriched for variants associated with schizophrenia. In addition to significantly expanding the translational landscape of the developing brain, this atlas will serve as a rich resource for the annotation and functional interrogation of thousands of previously unknown brain-specific protein products.

A major challenge to the characterization of intrinsically disordered regions (IDRs), which are widespread in the proteome, but relatively poorly understood, is the identification of molecular features that mediate functions of these regions, such as short motifs, amino acid repeats and physicochemical properties. Here, we introduce a proteome-scale feature discovery approach for IDRs. Our approach, which we call “reverse homology”, exploits the principle that important functional features are conserved over evolution. We use this as a contrastive learning signal for deep learning: given a set of homologous IDRs, the neural network has to correctly choose a held-out homologue from another set of IDRs sampled randomly from the proteome. We pair reverse homology with a simple architecture and standard interpretation techniques, and show that the network learns conserved features of IDRs that can be interpreted as motifs, repeats, or bulk features like charge or amino acid propensities. We also show that our model can be used to produce visualizations of what residues and regions are most important to IDR function, generating hypotheses for uncharacterized IDRs. Our results suggest that feature discovery using unsupervised neural networks is a promising avenue to gain systematic insight into poorly understood protein sequences.

Isotopically labeled methyl groups provide NMR probes in large, otherwise deuterated proteins. However, the resonance assignment constitutes a bottleneck for broader applicability of methyl-based NMR. Here, we present the automated MethylFLYA method for the assignment of methyl groups that is based on methyl-methyl nuclear Overhauser effect spectroscopy (NOESY) peak lists. MethylFLYA is applied to five proteins (28–358 kDa) comprising a total of 708 isotope-labeled methyl groups, of which 612 contribute NOESY cross peaks. MethylFLYA confidently assigns 488 methyl groups, i.e. 80% of those with NOESY data. Of these, 459 agree with the reference, 6 were different, and 23 were without reference assignment. MethylFLYA assigns significantly more methyl groups than alternative algorithms, has an average error rate of 1%, modest runtimes of 0.4–1.2 h, and can handle arbitrary isotope labeling patterns and data from other types of NMR spectra., Nature Communications, 10 (1), ISSN:2041-1723

HSP27 is a human molecular chaperone that forms large, dynamic oligomers and functions in many aspects of cellular homeostasis. Mutations in HSP27 cause Charcot‐Marie‐Tooth (CMT) disease, the most common inherited disorder of the peripheral nervous system. A particularly severe form of CMT disease is triggered by the P182L mutation in the highly conserved IxI/V motif of the disordered C‐terminal region, which interacts weakly with the structured core domain of HSP27. Here, we observed that the P182L mutation disrupts the chaperone activity and significantly increases the size of HSP27 oligomers formed in vivo, including in motor neurons differentiated from CMT patient‐derived stem cells. Using NMR spectroscopy, we determined that the P182L mutation decreases the affinity of the HSP27 IxI/V motif for its own core domain, leaving this binding site more accessible for other IxI/V‐containing proteins. We identified multiple IxI/V‐bearing proteins that bind with higher affinity to the P182L variant due to the increased availability of the IxI/V‐binding site. Our results provide a mechanistic basis for the impact of the P182L mutation on HSP27 and suggest that the IxI/V motif plays an important, regulatory role in modulating protein–protein interactions., NMR studies define how the P182L mutation alters intramolecular interactions to increase HSP27 accessibility for numerous binding proteins, compromising its chaperone function in patient cell‐derived motor neurons.

Alternative splicing plays a major role in shaping tissue-specific transcriptomes. Among the broad tissue types present in metazoans, the central nervous system contains some of the highest levels of alternative splicing. Although many documented examples of splicing differences between broad tissue types exist, there remains much to be understood about the splicing factors and the cis sequence elements controlling tissue and neuron subtype-specific splicing patterns. By using translating ribosome affinity purification coupled with deep-sequencing (TRAP-seq) in Caenorhabditis elegans, we have obtained high coverage profiles of ribosome-associated mRNA for three broad tissue classes (nervous system, muscle, and intestine) and two neuronal subtypes (dopaminergic and serotonergic neurons). We have identified hundreds of splice junctions that exhibit distinct splicing patterns between tissue types or within the nervous system. Alternative splicing events differentially regulated between tissues are more often frame-preserving, are more highly conserved across Caenorhabditis species, and are enriched in specific cis regulatory motifs, when compared with other types of exons. By using this information, we have identified a likely mechanism of splicing repression by the RNA-binding protein UNC-75/CELF via interactions with cis elements that overlap a 5′ splice site. Alternatively spliced exons also overlap more frequently with intrinsically disordered peptide regions than constitutive exons. Moreover, regulated exons are often shorter than constitutive exons but are flanked by longer intron sequences. Among these tissue-regulated exons are several highly conserved microexons C. elegans that parallels the evolutionary forces and constraints observed across metazoa.

It is unclear how disease mutations impact intrinsically disordered protein regions (IDRs), which lack a stable folded structure. These mutations, while prevalent in disease, are frequently neglected or annotated as variants of unknown significance. Biomolecular phase separation, a physical process often mediated by IDRs, has increasingly appreciated roles in cellular organization and regulation. We find that autism spectrum disorder (ASD)- and cancer-associated proteins are enriched for predicted phase separation propensities, suggesting that IDR mutations disrupt phase separation in key cellular processes. More generally, we hypothesize that combinations of small-effect IDR mutations perturb phase separation, potentially contributing to "missing heritability" in complex disease susceptibility.

What do we know about the molecular evolution of functional protein condensation? The capacity of proteins to form biomolecular condensates (compact, protein-rich states, not bound by membranes, but still separated from the rest of the contents of the cell) appears in many cases to be bestowed by weak, transient interactions within one or between proteins. Natural selection is expected to remove or fix amino acid changes, insertions or deletions that preserve and change this condensation capacity when doing so is beneficial to the cell. A few recent studies have begun to explore this frontier of phylogenetics at the intersection of biophysics and cell biology.

Previously, we showed that intrinsically disordered regions in proteins (IDRs) contain multiple sequence-distributed molecular features that are conserved over evolution, despite little sequence similarity that can be detected in alignments (Zarin et al. 2019). Here, we aim to use these molecular features to make specific functional predictions for individual IDRs and identify the molecular features within them that are responsible for specific functions. We find that the predictable functions are diverse, identifying previously known associated molecular features, as well as features that were previously not known to be associated with these functions. We experimentally confirm that elevated isoelectric point and hydrophobicity, features that are positively associated with mitochondrial localization, are necessary for mitochondrial targeting function. Remarkably, increasing isoelectric point in a synthetic IDR restores weak mitochondrial targeting. We believe feature analysis represents a new systematic approach to understand how biological functions of IDRs are specified by their protein sequences.

As structural biology trends towards larger and more complex biomolecular targets, a detailed understanding of their interactions and underlying structures and dynamics is required. The development of methyl-TROSY has enabled NMR spectroscopy to provide atomic-resolution insight into the mechanisms of large molecular assemblies in solution. However, the applicability of methyl-TROSY has been hindered by the laborious and time-consuming resonance assignment process, typically performed with domain fragmentation, site-directed mutagenesis, and analysis of NOE data in the context of a crystal structure. In response, several structure-based automatic methyl assignment strategies have been developed over the past decade. Here, we present a comprehensive analysis of all available methods and compare their input data requirements, algorithmic strategies, and reported performance. In general, the methods fall into two categories: those that primarily rely on inter-methyl NOEs, and those that utilize methyl PRE- and PCS-based restraints. We discuss their advantages and limitations, and highlight the potential benefits from standardizing and combining different methods.

Pulsed-field gradient NMR spectroscopy is widely used to measure the translational diffusion and hydrodynamic radius (R(h)) of biomolecules in solution. For unfolded proteins, the R(h) provides a sensitive reporter on the ensemble-averaged conformation and the extent of polypeptide chain expansion as a function of added denaturant. Hydrostatic pressure is a convenient and reversible alternative to chemical denaturants for the study of protein folding, and enables NMR measurements to be performed on a single sample. While the impact of pressure on the viscosity of water is well known, and our water diffusivity measurements agree closely with theoretical expectations, we find that elevated pressures increase the R(h) of dioxane and other small molecules by amounts that correlate with their hydrophobicity, with parallel increases in rotational friction indicated by (13)C longitudinal relaxation times. These data point to a tighter coupling with water for hydrophobic surfaces at elevated pressures. Translational diffusion measurement of the unfolded state of a pressure-sensitized ubiquitin mutant (VA2-ubiquitin) as a function of hydrostatic pressure or urea concentration shows that R(h) values of both the folded and the unfolded states remain nearly invariant. At ca 23 Å, the R(h) of the fully pressure-denatured state is essentially indistinguishable from the urea-denatured state, and close to the value expected for an idealized random coil of 76 residues. The intrinsically disordered protein (IDP) α-synuclein shows slight compaction at pressures above 2 kbar. Diffusion of unfolded ubiquitin and α-synuclein is significantly impacted by sample concentration, indicating that quantitative measurements need to be carried out under dilute conditions.

Methyl groups provide crucial NMR probes for investigating protein structure, dynamics and mechanisms in systems that are too large for NMR with uniform isotope labeling. This requires the assignment of methyl signals in the NMR spectra to specific methyl groups in the protein, an expensive and time-consuming endeavor that limits the use of methyl-based NMR for large proteins. To resolve this bottleneck, several methyl resonance assignment methods have been developed. These approaches remain limited with regard to complete automation and/or the extent and accuracy of the assignments. Here, we present the completely automated MethylFLYA method for the assignment of methyl groups. MethylFLYA requires as input exclusively methyl-methyl nuclear Overhauser effect spectroscopy (NOESY) peak lists. The algorithm was applied to five proteins of 28–358 kDa mass with a total of 708 isotope-labeled methyl groups. Manually made 1H/13C reference assignments were available for 674 methyls. The available experimental peak lists contained NOESY cross peaks for 614 methyls. MethylFLYA confidently assigned 488 methyls, i.e. 79% of those with NOESY data. Of these assignments, 460 agreed with the reference, 5 were different (and 23 concerned methyls without reference assignment). For three proteins of 28, 81, and 358 kDa, all confident assignments by MethylFLYA were correct. We furthermore show that, for high-quality NOESY spectra, automatic picking of NOE signals followed by resonance assignment with MethylFLYA can yield results that are comparable to those obtained for manually prepared peak lists, indicating the feasibility of unbiased, fully automatic methyl resonance assignment starting directly from the NMR spectra. This renders MethylFLYA an advantageous alternative to existing approaches for structure-based methyl assignment. MethylFLYA assigns, for most proteins, significantly more methyl groups than other algorithms, has an average error rate of 1%, modest runtimes of 0.4–1.2 h for the five proteins, and flexibility to handle arbitrary isotope labeling patterns and include data from other types of NMR spectra.

The small heat-shock protein HSP27 is a redox-sensitive molecular chaperone that is expressed throughout the human body. Here we describe redox-induced changes to the structure, dynamics, and function of HSP27 and its conserved α-crystallin domain, and provide the first structural characterization of a small heat-shock protein monomer. While HSP27 assembles into oligomers, we show that the transiently populated monomers released upon reduction are highly active chaperones in vitro, but are kinetically unstable and susceptible to uncontrolled aggregation. By using relaxation dispersion and high-pressure nuclear magnetic resonance spectroscopy, we reveal that the pair of β-strands that mediate dimerization become partially disordered in the monomer. Strikingly, we note that numerous HSP27 mutations associated with inherited neuropathies cluster to this unstructured region. The high degree of sequence conservation in the α-crystallin domain amongst mammalian sHSPs suggests that partially unfolded monomers may be a general, functional feature of these molecular chaperones.

The small heat-shock protein HSP27 is a redox-sensitive molecular chaperone that is expressed throughout the human body. Here, we describe redox-induced changes to the structure, dynamics, and function of HSP27 and its conserved α-crystallin domain (ACD). While HSP27 assembles into oligomers, we show that the monomers formed upon reduction are highly active chaperones in vitro, but are susceptible to self-aggregation. By using relaxation dispersion and high-pressure nuclear magnetic resonance (NMR) spectroscopy, we observe that the pair of β-strands that mediate dimerisation partially unfold in the monomer. We note that numerous HSP27 mutations associated with inherited neuropathies cluster to this dynamic region. High levels of sequence conservation in ACDs from mammalian sHSPs suggest that the exposed, disordered interface present in free monomers or oligomeric subunits may be a general, functional feature of sHSPs., The small heat-shock protein HSP27 occurs predominantly in oligomeric forms, which makes its structural characterisation challenging. Here the authors employ CPMG and high-pressure NMR with native mass spectrometry and biophysical assays to show that the active monomeric form of HSP27 is substantially disordered and highly chaperone-active.

Methyl groups are powerful probes for the analysis of structure, dynamics and function of supramolecular assemblies, using both solution- and solid-state NMR. Widespread application of the methodology has been limited due to the challenges associated with assigning spectral resonances to specific locations within a biomolecule. Here, we present Methyl Assignment by Graph MAtching (MAGMA), for the automatic assignment of methyl resonances. A graph matching protocol examines all possibilities for each resonance in order to determine an exact assignment that in-cludes a complete description of any ambiguity. MAGMA gives 100% accuracy in confident assignments when tested against both synthetic data, and 9 cross-validated examples using both solution- and solid-state NMR data. We show that this remarkable accuracy enables a user to distinguish between alternative protein structures. In a drug discovery application on HSP90, we show the method can rapidly and efficiently distinguish between possible ligand binding modes. By providing an exact and robust solution to methyl resonance assignment, MAGMA can facilitate significantly accelerated studies of supramolecular machines using methyl-based NMR spectroscopy.

High-resolution solution-state NMR spectroscopy studies of large proteins typically require uniform deuteration of the system and selective protonation and isotope labelling of methyl groups. Under such circumstances, the assignment of methyl resonances presents a considerable experimental challenge and automation of the process using computational algorithms has been actively sought. Through-space connectivities between the labelled methyl groups can be established through nuclear Overhauser enhancement spectroscopy (NOESY). If a high-resolution structure of the system is available, the sparse connectivity restraints derived from this information enable structure-based methyl resonance assignment. Here, we outline a protocol for full automation of the methyl resonance assignment process using the CYANA software package. We tested the protocol on three-dimensional (3D) 13C/13C-separated NOESY spectra of a dimer of regulatory chains of aspartate transcarbamoylase (ATCase-r2). We used CYPICK to detect NOE signals, followed by automatic resonance assignment with FLYA. On this dataset, FLYA generated highly similar results using either automatically or manually generated peak lists, confidently assigning ~60% of the methyl groups with high accuracy (95 ± 2% correctness). We compared this performance to two alternative automatic methyl assignment protocols, MAP-XSII and FLAMEnGO2.0, both of which, similarly to FLYA, support unassigned NOESY peak lists as input., Journal of Physics: Conference Series, 1036 (1), ISSN:1742-6588, ISSN:1742-6596