Structure optimization, which yields the relaxed structure (minimum‑energy state), is essential for reliable materials property calculations, yet traditional ab initio approaches such as density‑functional theory (DFT) are computationally intensive. Machine learning (ML) has emerged to alleviate this bottleneck but suffers from two major limitations: (i) existing models operate mainly on atoms, leaving lattice vectors implicit despite their critical role in structural optimization; and (ii) they often rely on multi-stage, non-end-to-end workflows that are prone to error accumulation. Here, we present E³Relax, an end-to-end equivariant graph neural network that maps an unrelaxed crystal directly to its relaxed structure. E³Relax promotes both atoms and lattice vectors to graph nodes endowed with dual scalar–vector features, enabling unified and symmetry‑preserving modeling of atomic displacements and lattice deformations. A layer‑wise supervision strategy forces every network depth to make a physically meaningful refinement, mimicking the incremental convergence of DFT while preserving a fully end‑to‑end pipeline. We evaluate E³Relax on four benchmark datasets and demonstrate that it achieves remarkable accuracy and efficiency. Through DFT validations, we show that the structures predicted by E³Relax are energetically favorable, making them suitable as high-quality initial configurations to accelerate DFT calculations.

Accurate segmentation of the fetal head (FH) and pubic symphysis (PS) in intrapartum ultrasound (IU) images is a crucial step for automatic angle of progression (AoP) measurement, which plays an essential role in predicting delivery outcomes and reducing maternal and fetal complications. Existing CNN–Transformer hybrid models often suffer from attention collapse due to limited medical data and tend to overlook boundary details that are critical yet highly degraded by artifacts and noise in IU images. To address these issues, we propose a CNN-stylized dual-path CNN–Transformer hybrid network tailored for IU segmentation. The encoder integrates a parallel CNN branch and a CNN-stylized Transformer branch to balance local feature extraction and global dependency modeling while mitigating attention collapse. A Transformer-to-CNN fusion module further enhances cross-branch information interaction. In the decoder, a Boundary Attention Residual Module captures subtle foreground-background transitions and progressively refines boundary features. In addition, an Adaptive Boundary Enhancement strategy is designed to emphasize challenging boundary regions during CutMix augmentation. Experiments on three datasets demonstrate that our method outperforms state-of-the-art approaches in both overall accuracy and boundary precision. The automatic AoP measurement analysis further validates its potential for clinical translation. Code is available at https://github.com/SakuraKong/Dual-Path-CNN-Stylized-Hybrid-Network.

Accurate computer-aided interpretation of fetal heart rate (FHR) recordings depends on detecting the baseline and transient accelerations (Acc) and decelerations (Dec) that deviate from it. Most deep learning models treat FHR as a simple 1-D time sequence, overlooking the spectral separation between the low-frequency baseline and high-frequency Acc/Dec patterns. Neglecting this clinically important time-frequency structure can result in missed detections of Acc and Dec events and increased susceptibility to noise. To overcome these limitations, we present WaveFHR-VNet-a U-Net-style, multi-scale, multi-basis wavelet-voting network that analyzes FHR signals in the joint time-frequency domain. WaveFHR-VNet embeds a discrete wavelet transform (DWT) in every encoder block. Each DWT splits the features into approximation (low-pass) coefficients, which preserve the low-frequency baseline trends, and detail (high-pass) coefficients, which preserve the high-frequency Acc/Dec edges. Cascading these decompositions through successive layers yields a hierarchical, multi-scale representation. The decoder uses inverse DWT for full-resolution reconstruction. Skip connections are equipped with an Interactive Coefficient Selection (ICS) module that learns attention masks to suppress Doppler noise and motion artefacts in the detail stream while amplifying diagnostically salient transients. To enhance spectral diversity, five complementary wavelet bases (db4, db6, sym4, sym5, bior3.5) operate in parallel; a simple voting layer fuses their outputs, eliminating manual basis tuning. Evaluated on four FHR datasets, WaveFHR-VNet achieved state-of-the-art performance, with improvements of up to 5.3995% Dice, 5.4758% IoU, and 4.6263% accuracy over the best baselines on LCU-DB, the most widely used public benchmark. It also demonstrates strong cross-dataset generalization, consistently outperforming all comparison models. These results suggest that WaveFHR-VNet can serve as a reliable tool for intrapartum monitoring.

Cardiotocography (CTG) is essential for monitoring fetal health. Current deep learning applications in this field face two challenges: 1) the scarcity of annotated CTG data; 2) the difficulty in simultaneously capturing the coupled dependencies within multivariate time-series CTG. To bridge this gap, we propose MCTG: a multimodal self-supervised contrastive learning framework. For positive samples, we integrate an augmentation method based on frequency shapelets. These short frequency-domain subsequences capture class-specific information, preserving the most discriminative components of CTG signals. Additionally, we construct a multimodal learning framework that integrates series-image modalities to obtain complementary information between different modalities. This approach compensates for the inherent limitations of single time series modality networks in terms of complex cross-variable dependencies. Specifically, we encode series into images through three-channel encoding (frequency, wavelet, and time) to capture global long-term features, cross-scale features, and temporal texture details. Concurrently, multi-scale convolution integrates each variable dimension of the time series into a single image, enhancing the model’s ability to capture dependency couplings between variables. Experimental results demonstrate that MCTG effectively learns crucial feature representations from CTG data and achieves state-of-the-art performance in fetal distress prediction. The codes are available at https://github.com/Ladyfish030/MCTG.

The right atrium (RA) is critical for cardiac hemodynamics but is often overlooked in clinical diagnostics. This study presents a benchmark framework for RA cavity segmentation from late gadolinium-enhanced magnetic resonance imaging (LGE-MRIs), leveraging a two-stage strategy and a novel 3D deep learning network, RASnet. The architecture addresses challenges in class imbalance and anatomical variability by incorporating multi-path input, multi-scale feature fusion modules, Vision Transformers, context interaction mechanisms, and deep supervision. Evaluated on datasets comprising 354 LGE-MRIs, RASnet achieves SOTA performance with a Dice score of 92.19% on a primary dataset and demonstrates robust generalizability on an independent dataset. The proposed framework establishes a benchmark for RA cavity segmentation, enabling accurate and efficient analysis for cardiac imaging applications. Open-source code (https://github.com/zjinw/RAS) and data (https://zenodo.org/records/15524472) are provided to facilitate further research and clinical adoption.

Predicting the pharmacological activity, toxicity, and pharmacokinetic properties of molecules is a central task in drug discovery. Existing machine learning methods are transferred from one resource rich molecular property to another data scarce property in the same scaffold dataset. However, existing models may produce fragile and highly uncertain predictions for new scaffold molecules. And these models were tested on different benchmarks, which seriously affected the quality of their evaluation results. In this article, we introduce Meta-MolNet, a collection of data benchmark and algorithms, which is a standard benchmark platform for measuring model generalization and uncertainty quantification capabilities. Meta-MolNet manages a wide range of molecular datasets with high ratio of molecules/scaffolds, which often leads to more difficult data shift and generalization problems. Furthermore, we propose a graph attention network based on cross-domain meta-learning, Meta-GAT, which uses bilevel optimization to learn meta-knowledge from the scaffold family molecular dataset in the source domain. Meta-GAT benefits from meta-knowledge that reduces the requirement of sample complexity to enable reliable predictions of new scaffold molecules in the target domain through internal iteration of a few examples. We evaluate existing methods as baselines for the community, and the Meta-MolNet benchmark demonstrates the effectiveness of measuring the proposed algorithm in domain generalization and uncertainty quantification. Extensive experiments demonstrate that the Meta-GAT model has state-of-the-art domain generalization performance and robustly estimates uncertainty under few examples constraints. By publishing AI-ready data, evaluation frameworks, and baseline results, we hope to see the Meta-MolNet suite become a comprehensive resource for the AI-assisted drug discovery community. Meta-MolNet is freely accessible at https://github.com/lol88/Meta-MolNet.

Inductive bias in machine learning (ML) is the set of assumptions describing how a model makes predictions. Different ML-based methods for protein-ligand binding affinity (PLA) prediction have different inductive biases, leading to different levels of generalization capability and interpretability. Intuitively, the inductive bias of an ML-based model for PLA prediction should fit in with biological mechanisms relevant for binding to achieve good predictions with meaningful reasons. To this end, we propose an interaction-based inductive bias to restrict neural networks to functions relevant for binding with two assumptions: 1) A protein-ligand complex can be naturally expressed as a heterogeneous graph with covalent and non-covalent interactions; 2) The predicted PLA is the sum of pairwise atom-atom affinities determined by non-covalent interactions. The interaction-based inductive bias is embodied by an explainable heterogeneous interaction graph neural network (EHIGN) for explicitly modeling pairwise atom-atom interactions to predict PLA from 3D structures. Extensive experiments demonstrate that EHIGN achieves better generalization capability than other state-of-the-art ML-based baselines in PLA prediction and structure-based virtual screening. More importantly, comprehensive analyses of distance-affinity, pose-affinity, and substructure-affinity relations suggest that the interaction-based inductive bias can guide the model to learn atomic interactions that are consistent with physical reality. As a case study to demonstrate practical usefulness, our method is tested for predicting the efficacy of Nirmatrelvir against SARS-CoV-2 variants. EHIGN successfully recognizes the changes in the efficacy of Nirmatrelvir for different SARS-CoV-2 variants with meaningful reasons.

Finding candidate molecules with favorable pharmacological activity, low toxicity, and proper pharmacokinetic properties is an important task in drug discovery. Deep neural networks have made impressive progress in accelerating and improving drug discovery. However, these techniques rely on a large amount of label data to form accurate predictions of molecular properties. At each stage of the drug discovery pipeline, usually, only a few biological data of candidate molecules and derivatives are available, indicating that the application of deep neural networks for low-data drug discovery is still a formidable challenge. Here, we propose a meta learning architecture with graph attention network, Meta-GAT, to predict molecular properties in low-data drug discovery. The GAT captures the local effects of atomic groups at the atom level through the triple attentional mechanism and implicitly captures the interactions between different atomic groups at the molecular level. GAT is used to perceive molecular chemical environment and connectivity, thereby effectively reducing sample complexity. Meta-GAT further develops a meta learning strategy based on bilevel optimization, which transfers meta knowledge from other attribute prediction tasks to low-data target tasks. In summary, our work demonstrates how meta learning can reduce the amount of data required to make meaningful predictions of molecules in low-data scenarios. Meta learning is likely to become the new learning paradigm in low-data drug discovery. The source code is publicly available at: https://github.com/lol88/Meta-GAT.

Chest computed tomography (CT) image data is necessary for early diagnosis, treatment, and prognosis of Coronavirus Disease 2019 (COVID-19). Artificial intelligence has been tried to help clinicians in improving the diagnostic accuracy and working efficiency of CT. Whereas, existing supervised approaches on CT image of COVID-19 pneumonia require voxel-based annotations for training, which take a lot of time and effort. This paper proposed a weakly-supervised method for COVID-19 lesion localization based on generative adversarial network (GAN) with image-level labels only. We first introduced a GAN-based framework to generate normal-looking CT slices from CT slices with COVID-19 lesions. We then developed a novel feature match strategy to improve the reality of generated images by guiding the generator to capture the complex texture of chest CT images. Finally, the localization map of lesions can be easily obtained by subtracting the output image from its corresponding input image. By adding a classifier branch to the GAN-based framework to classify localization maps, we can further develop a diagnosis system with improved classification accuracy. Three CT datasets from hospitals of Sao Paulo, Italian Society of Medical and Interventional Radiology, and China Medical University about COVID-19 were collected in this article for evaluation. Our weakly supervised learning method obtained AUC of 0.883, dice coefficient of 0.575, accuracy of 0.884, sensitivity of 0.647, specificity of 0.929, and F1-score of 0.640, which exceeded other widely used weakly supervised object localization methods by a significant margin. We also compared the proposed method with fully supervised learning methods in COVID-19 lesion segmentation task, the proposed weakly supervised method still leads to a competitive result with dice coefficient of 0.575. Furthermore, we also analyzed the association between illness severity and visual score, we found that the common severity cohort had the largest sample size as well as the highest visual score which suggests our method can help rapid diagnosis of COVID-19 patients, especially in massive common severity cohort. In conclusion, we proposed this novel method can serve as an accurate and efficient tool to alleviate the bottleneck of expert annotation cost and advance the progress of computer-aided COVID-19 diagnosis.