Predicting transcriptional activation domain function using Graph Neural Networks.

<h3>Abstract</h3> A variety of studies have suggested that low-complexity domains (LCDs) tend to be intrinsically disordered and are relatively rare within structured proteins in the protein data bank (PDB). Although LCDs are often treated as a single class, we previously found that LCDs enriched in different amino acids tend to exhibit substantial differences in protein metabolism and function. Therefore, we wondered whether the structural conformations of LCDs are likewise dependent on which specific amino acids are enriched within each LCD. Here, we directly examined relationships between enrichment of individual amino acids and secondary structure preferences across the entire PDB proteome. Secondary structure preferences varied as a function of the identity of the amino acid enriched and its degree of enrichment. Furthermore, divergence in secondary structure profiles often occurred for LCDs enriched in physicochemically similar amino acids (e.g. valine vs. leucine), indicating that LCDs composed of related amino acids can have distinct secondary structure preferences. Comparison of LCD secondary structure preferences with numerous pre-existing secondary structure propensity scales resulted in relatively poor correlations for certain types of LCDs, indicating that these scales may not capture secondary structure preferences as sequence complexity decreases. Collectively, these observations provide a highly resolved view of structural preferences among LCDs parsed by the nature and magnitude of single amino acid enrichment. <h3>Author Summary</h3> The structures that proteins adopt are directly related to their amino acid sequences. Low-complexity domains (LCDs) in protein sequences are unusual regions made up of only a few different types of amino acids. Although this is the key feature that classifies sequences as LCDs, the physical properties of LCDs will differ based on the types of amino acids that are found in each domain. For example, the sequences “AAAAAAAAAA”, “EEEEEEEEEE”, and “EEKRKEEEKE” will have very different properties, even though they would all be classified as LCDs by traditional methods. In a previous study, we developed a new method to further divide LCDs into categories that more closely reflect the differences in their physical properties. In this study, we apply that approach to examine the structures of LCDs when sorted into different categories based on their amino acids. This allowed us to define relationships between the types of amino acids in the LCDs and their corresponding structures. Since protein structure is closely related to protein function, this has important implications for understanding the basic functions and properties of LCDs in a variety of proteins.

The antioxidant-activated transcription factor nuclear factor erythroid 2-related factor 2 (Nrf2) regulates the induction of cytoprotective genes against chemical toxicity and oxidative injuries. The role of phosphorylation in Nrf2 activation has been suggested but remains elusive. We report that phenolic antioxidant/pro-oxidant tert-butylhydroquinone (tBHQ) induced two forms of the Nrf2 protein in neuroblastoma cells (IMR-32), which migrated as distinctive bands on SDS-PAGE. In vitro treatment with lambda phosphatase eliminated the slower migrating form and increased the amount of the faster migrating form of Nrf2. In vivo (32)Pi-phosphorylation resulted in (32)Pi-labeling of the Nrf2 protein in the presence of tBHQ that can be dephosphorylated by lambda phosphotase, indicating that the slower migrating form is a phosphorylated Nrf2 protein and the faster form an unphosphorylated Nrf2. Unphosphorylated Nrf2 predominated in the cytoplasm, whereas the phosphorylated form preferentially localized in the nucleus. Nuclear Nrf2 can be dephosphorylated by lambda phosphotase in vitro and be converted to the faster migrating form, implicating phosphorylation of Nrf2 in the cytoplasmic-nuclear translocation of the protein. Deletional analyses from both the carboxyl- and amino-ends revealed the transcription activation (TA) domains Neh4 (Nrf2-ECH homology 4) and Neh5 (Nrf2-ECH homology 5) as a major region necessary for the phosphorylation. The TA domains are characterized by the presence of multiple phosphorylation sites of casein kinase 2 (CK2). Moreover, CK2 phosphorylated the TA domains in vitro. Treatment with CK2 inhibitor 2-dimethylamino-4,5,6,7,-tetrabromo-1H-benzimidazole (DMAT) blocked the induction of endogenous target genes of Nrf2 in cells and inhibited the TA activities of both the full length and the TA domains of Nrf2 to a large extent. Finally, phosphorylation of the TA domains correlated with the nuclear translocation of Nrf2 that was inhibited by DMAT in a concentration-dependent manner. The findings demonstrated that phosphorylation of Nrf2 at the TA domains by CK2 is an integral component of Nrf2 activation necessary for the nuclear localization and transcription activation function of Nrf2 in neuroblastoma cells.

SUMO-Forest: A Cascade Forest based method for the prediction of SUMOylation sites on imbalanced data.

In a statistical approach to protein structure analysis, Miyazawa and Jernigan (MJ) derived a $20\times 20$ matrix of inter-residue contact energies between different types of amino acids. Using the method of eigenvalue decomposition, we find that the MJ matrix can be accurately reconstructed from its first two principal component vectors as $M_{ij}=C_0+C_1(q_i+q_j)+C_2 q_i q_j$, with constant $C$'s, and 20 $q$ values associated with the 20 amino acids. This regularity is due to hydrophobic interactions and a force of demixing, the latter obeying Hildebrand's solubility theory of simple liquids.

Porphyrin‐based nanomaterials have attracted much attention in photocatalysis due to their excellent photoelectrocatalytic ability and strong intermolecular interaction forces. Herein, a series of chiral assemblies of porphyrin–amino acid derivatives are prepared by linking different types of amino acids with terminal modifications to tetrakis(4‐carboxyphenyl)porphyrin via covalent bonding and their catalytic ability in enantioselective photocatalytic oxidation of tyrosine is investigated. The results show that the amino acid type and terminal groups significantly affect the macroscopic assembly morphology and microscopic chiral space of the porphyrin–amino acid derivatives, which is consistent with the results of molecular dynamics simulations and further impacts the rate and enantioselectivity of tyrosine chiral photocatalytic oxidation. These results provide new avenues for applying porphyrin derivatives in chiral photocatalysis.

Condensates formed by intrinsically disordered proteins mediate a myriad of cellular processes and are linked to pathological conditions including neurodegeneration. Rules of how different types of amino acids (e.g., π-π pairs) dictate the physical properties of biomolecular condensates are emerging, but our understanding of the roles of different amino acids is far from complete. Here we studied condensates formed by tetrapeptides of the form XXssXX, where X is an amino acid and ss represents a disulfide bond along the backbone. Eight peptides form four types of condensates at different concentrations and pH values: droplets (X = F, L, M, P, V, A); amorphous dense liquids (X = L, M, P, V, A); amorphous aggregates (X = W), and gels (X = I, V, A). The peptides exhibit enormous differences in phase equilibrium and material properties, including a 368-fold range in the threshold concentration for phase separation and a 3856-fold range in viscosity. All-atom molecular dynamics simulations provide physical explanations of these results. The present work also reveals widespread critical behaviors, including critical slowing down manifested by the formation of amorphous dense liquids and critical scaling obeyed by fusion speed, with broad implications for condensate function.

Enzymes are highly efficient bio-catalysts interesting for industries and medicine. Therefore, a goal of utmost importance in biochemical research is to understand how an enzyme catalyzes a chemical reaction. Here, the computational identification of functionally or structurally important residue positions can be of tremendous help. The datasets that are most informative for the algorithms are the 3D structure of a protein and a multiple sequence alignment (MSA) composed of homologous sequences. For example, an MSA allows for the quantification of residue conservation. Residue conservation at a given position indicates that only one type of amino acid fulfills all constraints imposed by protein structure or function. Furthermore, a detailed analysis of less strictly conserved residue positions may identify pairs, whose orchestration is mutually dependent and induces correlated mutations. Both of these conservation signals are indicative of functionally or structurally important positions. In the first part of this thesis, methods of machine learning were used to identify and classify these residue positions. It was the aim to predict in a mutually exclusively manner a role in catalysis, ligand-binding or protein stability for each residue position of a protein. Unfortunately, for many proteins the 3D structure is unknown. For other proteins, the number of known homologs is not sufficient to compile a meaningful MSA. Therefore, three variants of a classifier were designed and implemented, named CLIPS-1D, CLIPS-3D, and CLIPS-4D. These multi-class support vector machines allow for a classification based on an MSA (CLIPS-1D), a 3D structure (CLIPS-3D), and a combination of both (CLIPS-4D). CLIPS-1D exploits seven sequence-based features, whereas CLIPS-3D utilizes seven structure-based features. CLIPS-4D combines the seven sequence-based features of CLIPS-1D with those two structure-based features that increased its classification performance. A comparison with existing methods and a detailed analysis on a well-studied enzyme confirmed state-of-the-art prediction quality for CLIPS-1D and CLIPS-4D. In the second part of this thesis an algorithm for the identification of correlated mutations was improved. A common method for the identification of correlated mutations is to deduce the mutual information (MI) of a pair of residue positions from an MSA. The classical MI is based on Shannon’s information theory that utilizes probabilities only. Consequently, these approaches do not consider the similarity of residue pairs, which is a severe limitation. In order to improve these algorithms, H2rs was developed for this thesis. Thus, the MIvalues originate from the von Neumann entropy (vNE), which takes into account amino acid similarities modeled by means of a substitution matrix. To further improve the specificity of H2rs, the significance of MIvNE-values was assessed with a bootstrapping approach. The analysis of a large in silico testbed and the detailed assessment of five well-studied enzymes demonstrated state-of-the-art performance.

Deep neural networks have improved the accuracy of many difficult prediction tasks in biology, but it remains challenging to interpret these networks and learn molecular mechanisms. Here, we address the interpretability challenges associated with predicting transcriptional activation domains from protein sequence. Activation domains, regions within transcription factors that drive gene expression, were traditionally difficult to predict due to their sequence diversity and poor conservation. Multiple deep neural networks can now accurately predict activation domains, but these predictors are difficult to interpret. With the goal of interpretability, we designed simple neural networks that incorporated biophysical models of activation domains. The simplicity of these neural networks allowed us to visualize their parameters and directly interpret what the networks learned. The biophysical neural networks revealed two new ways that arrangement (i.e. the sequence grammar) of activation domain controlled function: 1) hydrophobic residues both increase activation domain strength and decrease protein abundance, and 2) acidic residues control both activation domain strength and protein abundance. Notably, the biophysical neural networks helped us to recognize the same signatures in complex interpreters of the deeper neural networks. We demonstrate how combining biophysical and deep neural networks maximizes both prediction accuracy and interpretability to yield insights into biological mechanisms.

A Potential New Therapeutic Approach for Friedreich Ataxia: Induction of Frataxin Expression With TALE Proteins

Modern computer networks' increasing complexity and scale need serious attention towards network anomaly detection. Software‐defined networking (SDN) and graph neural networks (GNN) have emerged as promising approaches for anomaly detection due to their ability to capture dynamic network behavior and learn complex patterns from large‐scale network data. The amalgamation of SDN and GNN for network anomaly detection presents promising opportunities for improving the accuracy and efficiency of network anomaly detection. This paper focuses on various trends, issues, and challenges by integrating GNN on the top of SDN for network anomaly detection. The article highlights the advantages of using SDN for providing fine‐grained control and programmability in network monitoring. At the same time, GNN can model network behavior as a graph and learn representations from graph‐structured data. The authors also discuss the limitations of traditional anomaly detection methods in SDN, such as rule‐based approaches, and the potential of GNN to overcome these limitations by leveraging their ability to capture non‐linear and dynamic patterns in network data. This paper also presents a case study of DoS attack detection using SDN. The result shows that SDN based approach helps to detect attacks with an accuracy of 97% with future research directions.

A High-Throughput Mutational Scan of an Intrinsically Disordered Acidic Transcriptional Activation Domain.

Nucleosome Transactions on the Promoters of the YeastGAL and PHO Genes

Eukaryotic transcriptional regulatory proteins typically comprise both a DNA binding domain and a regulatory domain. Although the structures of many DNA binding domains have been elucidated, no detailed structures are yet available for transcriptional activation domains. The activation domain of the herpesvirus protein VP16 has been an important model in mutational and biochemical studies. Here, we characterize the VP16 activation domain using time-resolved and steady-state fluorescence. Unique intrinsic fluorescent probes were obtained by replacing phenylalanine residues with tryptophan at position 442 or 473 of the activation domain of VP16 (residues 413-490, or subdomains thereof), linked to the DNA binding domain of the yeast protein GAL4. Emission spectra and quenching properties of Trp at either position were characteristic of fully exposed Trp. Time-resolved anisotropy decay measurements suggested that both Trp residues were associated with substantial segmental motion. The Trp residues at either position showed nearly identical fluorescence properties in either the full-length activation domain or relevant subdomains, suggesting that the two subdomains are similarly unstructured and have little effect on each other. As this domain may directly interact with several target proteins, it is likely that a significant structural transition accompanies these interactions.

This paper proposes an innovative solution based on deep learning for the anomaly detection challenge of big data in the industrial Internet of Things environment. Through in-depth analysis of time series and image data characteristics, LSTM and GRU networks are introduced to process time series data, and CNN models (such as ResNet and InceptionNet) are introduced to deal with image analysis and effectively capture complex patterns in data. Model-building involves not only architectural design, but also optimization strategies such as Adam Optimizer and loss function selection. In the preprocessing stage, the data are cleaned carefully through standardization, normalization, de-trending and de-noising to improve the learning efficiency of the model. The dataset used in the study is from a real - world intelligent manufacturing plant, with 1,000,000 records over three years. It contains 12 - dimensional sensor data (such as temperature, vibration frequency, current intensity), and outliers account for approximately 5% of the total data. Experimental results demonstrate that the distributed autoencoder under the federated learning framework outperforms traditional methods. It achieves an accuracy of 0.97, a recall of 0.94, an F1 - score of 0.95, and an AUC value of 0.96.However, it has a higher training time of 45 minutes and high communication costs during the training process due to the exchange of more data. In the federated learning process, each participating node independently trains the model based on local data, and uses the FedAvg strategy to aggregate parameters on the central server. Homomorphic encryption technology is used to ensure data privacy and prevent the leakage of original data. Compared with baseline methods such as IQR and Isolation Forest, the accuracy of this method is improved by 8%, which is statistically significant after the t-test (p&lt;0.05). Although the training time is 45 minutes, which is higher than the traditional method, it has obvious advantages in complex industrial data processing and privacy protection, and achieves a trade-off between computational efficiency and privacy protection and detection performance. The 95% confidence interval of the precision rate of 0.97 is [0.962, 0.978], the 95% confidence interval of the recall rate of 0.94 is [0.931, 0.949], the 95% confidence interval of the F1 value of 0.95 is [0.943, 0.957], and the 95% confidence interval of the AUC value of 0.96 is [0.952, 0.968].

Graphs effectively model interactions in real-world applications such as social and trade networks, where Graph Neural Networks (GNNs) excel at tasks such as link prediction to enhance user experiences. Despite these benefits, users raise privacy concerns as user data can be exploited to improve GNN performance without consent. Accordingly, various graph unlearning methods have been developed. Prior work shows that comparing models before and after unlearning enables attackers to launch former membership inference attacks (FMIA) on unlearned data. However, the imprint of unlearned data left in the unlearned model itself remains underexplored, and existing membership inference methods mainly exploit overfitting, making them ineffective for identifying unlearned data. To address this, we conducted theoretical analysis and proposed an attack framework targeting unlearned GNNs by learning the distribution patterns of unlearned data to distinguish them from normal test data. Extensive experiments on four real-world datasets and GNN architectures confirm our framework's effectiveness and reveal significant vulnerabilities in current graph unlearning methods.