Predicting the physicochemical properties of molecules is a cornerstone of computational chemistry and drug discovery. While quantum mechanical methods provide high accuracy, their computational expense limits their use in high-throughput screening. To bridge this gap, we introduce a novel, disentangled deep learning architecture that adaptively predicts molecular properties by leveraging either 2D topological graphs or 3D geometric information. Our framework consists of two specialized models: a Message Passing Neural Network (MPNN) enhanced with cycle-based semimaster nodes for robust 2D feature extraction, and an equivariant network with a disentangled update mechanism for high-fidelity 3D-aware predictions. The architecture's built-in hierarchical attention mechanism provides interpretability by highlighting salient atomic and substructure features. This adaptive, high-performance, and interpretable framework offers a versatile solution for accelerating molecular discovery, irrespective of the data dimensionality.

Proxy applications play a fundamental role in high-performance computing (HPC) by providing simplified representations of production applications for performance evaluation, bottleneck analysis, and optimization. Developing high-fidelity proxy applications remains a labor-intensive process, while existing automated methods often failing to maintain fidelity across diverse hardware and software platforms. To address these challenges, we propose HFProxy, a novel framework for synthesizing portable and high-fidelity proxy applications for Message Passing Interface (MPI) programs. HFProxy traces both communication and computation events in MPI programs, models representative computation and memory access patterns, and employs a grammar-based approach to compress trace data. By combining these techniques with contention-aware performance modeling and workload shrinkage, HFProxy generates lightweight proxies that faithfully replicate the performance characteristics of the original programs across platforms. We evaluate HFProxy on a diverse set of MPI programs, demonstrating its ability to achieve high fidelity with time prediction errors under 8% across platforms. HFProxy also enables up to 10-fold workload shrinkage while maintaining critical performance metrics. Additionally, case studies show that HFProxy provides actionable insights for performance analysis and bottleneck diagnosis, proving its utility for both developers and HPC system administrators.

Current High-Performance Computing systems rely on massive parallelism to achieve exascale performance. They use task scheduling and message-passing programming models to explore complementary sources of parallelism. Combining the two holds the promise of allowing seamless exploitation of both intra- and inter-node concurrency while leveraging widely-known programming abstractions. Still, the interaction between the two raises coordination problems that could make work distribution excessively costly, limiting performance. This work is the first to evaluate comprehensive hardware acceleration of their combined use integrating them in a programming model that further exploits their synergies. The hardware/software co-design approach proposed for this purpose is prototyped on a cluster of 64 FPGA nodes, where each holds a RISC-V Rocket Chip CPU with 8 cores. On one hand, this article combines OMPIF and Picos, which are hardware accelerators for message passing and task scheduling respectively. They interface with the CPU through RoCC-based custom RISC-V instructions. On the other hand, we present the Implicit Message Passing (IMP) programming model, that extends task scheduling abstractions to leverage MPI-mediated inter-node parallelism without requiring explicit MPI calls. Thus, IMP transparently allows the dataflow-style execution induced by task scheduling to span multiple nodes. We implement three benchmarks, N-body, Heat, and Cholesky, each with two different strategies, IMP and explicit MPI, and evaluate them on the multi-core FPGA-based cluster. We demonstrate our hardware-software co-design approach achieves near-linear scalability with IMP and the OMPIF/Picos accelerators, and reduces task management overhead from 2200 to 300 cycles per task. Furthermore, when leveraging all 512 cores (split among the 64 nodes), we measure speedups of 2.04x (40x in communication), 1.25x (7x in communication), and 7.29x (25x in communication) compared with unaccelerated MPI for N-body, Heat, and Cholesky respectively. Finally, at 64 nodes, we respectively achieve 99%, 83%, and 79% of weak scaling efficiency.

Enhancing Low-Degree Graph Neural Networks via Joint Training and Improved Message Passing

ABSTRACT The era of generative artificial intelligence (AI) in drug design is here, and in this study, we applied a generative AI model to design potential non‐nucleoside inhibitors of HIV‐1 reverse transcriptase (RT). RT converts viral RNA into DNA, facilitating viral integration into a host's genome. Inhibiting RT is therefore critical in combating HIV‐1. Using the REINVENT prior model, we fine‐tuned it with classified RT inhibitors (pIC50 > 8.5) from ChEMBL, generating over 6000 compounds. These compounds underwent rigorous filtering, including classification by a Message Passing Neural Network (MPNN), PAINS filtering, pharmacophore modeling, structure‐based screening, MMGBSA scoring, and ADMET prediction. We identified five top‐performing compounds, which demonstrated superior docking scores (≤ −13.22 kcal/mol) and binding free energies (MMGBSA dG Bind ≤ −77.48 kcal/mol) compared to the reference ligand (−8.42 kcal/mol). Further ADMET predictions and molecular dynamics simulation at 500 ns revealed that these compounds had better drug‐like properties and comparable stability at the active site compared to the reference ligand. These findings suggest that the identified compounds are promising candidates for further in vitro validation to confirm their therapeutic potential against HIV‐related proteins. This study highlights the transformative role of AI‐driven drug discovery in addressing HIV drug resistance.

AI-Driven molecule generation and bioactivity prediction: A multi-model approach combining VAE, graph and language-based neural networks.

Millimeter-wave (mmWave) massive MIMO systems use many antennas. These systems offer high data rates. But using many radio frequency (RF) chains increases cost and power use. To solve this, lens antenna arrays are used. Energy is focused, allowing the use of fewer RF chains. However, this creates a new challenge. With fewer RF chains, it is hard to estimate the wireless channel. Accurate channel estimation is needed for good system performance. In beamspace, the channel is sparse. This shows that only a few values are large. The rest are close to zero. Because of this, the problem is seen as sparse signal recovery. AMP (Approximate Message Passing) is one popular algorithm used for this. A better version named LAMP (Learned AMP) uses deep learning. But it still does not give the best results. This paper proposes a new method GM-LAMP. It improves the channel estimation accuracy. It uses prior knowledge about the channel. It assumes that the beamspace channel follows a Gaussian mixture distribution. First, a new shrinkage function is created based on this distribution. Then, the original function in the LAMP network is replaced with the new one. As a result, a better deep learning model is developed. The final GM-LAMP network estimates the beamspace channel more precisely. It works well with both theoretical models and real-world data. Simulations show that GM-LAMP performs better than earlier methods. This approach combines math knowledge and deep learning. It shows that using prior information helps deep networks make smarter predictions. The proposed method offers better accuracy and is useful for future mmWave systems.

ABSTRACT Parallel simulation has been widely utilized in large-scale microscopic traffic simulation by dividing the road network into partitions. However, in congested scenarios, both the efficiency and accuracy of parallel simulation are highly challenging due to the loss of traffic information on the boundary between subnetworks. This paper proposes a tailored parallel simulation method for congestion scenarios. A network partitioning approach using the Spectral Partitioning (SP) algorithm and congestion information is developed with the aim of achieving load balancing across simulation processes. A new boundary transition strategy is introduced to assess road congestion states. The Message Passing Interface (MPI) is applied to relay congested vehicle information between processes. The results demonstrate that the proposed parallel simulation method is capable of achieving load balancing, where each process bears a similar simulation load. Meanwhile, the accuracy of the proposed parallel simulation method is increased from 78% to 91%.

Graph neural networks (GNNs) achieve deep representation learning of graph-structured data through a message passing mechanism, becoming a core technical paradigm for processing structured data. This paper systematically analyzes the architectural principles of the message passing mechanism, illustrating its three-step core process of "message generation-aggregation-update" and its mathematical implementation. The paper focuses on comparing the mechanism with typical models such as GCN and GAT. Drawing on empirical examples from fields such as bioinformatics, financial risk control, and retail supply chain, the paper reveals the unique advantages of this mechanism in capturing entity associations and discovering hidden patterns. In protein function prediction, the message passing-based DeepFRI model achieves accurate mapping between structure and function; in financial fraud detection, aggregating multi-hop transaction information significantly improves the ability to identify hidden criminal networks. The paper also analyzes current challenges such as oversmoothing and large-scale graph adaptation, proposing solutions such as residual connection optimization and neighbor sampling. This research demonstrates that the message passing mechanism provides a modeling approach that captures the relational nature of structured data, demonstrating strong potential for application in complex, cross-domain problems. Future applications will be further overcome by combining self-supervised learning with large-scale model technologies.

Context . Modeling the solar atmosphere is challenging due to its layered structure and dynamic multi-scale processes. Aims . We aim to validate the new radiative magnetohydrodynamic (MHD) code MAGEC—built by integrating the M ANCHA 3D and MAGNUS codes into a finite-volume, shock-capturing framework—and to explore its capabilities through 2D simulations of magnetoconvection in the solar atmosphere. Methods . The MAGEC code is parallelized with Message Passing Interface (MPI), enabling efficient scalability for large-scale simulations. We have enhanced it with advanced numerical techniques to address the specific complexities of the solar corona, including a module for local thermodynamic equilibrium (LTE) radiative coronal losses. To address the small time steps due to large heat flux values, we adopted the hyperbolic treatment for the thermal conduction of M ANCHA 3D, which significantly improves the computational times. In addition, we estimated the effective numerical resistivity and viscosity through a dedicated set of experiments. To evaluate the robustness and accuracy of MAGEC, we performed a series of 2D simulations covering a domain extending from 2 Mm below the solar surface to 18.16 Mm into the corona. Simulations were conducted with both open and closed magnetic field configurations. For each case, we analyzed the resulting steady-state temperature profiles and examined the energy contributions at different heights. In addition, we investigated the influence of the perpendicular component of thermal conduction in a dedicated simulation. Results . The MAGEC code effectively reproduced expected temperature profiles based on the boundary conditions applied and the imposed magnetic field configuration. All simulations reached a thermally stable state. When using an open vertical magnetic field, the temperature in the middle corona was higher than in the case with a closed, arcade-like magnetic field structure. We quantified the contributions to the internal energy from all explicit and implicit terms in the steady state, both in terms of temporal averages and as functions of height, as well as their relative contributions to total heating and cooling. In a second phase of the study, we investigated the role of the perpendicular component of thermal conduction, which is often neglected in coronal models, and found that it can influence plasma dynamics around reconnection events. Although local effects are modest, their cumulative impact can lead to measurable changes in the average temperature profile. Conclusions . Through detailed validation, MAGEC is a reliable and efficient code for radiative MHD simulations of the solar atmosphere. The integration of shock-capturing methods is particularly well suited to modeling the plasma environment, effectively handling the shocks and discontinuities characteristic of the solar atmosphere. MAGEC is a robust tool for high-fidelity magneto-convection simulations of the solar atmospheric dynamics.

Graph Neural Networks (GNNs) exhibit satisfactory performance on homophilic networks, where most edges connect two nodes with the same label. However, their effectiveness diminishes as the graphs become heterophilic (low homophily), prompting the exploration of various message-passing schemes. In particular, assigning negative weights to heterophilic edges (signed propagation) for message-passing has gained significant attention, and some studies theoretically confirm its effectiveness. Nevertheless, prior theorems assume binary classification scenarios, which may not hold well for graphs with multiple classes. To solve this limitation, we offer new theoretical insights into GNNs in multi-class environments and identify the drawbacks of employing signed propagation from two perspectives: message-passing and parameter update. We found that signed propagation without considering feature distribution can degrade the separability of dissimilar neighbors, which also increases prediction uncertainty (e.g., conflicting evidence) that can cause instability. To address these limitations, we introduce two novel calibration strategies aiming to improve discrimination power while reducing entropy in predictions. Through theoretical and extensive experimental analysis, we demonstrate that the proposed schemes enhance the performance of both signed and general message-passing neural networks (Choi et al. 2023).

Hepatocyte Growth Factor Receptor (HGFR) overexpression plays a critical role in ovarian cancer progression by promoting cell proliferation, survival, and metastasis. Despite the therapeutic potential of existing HGFR inhibitors, such as crizotinib, concerns regarding low potency and high toxicity require safer alternatives. This study establishes an integrative in silico framework combining deep learning-based bioactivity prediction, structure-based drug repurposing, and toxicity profiling via Directed Message Passing Neural Network (D-MPNN). A rigorously filtered dataset of HGFR-targeting bioactives was used to train an artificial neural network (ANN), which was subsequently applied to evaluate the bioactivity of 1,040 FDA-approved drugs. Highly potent candidates underwent molecular docking, identifying venetoclax (S-score: -8.78, RMSD: 1.32), LSM-5313 (S-score: -8.50, RMSD: 1.89), and cefoperazone (S-score: -8.24, RMSD: 1.82) as the lead compounds. Micro-scale molecular dynamics simulations (2µs) and post-trajectory analyses including RMSD, RMSF, Rg, hydrogen bonding, PCA, FEL, DCCM, and MMGBSA confirmed their stable and favorable binding at the HGFR active site. Finally, the D-MPNN-driven toxicity assessment revealed no significant toxic liabilities in the proposed compounds. Overall, this multi-tiered computational approach offers reliable, mechanistically supported candidates for HGFR inhibition. The identified FDA-approved drugs represent promising, non-toxic therapeutic options for ovarian cancer, encouraging further preclinical and clinical investigation.

Massive machine-type communications (mMTC) in future 6G networks will involve a vast number of devices with sporadic traffic. Grant-free access has emerged as an effective strategy to reduce the access latency and processing overhead by allowing devices to transmit without prior permission, making accurate active user detection and channel estimation (AUDCE) crucial. In this paper, we investigate the joint AUDCE problem in wideband massive access systems. We develop an innovative channel prior model that captures the dual correlation structure of the channel using three state variables: active indication, channel supports, and channel values. By integrating Markov chains with coupled Gaussian distributions, the model effectively describes both the structural and numerical dependencies within the channel. We propose the correlated hybrid message passing (CHMP) algorithm based on Bethe free energy (BFE) minimization, which adaptively updates model parameters without requiring prior knowledge of user sparsity or channel priors. Simulation results show that the CHMP algorithm accurately detects active users and achieves precise channel estimation.

Abstract Coronal mass ejections (CMEs), as crucial drivers of space weather, necessitate a comprehensive understanding of their initiation and evolution in the solar corona in order to better predict their propagation. Solar Cycle 24 exhibited lower sunspot numbers compared to Solar Cycle 23, along with a decrease in the heliospheric magnetic pressure. Consequently, a higher frequency of weak CMEs was observed during Solar Cycle 24. Forecasting CMEs is vital, and various methods, primarily involving the study of the global magnetic parameters using data sets like Space-weather Helioseismic and Magnetic Imager Active Region Patches, have been employed in earlier works. In this study, we perform numerical simulations of CMEs within a magnetohydrodynamics framework using Message Passing Interface–Adaptive Mesh Refinement Versatile Advection Code in 2.5D. By employing the breakout model for CME initiation, we introduce a multipolar magnetic field configuration within a background bipolar magnetic field, inducing shear to trigger the CME eruption. Our investigation focuses on understanding the impact of the background global magnetic field on CME eruptions. Furthermore, we analyze the evolution of various global magnetic parameters in distinct scenarios (failed eruption, single eruption, and multiple eruptions) resulting from varying amounts of helicity injection in the form of shear at the base of the magnetic arcade system. Our findings reveal that an increase in the strength of the background poloidal magnetic field constrains CME eruptions. Furthermore, we establish that the growth rate of absolute net current helicity is the crucial factor that determines the likelihood of CME eruptions.

MPNN-CWExplainer: An enhanced deep learning framework for HIV drug bioactivity prediction with class-weighted loss and explainability.

Unified Cooperative Localization via Augmented Factor Graph and Error State Message Passing

We present a matrix-free multigrid method for high-order Discontinuous Galerkin (DG) finite element methods with GPU acceleration. A performance analysis is conducted, comparing various data and compute layouts. Smoother implementations are optimized through localization and fast diagonalization techniques. Leveraging conflict-free access patterns in shared memory, arithmetic throughput of up to 40% of the peak performance on NVIDIA A100 GPUs are achieved. Experimental results affirm the effectiveness of mixed-precision approaches and Message Passing Interface (MPI) parallelization in accelerating algorithms. Furthermore, an assessment of solver efficiency and robustness is provided across both two and three dimensions, with applications to Poisson problems.

Satellite Internet of Things (IoT) networks based on satellites are becoming increasingly critical for mission-critical applications, including disaster recovery, environmental surveillance, and remote sensing. While becoming more widespread, they are also more vulnerable to various risks, particularly due to the heterogeneous communication technologies they support and the limited computing capacity on each device. When such IoT systems are connected with central HighPerformance Computing (HPC) clouds, particularly by satellite links, new security issues arise, the primary one being the secure transmission of confidential information. To overcome such challenges, this research proposes a new security framework termed DLGAN (Deep Learning-based Generative Adversarial Network), specially designed for satellite-based IoT scenarios. The model leverages the strengths of Convolutional Neural Networks (CNNs) for real-time anomaly detection, combined with Generative Adversarial Networks (GANs) to generate realistic synthetic attack data, thereby addressing the challenge of skewed datasets prevalent in cybersecurity research. Since training GANs may be computationally expensive, the model is optimized to run on an HPC system via the Message Passing Interface (MPI) to enable scalable parallel processing of huge IoT data. Fundamentally, the DLGAN model is based on a generator/discriminator mechanism for effectively distinguishing network traffic as either benign or malicious, with the capability to detect 14 different types of attacks. By harnessingAI-enabled GPUs in the HPC cloud, the system can provide fast and accurate detection while maintaining low computational costs. Experimental evaluations demonstrate that the framework significantly enhances detection accuracy, reduces training time, and scales well with large data volumes, making it highly suitable for real-time security operations. In total, this study highlights how integrating advanced deep learning technologies with HPC-based distributed environments can deliver an efficient and dynamic defense mechanism for contemporary IoT networks. The envisaged solution is unique in its ability to scale, maximize efficiency, and resist attacks while securing satellite-based IoT infrastructures.

Nephrotoxicity remains a critical safety concern in drug development and clinical practice. Despite their significance, existing computational models for nephrotoxicity prediction face challenges related to limited precision and reliability. To address these challenges, this study constructed the largest publicly available database to date, comprising 1831 high-quality nephrotoxicity-related compounds. Using this dataset, we developed classification models employing both traditional machine learning algorithms and graph-based deep learning methods. Our results demonstrate that the Directed Message Passing Neural Network model, combined with molecular graphs and ChemoPy2D descriptors, outperformed other models, achieving a mean Kappa value of 70.3%. To improve the reliability of the model, we implemented an uncertainty quantification method to define the model's applicability domain and quantify the confidence of prediction. On the representative model, this approach significantly enhanced predictive performance within the applicability domain, yielding a Kappa metric of 90.4%. Notably, our model also achieved the highest performance on an external test set compared to existing models. Finally, we performed multiscale feature analysis to provide actionable insights for safer drug design. This analysis integrated dataset -centric methods to identify structural alerts and beneficial transformations, alongside model-centric techniques to reveal the key global and local features driving nephrotoxicity. The established prediction models, combined with uncertainty quantification and structural feature insights derived in this study, offer valuable tools for the development of safer and more effective drugs.

The confined space of coal mines characterized by curved tunnels with rough surfaces and a variety of deployed production equipment induces severe signal attenuation and interruption, which significantly degrades the accuracy of conventional channel estimation algorithms applied in coal mine wireless communication systems. To address these challenges, we propose a modified Bilinear Generalized Approximate Message Passing (mBiGAMP) algorithm enhanced by intelligent reflecting surface (IRS) technology to improve channel estimation accuracy in coal mine scenarios. Due to the presence of abundant coal-carrying belt conveyors, we establish a hybrid channel model integrating both fast-varying and quasi-static components to accurately model the unique propagation environment in coal mines. Specifically, the fast-varying channel captures the varying signal paths affected by moving conveyors, while the quasi-static channel represents stable direct links. Since this hybrid structure necessitates an augmented factor graph, we introduce two additional factor nodes and variable nodes to characterize the distinct message-passing behaviors and then rigorously derive the mBiGAMP algorithm. Simulation results demonstrate that the proposed mBiGAMP algorithm achieves superior channel estimation accuracy in dynamic conveyor-affected coal mine scenarios compared with other state-of-the-art methods, showing significant improvements in both separated and cascaded channel estimation. Specifically, when the NMSE is , the SNR of mBiGAMP is improved by approximately 5 dB, 6 dB, and 14 dB compared with the Dual-Structure Orthogonal Matching Pursuit (DS-OMP), Parallel Factor (PARAFAC), and Least Squares (LS) algorithms, respectively. We also verify the convergence behavior of the proposed mBiGAMP algorithm across the operational signal-to-noise ratios range. Furthermore, we investigate the impact of the number of pilots on the channel estimation performance, which reveals that the proposed mBiGAMP algorithm consumes fewer number of pilots to accurately recover channel state information than other methods while preserving estimation fidelity.