Web Information System Design for Fast Protein Post-Translational Modification Site Prediction

MusiteDeep is an online resource providing a deep-learning framework for protein post-translational modification (PTM) site prediction and visualization. The predictor only uses protein sequences as input and no complex features are needed, which results in a real-time prediction for a large number of proteins. It takes less than three minutes to predict for 1000 sequences per PTM type. The output is presented at the amino acid level for the user-selected PTM types. The framework has been benchmarked and has demonstrated competitive performance in PTM site predictions by other researchers. In this webserver, we updated the previous framework by utilizing more advanced ensemble techniques, and providing prediction and visualization for multiple PTMs simultaneously for users to analyze potential PTM cross-talks directly. Besides prediction, users can interactively review the predicted PTM sites in the context of known PTM annotations and protein 3D structures through homology-based search. In addition, the server maintains a local database providing pre-processed PTM annotations from Uniport/Swiss-Prot for users to download. This database will be updated every three months. The MusiteDeep server is available at https://www.musite.net. The stand-alone tools for locally using MusiteDeep are available at https://github.com/duolinwang/MusiteDeep_web.

In heuristic search and especially in optimal classical planning the computation of accurate heuristic values can take up the majority of runtime. In many cases, the heuristic computations for a search node and its successors are very similar, leading to significant duplication of effort. For example most landmarks of a node that are computed by the LM-cut algorithm are also landmarks for the node's successors. We propose to reuse these landmarks and incrementally compute new ones to speed up the LM-cut calculation. The speed advantage obtained by incremental computation is offset by higher memory usage. We investigate different search algorithms that reduce memory usage without sacrificing the faster computation, leading to a substantial increase in coverage for benchmark domains from the International Planning Competitions.

Linear search algorithms have been shown to be particularly effective for solving partial Maximum Satisfiability (MaxSAT) problem instances. These algorithms start by adding a new relaxation variable to each soft clause and solving the resulting formula with a SAT solver. Whenever a Model is found, a new constraint on the relaxation variables is added such that Models with a greater or equal value are excluded. However, if the problem instance has a large number of relaxation variables, then adding a new constraint over these variables can lead to the exploration of a much larger search space. This article proposes new algorithms to improve the performance of linear search algorithms for MaxSAT by using Models found by the SAT solver to partition the relaxation variables. These algorithms add a new constraint on a subset of relaxation variables, thus intensifying the search on that subspace. In addition, the proposed algorithms are further enhanced by using incremental approaches where learned clauses are kept between calls to the SAT solver. Experimental results show that Model-based algorithms can outperform a classic linear search algorithm in several problem instances. Moreover, incremental approaches further improve the performance of linear search and Model-based algorithms for MaxSAT. Overall, Model-based algorithms are competitive with state-of-the-art MaxSAT solvers, and can improve their performance when solving problem instances with a large number of soft clauses.

We present here the recent update of AutoMotif Server (AMS 2.0) that predicts post-translational modification sites in protein sequences. The support vector machine (SVM) algorithm was trained on data gathered in 2007 from various sets of proteins containing experimentally verified chemical modifications of proteins. Short sequence segments around a modification site were dissected from a parent protein, and represented in the training set as binary or profile vectors. The updated efficiency of the SVM classification for each type of modification and the predictive power of both representations were estimated using leave-one-out tests for model of general phosphorylation and for modifications catalyzed by several specific protein kinases. The accuracy of the method was improved in comparison to the previous version of the service (Plewczynski et al., "AutoMotif server: prediction of single residue post-translational modifications in proteins", Bioinformatics 21: 2525-7, 2005). The precision of the updated version reached over 90% for selected types of phosphorylation and was optimized in trade of lower recall value of the classification model. The AutoMotif Server version 2007 is freely available at http://ams2.bioinfo.pl/ . Additionally, the reference dataset for optimization of prediction of phosphorylation sites, collected from the UniProtKB was also provided and can be accessed at http://ams2.bioinfo.pl/data/ .

Packet classification categorizes incoming packets by searching packet field values in rule database. This being highly researched domain, we have many algorithms providing faster lookup. However, for large rule-set up to 1M and larger field bytes up to 48, we observe surge in memory usage and drop in search performance for those algorithms. We take Aggregated Bit Vector (ABV) as base algorithm and optimize it for better classification speed and memory usage. Our solution maintains dynamic ABV intersection result table as a cache table and uses it instead of computing same ABV vector intersection for subsequent searches. This helps in getting cache hits even when subsequent search key values are different and hence leads to higher search optimization. We propose another search optimization for computing ABV vector intersection using ABV min-max index. We show memory efficient way to handle rules with more range fields, which reduces memory usage and improves pre-processing time. We achieve around 10-50% search optimization with various cases of search key values, 17–92 % preprocessing optimization and 65-80% memory optimization for rules with higher range fields.

We can implement relocatable, bus-based communication structures on Xilinx FPGAs using RapidWright while delivering competitive frequency, single digit speedups in execution time, and orders of magnitude reduction in memory usage over Xilinx Vivado 2017.2. We develop RapidRoute, a custom router that exploits symmetry in placement and routing of bus endpoints, caching of reusable route segments, selective multi-threading of the router engine, and abutment-friendly tiling heuristics. The key idea is to reduce the amount of work necessary to generate these communication structures through the use of search heuristics, parallelism, and reuse. We are able to outperform Vivado router by as much as 8× for topologies ranging from 1D rings, torii, and meshes, while taking 1000× lower memory footprint, and delivering timing with 0.2ns of Vivado. RapidRoute opens the door to building a family of custom routing tools for constructing FPGA overlays for various application domains.

Neural networks are widely used as a model for classification in a large variety of tasks. Typically, a learnable transformation (i.e., the classifier) is placed at the end of such models returning a value for each class used for classification. This transformation plays an important role in determining how the generated features change during the learning process. In this work, we argue that this transformation not only can be fixed (i.e., set as nontrainable) with no loss of accuracy and with a reduction in memory usage, but it can also be used to learn stationary and maximally separated embeddings. We show that the stationarity of the embedding and its maximal separated representation can be theoretically justified by setting the weights of the fixed classifier to values taken from the coordinate vertices of the three regular polytopes available in [Formula: see text], namely, the d -Simplex, the d -Cube, and the d -Orthoplex. These regular polytopes have the maximal amount of symmetry that can be exploited to generate stationary features angularly centered around their corresponding fixed weights. Our approach improves and broadens the concept of a fixed classifier, recently proposed by Hoffer et al., to a larger class of fixed classifier models. Experimental results confirm the theoretical analysis, the generalization capability, the faster convergence, and the improved performance of the proposed method. Code will be publicly available.

Focusing on the huge memory requirements of finite element method, when solving complex electric large size problem, a memory reduction scheme based on schur complement (MRS-SC) for parallel finite element method sparse matrix equation solving is proposed in this paper. The scheme's main idea is to convert the global system of equations to a local system of equations for each MPI (message passing interface) process through schur complement. The accuracy of the method and the memory saving efficiency are verified by numerical examples. As the scale of the problem increases, this method can achieve about a 40% reduction in memory usage over the traditional approach.

The identification of tunnel pipe jacking obstacles is usually carried out in harsh environments, and timely and accurate identification can avoid unnecessary economic and labor losses. However, the currently commonly used obstacle recognition models are complex in structure and have long recognition feedback times. Therefore, this study proposes a tunnel pipe jacking obstacle recognition model based on an improved You Only Look Once version 5. The model is optimized through pruning and knowledge distillation techniques to enhance its lightweight characteristics and accuracy. The experiments were conducted using the COCO dataset and a custom dataset consisting of 2667 tunnel shield tunneling obstacle images. The optimized model achieved an 88.6% reduction in the number of parameters, an 84.2% reduction in floating-point operations, a 62.5% reduction in memory usage, and a 90.1% reduction in response time. In real-world testing, the model achieved an accuracy of 94.0% and a processing speed of 75 frames per second (FPS), outperforming traditional YOLOv5 and other lightweight YOLOv5 variants such as M-YOLOv5, S-YOLOv5, PL-YOLOv5, and C-YOLOv5. Using evaluation metrics such as mean Average Precision (mAP), the proposed model demonstrated high efficiency and effectiveness in real-time obstacle recognition for tunnel construction. The model provides a new technological approach for safety management in tunnel construction while improving computational efficiency and maintaining high recognition accuracy.

Real-time image classification plays a pivotal role in applications such as autonomous driving, medical imaging, and surveillance, where both high accuracy and low-latency are critical. This study presents a performance-centric optimization framework for real-time image classification using TensorFlow, aiming to balance inference speed and model accuracy. The proposed approach integrates model pruning, quantization, and GPU acceleration to optimize Convolutional Neural Networks (CNNs) for deployment in resource-constrained environments. Experimental results on the ImageNet dataset demonstrate that the optimized model achieves 92% accuracy while reducing inference time to 45 milliseconds per image, outperforming both baseline CNN and MobileNetV2 models. Additionally, a 30% reduction in memory usage is observed, highlighting the efficiency of the optimization techniques. These results underscore the feasibility of deploying deep learning models in real-time applications, offering a viable solution for sectors where rapid decision-making and high classification performance are essential.

Non-uniform sampling is essential in NMR spectroscopy to accelerate data acquisition without compromising spectral information. Among the various NUS reconstruction techniques, such as compressive sensing and maximum entropy, structured low-rank methods are particularly effective in preserving spectral fidelity by exploiting the inherent redundancies in NMR data. However, their application to high-dimensional NMR remains limited due to the high memory and computational cost associated with nuclear norm minimization or matrix factorization methods, both of which rely on large intermediate variables. In this work, we propose a memory-efficient structured low-rank reconstruction algorithm that avoids explicit construction of block Hankel matrices and reduces the dependence on large matrix multiplications through a reformulated Alternating Direction Method of Multipliers scheme. In contrast to existing matrix factorization approaches, our algorithm significantly lowers both computational and memory complexity, achieving four-dimensional (4D) NMR reconstructions with more than a 90% reduction in memory usage while maintaining accuracy. This advancement enables practical and scalable applications of structured low-rank reconstruction to high-dimensional NMR, greatly enhancing the utility of NUS in real-world scenarios. The code is available at https://github.com/EricLin1993/HERO.

Analysis of Vector Particle-In-Cell (VPIC) memory usage optimizations on cutting-edge computer architectures

Patch-based anomaly detection methods have recently achieved strong performance by modeling localized feature embeddings extracted from pretrained neural networks. However, existing approaches often rely on memory-intensive storage of patch features or complex modeling techniques, which can be inefficient and overlook global structural relationships within an image. In this paper, we propose a novel method that detects anomalies by modeling the latent similarity structure among image patches. Specifically, we estimate conditional distributions over patches using feature embeddings and aggregate this information into a compact structural descriptor. During inference, anomalies are identified by measuring the divergence of a test image's descriptor from normal training samples using Kullback-Leibler divergence. Our method is fully unsupervised, lightweight, and highly efficient, achieving reductions in memory usage and inference time compared to existing methods, while maintaining competitive detection accuracy. We evaluate the proposed approach on the MVTec AD dataset, demonstrating that latent similarity modeling provides an attractive speedaccuracy trade-off for real-world industrial and robotic anomaly detection systems.

Swarm robotics is an emerging field capable of accomplishing complex tasks through collective behaviour. However, it continues to face persistent challenges in secure communication, decentralized decision-making, and scalability. To operate effectively in resource-constrained environments, swarm networks require a decentralized mechanism that is secure, fast, and efficient. Although many studies have explored the use of blockchain technology for swarm robotics, existing blockchain consensus algorithms such as Proof of Work (PoW), Proof of Stake (PoS), and their variants remain unsuitable due to high computational complexity and risk of stake centralization. To address these challenges, we introduce the blockchain-based Rotational Leadership Role (RLR) consensus algorithm, a voting-based consensus re-engineered from the Raft approach, together with Decentralized Task Authorization and Validation (DeTAV), a token-based mechanism for context-aware task validation. This design ensures efficiency, security, and scalability in swarm robotics and drone systems. RLR is lightweight and well suited to operate within the limited computing resources of small robots or aerial drones. To validate its performance, a custom-built robotic simulator was developed as part of this research. Experiments conducted with up to 70 concurrent robots demonstrated that RLR consumed under 90 MB Random Access Memory (RAM) and 12% Central Processing Unit (CPU), whereas PoW required 460 MB RAM and 27% CPU with a minimum difficulty level of 21, reflecting an 80% reduction in memory usage and a 55% reduction in CPU consumption. Scalability tests with 4 to 70 robots further revealed RLR’s scalability with an average of 78% higher throughput, 47% lower election latency, and 34% lower consensus latency. Additionally, under the simulated attack scenarios and assuming uncompromised cryptographic keys, DeTAV’s context-based validation consistently achieved 100% success in detecting and isolating Byzantine nodes, while reducing Quality of Detection (QoD) time by 67%. Collectively, these results confirm that RLR with DeTAV effectively meets the efficiency, security, and scalability requirements of swarm robotic and drone networks.

Farmers started to adopt smart gadgets like automated pest detection and kiosks for their work in light of the AI revolution in agriculture. These kinds of technologies help farmers identify pest infestations early, which allows the farmers to take action accordingly and prevent crop loss. This results in improvement in both yield and sustainability. However, running an advanced model such as YOLOv7 on low-power edge devices remains a challenge, especially in real time. To address the problem, our research brings an innovative Advanced Edge-Based Vision System designed for detecting small insects and classification in agricultural environments. The model built utilizes the Jetson Nano platform, TensorRT optimization, and an enhanced YOLOv8 model. To optimize the performance, the model uses compression techniques, such as INT8 quantization, and is accelerated by TensorRT. The YOLOv8 model is trained on a specialised insect dataset, and this optimization technique results in a tenfold reduction in memory usage, yet performs efficiently and also improves the inference speed using Jetson Nano by achieving a processing rate of 45 Frames Per Second (FPS). Despite these optimizations and memory reduction, the model serves at its best where the F1-scores exceeding 95% ensure accurate detection of the pest. This study demonstrates the successful deployment of lightweight, high-performance vision models on edge devices, enabling real-time, accurate pest detection. The findings contribute to advancing precision agriculture by making intelligent pest management more accessible and efficient.