Diffusion and flow matching models have recently emerged as promising approaches for peptide binder design. Despite their progress, these models still face two major challenges. First, categorical sampling of discrete residue types collapses their continuous parameters into one-hot assignments, while continuous variables (e.g., atom positions) evolve smoothly throughout the generation process. This mismatch disrupts the update dynamics and results in suboptimal performance. Second, current models assume unimodal distributions for side-chain torsion angles, which conflicts with the inherently multimodal nature of side-chain rotameric states and limits prediction accuracy. To address these limitations, we introduce PepBFN, the first Bayesian flow network for full-atom peptide design that directly models parameter distributions in fully continuous space. Specifically, PepBFN models discrete residue types by learning their continuous parameter distributions, enabling joint and smooth Bayesian updates with other continuous structural parameters. It further employs a novel Gaussian mixture-based Bayesian flow to capture the multimodal side-chain rotameric states and a Matrix Fisher-based Riemannian flow to directly model residue orientations on the SO(3) manifold. Together, these parameter distributions are progressively refined via Bayesian updates, yielding smooth and coherent peptide generation. Experiments on side-chain packing, reverse folding, and binder design tasks demonstrate the strong potential of PepBFN in computational peptide design.

Electroencephalography (EEG) plays a vital role in clinical and cognitive applications such as epilepsy diagnosis and emotion recognition. However, the low signal-to-noise ratio, inter-subject variability, and inherent non-stationarity of EEG signals present substantial modeling challenges. While recent Transformer-based models offer promising long-range modeling capabilities, their self-attention mechanism behaves as a low-pass filter, suppressing high-frequency neural patterns critical for decoding transient events. In this work, we provide the first formal analysis demonstrating this low-pass behavior in self-attention mechanisms when applied to EEG signals, revealing a fundamental limitation of deep attention-based EEG models. To address this, we propose SEBSFormer, a spectral-enhanced bi-Stream Transformer that jointly models temporal dependencies and spectral structures. SEBSFormer integrates three key modules: a spectral compensation module that restores high-frequency components via residual correction in the Fourier domain; a multi-scale temporal attention module for saliency-guided temporal compression; and a graph-guided dynamic fusion module for adaptive spatial aggregation across electrodes. Extensive experiments on three benchmark datasets—TUAB, TUEV, and SEED—demonstrate that SEBSFormer consistently outperforms existing state-of-the-art models across both clinical and affective tasks. Our findings establish a new paradigm for frequency-aware EEG modeling.

PIWI proteins maintain genome integrity by piRNA-guided cleavage of complementary RNA targets. While Cleave-N'-Seq (CNS-seq) has advanced our understanding of PIWI targeting logic through quantitative mapping of cleavage rates and pairing rules, its labor-intensive workflows hinder systematic exploration of sequence determinants. Here, we present PAIRNet, a deep learning framework that predicts PIWI-mediated RNA cleavage rates by explicitly modeling guide-target interactions. Recognizing that interaction geometry, not just sequence, dictates cleavage efficiency, PAIRNet integrates biochemical insights with computational innovation: it encodes pairing states, mismatch types, insertions, and deletions alongside learnable positional embeddings to quantify spatial dependencies; employs a hybrid CNN-Transformer architecture prioritizing duplex dynamics over static sequence features to resolve both local catalytic motifs (e.g., contiguous base-pairing at g10-g11) and distal structural perturbations; and incorporates interpretability modules (saliency maps, counterfactual analysis) to link interaction patterns to biochemical insights and uncover position-specific cleavage rules. Validated across four PIWI-guide datasets, PAIRNet consistently ranks among the top two performers in all experimental conditions, achieving the most pronounced relative improvements in PCC, 34.7% for MILI and 14.6% for MIWI, over second-ranking methods. Critically, PAIRNet recapitulates key biological principles-stringent complementarity at catalytic residues (g10-g11) and tolerance for 3' mismatches-aligning with structural studies of PIWI dynamics. By bridging biochemical precision with computational scalability, PAIRNet establishes a roadmap for designing high-specificity piRNA silencing tools while accelerating mechanistic studies of RNA-guided genome defense.

Controllable image structure editing has attracted increasing attention. While recent interactive point-based methods are convenient and realistic, they often lack fine-grained control over localized content. Partial sketches provide a simple yet expressive interface for local structure manipulation. However, existing partial-sketch-based manipulation methods relying on generative adversarial networks (GANs) suffer from limited generalization and fidelity. Moreover, although diffusion-based adapters excel at global conditioning (e.g., edge maps), localized editing with partial strokes remains challenging due to two key issues: effectively injecting sparse stroke conditions during denoising and preserving non-edited regions to avoid unintended changes. To address these challenges, we propose DiffStroke, a mask-free framework for localized image manipulation with partial sketches. We introduce trainable Image-Stroke Fusion (ISF) blocks to fuse source images and strokes at the feature level, enabling precise local shape control while maintaining appearance consistency. We further develop a self-supervised mask estimator to protect irrelevant regions without manual input. Specifically, we leverage Tweedie's formula to estimate a clean latent image from noisy latents, blend the denoised result with the source, and train the mask estimator by minimizing the error between the blended latent and the target latent. Experiments on natural and facial images demonstrate that DiffStroke outperforms state-of-the-art methods on both simple and complex stroke-based editing tasks. DiffStroke can also be combined with text prompts to produce diverse and creative results. Code is available at https://github.com/CMACH508/DiffStroke.

Deep generative models provide a promising approach to de novo 3D peptide design. Most of them jointly model the distributions of peptide's position, orientation, and conformation, attempting to simultaneously converge to the target pocket. However, in the early stage of docking, optimizing conformation-only modalities such as rotation and torsion can be physically meaningless, as the peptide is initialized far from the protein pocket and no interaction field is present. We define this problem as the multimodal temporal inconsistency problem and claim it is a key factor contributing to low binding affinity in generated peptides. To address this challenge, we propose THFlow, a novel flow matching-based multimodal generative model that explicitly models the temporal hierarchy between peptide position and conformation. It employs a polynomial-based conditional flow to accelerate positional convergence early on, and later aligns it with rotation and torsion for coordinated conformation refinement under the emerging interaction field. Additionally, we incorporate interaction-related features, such as polarity, to further enhance the model's understanding of peptide-protein binding. Extensive experiments demonstrate that THFlow outperforms existing methods in generating peptides with superior stability, affinity, and diversity, offering an effective and accurate solution for advancing peptide-based therapeutic development.

PIWI-interacting RNAs (piRNAs) are critical genome guardians that guide PIWI proteins to cleave trans-posable element transcripts. Predicting these cleavage events is vital for understanding gene regulation but is challenged by the complex, “relaxed” rules of piRNA targeting and the noisy cellular context of in vivo data. Previous models, trained solely on in vivo data, struggle with high noise, non-functional binding events, and binary outputs that fail to capture the quantitative biophysics of piRNA-target interactions. Recognizing that precise, quantitative in vitro data directly quantifies biophysical parameters of cleavage, we aimed to leverage this information to overcome the noise and complexity inherent in in vivo observations. To achieve this, we developed piR-DANN, a deep learning framework based on a novel, biologically-informed adversarial domain adaptation strategy. Our core innovation moves beyond conventional approaches by providing the domain classifier with structural and sequence determinants of targeting summarizing known biological rules, in addition to the deep features learned by the model. This asymmetric design compels the feature extractor to learn the fundamental, domain-invariant principles of piRNA targeting. piR-DANN outperforms existing benchmarks predictive performance, with AUROC scores of 94.7% and 99.7% on two independent test sets. Furthermore, counterfactual analysis reveals a positional importance map concordant with PIWI catalytic core constraints, validating its extraction of biological signal from noise. By integrating heterogeneous data into an accurate, interpretable framework, our work explores piRNA biology and proposes a generalizable approach for deciphering gene regulatory systems.

Accurately predicting drug synergy in cancer therapy remains challenging due to the strong dependence of drug effectiveness on cell-line context. Most existing models overlook this variability and fail to fully capture drug-cell line interactions. We present CATSyn, a Context-Aware heTerogeneous graph model for synergistic drug combination prediction. CATSyn introduces a context-aware attention mechanism that dynamically adjusts network weights based on cell-line environments, capturing cellspecific drug effects while maintaining generalization. To model these effects, we construct heterogeneous graphs that integrate drug and cell-line features into composite nodes, supported by universal nodes to share global information. Experiments on benchmark datasets demonstrate that CATSyn achieves state-of-the-art performance in both standard and unseen cell-line settings, highlighting its ability to balance specificity and generalization in synergy prediction.

Magnetic resonance imaging (MRI) provides highquality soft tissue contrast images and is crucial in medical diagnosis. However, systems face trade-offs between image resolution and scan time. Low-resolution MRI scans reduce scan time and patient burden but lose critical details needed for accurate diagnosis. To address this problem, super-resolution techniques have been developed to improve the clarity of lowresolution input images. Single-image super-resolution (SISR), which minimizes patient scanning time, has gradually become a research focus, but existing methods often struggle to balance the reconstruction of low-frequency structural information and high-frequency details. In this paper, we propose a novel superresolution up-sampling pipeline that enhances both the highfrequency and low-frequency components of magnetic resonance imaging. In addition, we introduce an enhanced loss function that includes symmetry and edge constraints to preserve critical structural details for improved diagnostic accuracy. The extensive experiments across multiple datasets validate the effectiveness of our SISR model. Source code will be made publicly available.

General intelligence requires quick adaptation across tasks. While existing reinforcement learning (RL) methods have made progress in generalization, they typically assume only distribution changes between source and target domains. In this paper, we explore a wider range of scenarios where not only the distribution but also the environment spaces may change. For example, in the CoinRun environment, we train agents from easy levels and generalize them to difficulty levels where there could be new enemies that have never occurred before. To address this challenging setting, we introduce a causality-guided self-adaptive representation-based approach, called CSR, that equips the agent to generalize effectively across tasks with evolving dynamics. Specifically, we employ causal representation learning to characterize the latent causal variables within the RL system. Such compact causal representations uncover the structural relationships among variables, enabling the agent to autonomously determine whether changes in the environment stem from distribution shifts or variations in space, and to precisely locate these changes. We then devise a three-step strategy to fine-tune the causal model under different scenarios accordingly. Empirical experiments show that CSR efficiently adapts to the target domains with only a few samples and outperforms state-of-the-art baselines on a wide range of scenarios, including our simulated environments, CartPole, CoinRun and Atari games.

Graph Alignment (GA) is an NP-hard combinatorial challenge. The existing methods usually work on small-scale graphs because their computational complexity grows as the square of the number of nodes. Recent techniques relied on alignment labels to divide the large source and target graphs into small ones respectively, which are then aligned separately. However, the alignment labels in expansive real-world graphs are often scarce, which makes GA even more challenging. To address these issues, we propose a novel self-supervised learning framework that is able to work on the graphs of million nodes without any alignment labels, where the existing methods usually fail to output an answer within a reasonable time. Our method also falls in the divide-and-conquer paradigm. We do not cluster the source and target graph separately by the existing graph clustering algorithms, because it is difficult to pair and then align the cluster-induced source and target subgraphs in the absence of alignment labels. We devise an Alignment-Aware Clustering (AAC) method to partition the source and target nodes jointly and keep the ought-to-be-aligned pairs of nodes within the same cluster as much as possible. Furthermore, we develop a Topology-Aware Repartition (TAR) that leverages pseudo-alignment labels as anchors for re-clustering, preserving detailed structural information within clusters. This enables graph neural networks to effectively enhance the graph representation learning for node alignments which are self-supervised by a graph-matching solver. Extensive experiments have demonstrated that our method greatly reduces computational demands and sustains high alignment accuracies in large-scale graph applications.

Recent advances in vision-language models have facilitated progress in sketch generation. However, existing specialized methods primarily focus on generic synthesis and lack mechanisms for precise control over sketch styles. In this work, we propose a training-free framework based on diffusion models that enables explicit style guidance via textual prompts and referenced style sketches. Unlike previous style transfer methods that overwrite key and value matrices in self-attention, we incorporate the reference features as auxiliary information with linear smoothing and leverage a style-content guidance mechanism. This design effectively reduces content leakage from reference sketches and enhances synthesis quality, especially in cases with low structural similarity between reference and target sketches. Furthermore, we extend our framework to support controllable multi-style generation by integrating features from multiple reference sketches, coordinated via a joint AdaIN module. Extensive experiments demonstrate that our approach achieves high-quality sketch generation with accurate style alignment and improved flexibility in style control. The official implementation of M3S is available at https://github.com/CMACH508/M3S.

In recent years, neural network-guided heuristic search algorithms, such as Monte-Carlo tree search and A$^\*$ search, have achieved significant advancements across diverse practical applications. Due to the challenges stemming from high state-space complexity, sparse training datasets, and incomplete environmental modeling, heuristic estimations manifest uncontrolled inherent biases towards the actual expected evaluations, thereby compromising the decision-making quality of search algorithms. Sampling exploration enhanced A$^\*$ (SeeA$^\*$) was proposed to improve the efficiency of A$^\*$ search by constructing an dynamic candidate subset through random sampling, from which the expanded node was selected. However, uniform sampling strategy utilized by SeeA$^\*$ facilitates exploration exclusively through the injection of randomness, which completely neglects the heuristic knowledge relevant to open nodes. Moreover, the theoretical support of cluster sampling remains ambiguous. Despite the existence of potential biases, heuristic estimations still encapsulate certain valuable information. In this paper, epistemic exploratory A$^\*$ search (KeeA$^\*$) is proposed to integrate heuristic knowledge for calibrating the sampling process. We first theoretically demonstrate that SeeA$^\*$ with cluster sampling outperforms uniform sampling due to the distribution-aware selection with higher variance. Building on this insight, cluster scouting and path-aware sampling are introduced in KeeA$^\*$ to further exploit heuristic knowledge to increase the sampling mean and variance, respectively, thereby generating higher-quality extreme candidates and enhancing overall decision-making performance. Finally, empirical results on retrosynthetic planning and logic synthesis demonstrate superior performance of KeeA$^*$ compared to state-of-the-art heuristic search algorithms.

Structure-based drug design (SBDD), aiming to generate 3D molecules with high binding affinity toward target proteins, is a vital approach in novel drug discovery. Although recent generative models have shown great potential, they suffer from unstable probability dynamics and mismatch between generated molecule size and the protein pockets geometry, resulting in inconsistent quality and off-target effects. We propose PAFlow, a novel target-aware molecular generation model featuring prior interaction guidance and a learnable atom number predictor. PAFlow adopts the efficient flow matching framework to model the generation process and constructs a new form of conditional flow matching for discrete atom types. A protein–ligand interaction predictor is incorporated to guide the vector field toward higher-affinity regions during generation, while an atom number predictor based on protein pocket information is designed to better align generated molecule size with target geometry. Extensive experiments on the CrossDocked2020 benchmark show that PAFlow achieves a new state-of-the-art in binding affinity (up to -8.31 Avg. Vina Score), simultaneously maintains favorable molecular properties.

Structure-based drug design (SBDD) is a critical subtask in the drug discovery process, with deep generative models playing a pivotal role. Inherently, drug design is a multi-objective task given the fact that a promising drug candidate must satisfy multiple properties. However, existing SBDD methods either focus solely on the binding affinity between molecules and target proteins while neglecting other crucial properties, or they assume that objective properties are independent of each other. Yet there are often potential relationships among properties, which can be conflicting-improving one property may lead to the deterioration of another. The lack of consideration for these relationships in current methods makes it unfeasible to generate molecules that simultaneously meet multiple objectives. To address the above issues, a multi-objective SBDD algorithm is proposed based on the diffusion model to optimize binding affinity and other drug properties simultaneously. Multiple expert networks are trained in parallel to predict properties for molecules in intermediate states and transmit gradients, and a causal graph is constructed through the causal discovery algorithm to unveil the underlying relationships among target properties. During the entire generation process, the joint distribution of target properties is decomposed in a reasonable manner according to the casual graph, and then the gradients of each property are applied to guide the optimizing direction of generation. Experimental results indicate that our model effectively optimizes multiple objectives simultaneously, generating molecules with greater drug potential compared to baseline models.

Prediction of drug-target interactions (DTIs) is a crucial step in drug discovery, and deep learning methods have shown great promise on various DTI datasets. However, existing approaches still face several challenges, including limited labeled data, hidden bias issue, and a lack of generalization ability to out-of-domain data. These challenges hinder the model's capacity to learn truly informative interaction features, leading to shortcut learning and inferior predictive performance on novel drug-target pairs. To address these issues, we propose MlanDTI, a semi-supervised domain adaptive multilevel attention network (Mlan) for DTI prediction. We utilize two pre-trained BERT models to acquire bidirectional representations enriched with information from unlabeled data. Then, we introduce a multilevel attention mechanism, enabling the model to learn domain-invariant DTIs at different hierarchical levels. Moreover, we present a simple yet effective semi-supervised pseudo-labeling method to further enhance our model's predictive ability in cross-domain scenarios. Experiments on four datasets show that MlanDTI achieves state-of-the-art performances over other methods under intra-domain settings and outperforms all other approaches under cross-domain settings. The source code is available at https://github.com/CMACH508/MlanDTI.

Molecular representations are crucial for accurately predicting molecular properties and are fundamental in drug design and related fields. Recently, self-supervised models have leveraged large-scale unlabeled molecular data, with transformers extending learning capabilities. However, existing frameworks predominantly rely on training data for understanding molecular structures, leading to limited generalization on diverse downstream tasks. Moreover, these models often fail to integrate molecular property knowledge effectively, further constraining their ability to generalize. To address these limitations, we propose a multi-task pre-training strategy, MulMol, which combines masked molecular structure reconstruction with property reconstruction, and incorporates contrastive learning to enhance performance. Furthermore, we introduce a similarity comparison task to ensure that the model maintains property consistency despite structural deficiencies at different positions. Evaluations on 11 benchmark datasets demonstrate that MulMol outperforms existing baselines in molecular property prediction across various domains, offering a robust AI-driven tool for drug discovery. The source code is publicly accessible on https://github.com/CMACH508/MulMol.

Causal intervention has been widely used in deep learning to tackle the confounding problems when the data are out of distribution. A representative class of causal-intervention-based strategy is invariant risk minimization, yet manually annotating is indispensable to indicate the environmental splits. A feasible solution is to use a pair of complementary attention, one of which focuses on the foreground target and another extracts the background features. The environments are then split unsupervisedly. However, in terms of medical images, the background is not as simple as normal images (e.g., a camel in desert), and using all the background feature as confounder is redundant and will degrade the expected generalization performance. In this paper, we develop a novel multimodal deep neural network which automatically extracts the confounded background features of chest X-ray (CXR) images by a multimodal approach, i.e., feature fusion with explicit confounders such as demographic information. To achieve this, we design an architecture with two modules, a classifier module taking images as input and an environment split module taking demographic tables as input. The cross attention is then conducted between the encoded background features and demographic features to extract the actual confounded background features. The proposed method meaningfully improves the generalization performance on multiple CXR datasets and accurately locates the lesions. The source code is available at https://github.com/CMACH508/CausalCXR.

The generation of drug-like, high-quality molecules from scratch within the expansive chemical space is a significant challenge in drug discovery. In previous research, value-based reinforcement learning algorithms have been utilized to optimize multiple desired properties simultaneously. Randomness is injected into the decision-making process through ε-Greedy or stochastic sampling to enable the generation of a diverse ensemble of molecules, which usually encounters a trade-off between the optimality and diversity of these generated molecules. Moreover, novelty has not been explicitly addressed as an optimization objective, and the distinctiveness of generated molecules from the reference molecules is not guaranteed. In this paper, novelty-encouraged beam clustering (NeBC) search algorithm is proposed for de novo drug design. A clustering strategy is integrated with heuristic value-guided beam search to strike a balance between the optimality and diversity of the generated molecules. An intrinsic reward, which is measured by the disagreement of a group of experts trained on reference molecules, is proposed to encourage novelty explicitly. Experimental results demonstrate that NeBC search not only achieves a balanced trade-off between optimality and diversity but also effectively enhances the novelty of the generated molecules. The source code is publicly accessible on https://github.com/CMACH508/NeBC.

Monte-Carlo tree search (MCTS) and reinforcement learning contributed crucially to the success of AlphaGo and AlphaZero, and A$^*$ is a tree search algorithm among the most well-known ones in the classical AI literature. MCTS and A$^*$ both perform heuristic search and are mutually beneficial. Efforts have been made to the renaissance of A$^*$ from three possible aspects, two of which have been confirmed by studies in recent years, while the third is about the OPEN list that consists of open nodes of A$^*$ search, but still lacks deep investigation. This paper aims at the third, i.e., developing the Sampling-exploration enhanced A$^*$ (SeeA$^*$) search by constructing a dynamic subset of OPEN through a selective sampling process, such that the node with the best heuristic value in this subset instead of in the OPEN is expanded. Nodes with the best heuristic values in OPEN are most probably picked into this subset, but sometimes may not be included, which enables SeeA$^*$ to explore other promising branches. Three sampling techniques are presented for comparative investigations. Moreover, under the assumption about the distribution of prediction errors, we have theoretically shown the superior efficiency of SeeA$^*$ over A$^*$ search, particularly when the accuracy of the guiding heuristic function is insufficient. Experimental results on retrosynthetic planning in organic chemistry, logic synthesis in integrated circuit design, and the classical Sokoban game empirically demonstrate the efficiency of SeeA$^*$, in comparison with the state-of-the-art heuristic search algorithms.

Designing 3D molecules with high binding affinity for specific protein targets is crucial in drug design. One challenge is that the atomic interaction between molecules and proteins in 3D space has to be taken into account. However, the existing target-aware methods solely model the joint distribution between the molecules and proteins, disregarding the binding affinities between them, which leads to limited performance. In this paper, we propose an explainable diffusion model to generate molecules that can be bound to a given protein target with high affinity. Our method explicitly incorporates the chemical knowledge of protein-ligand binding affinity into the diffusion model, and uses the knowledge to guide the denoising process towards the direction of high binding affinity. Specifically, an SE(3)-invariant expert network is developed to fit the Vina scoring functions and jointly trained with the denoising network, while the domain knowledge is distilled and conveyed from Vina functions to the expert network. An effective guidance is proposed on both continuous atom coordinates and discrete atom types by taking advantages of the gradient of the expert network. Experiments on the benchmark CrossDocked2020 demonstrate the superiority of our method. Additionally, an atom-level explanation of the generated molecules is provided, and the connections with the domain knowledge are established.

Reinforcement learning (RL) has been applied to financial portfolio management in recent years. Current studies mostly focus on profit accumulation without much consideration of risk. Some risk‐return balanced studies extract features from price and volume data only, which is highly correlated and missing representation of risk features. To tackle these problems, we propose a weight control unit (WCU) to effectively manage the position of portfolio management in different market statuses. A loss penalty term is also designed in the reward function to prevent sharp drawdown during trading. Moreover, stock spatial interrelation representing the correlation between two different stocks is captured by a graph convolution network based on fundamental data. Temporal interrelation is also captured by a temporal convolutional network based on new factors designed with price and volume data. Both spatial and temporal interrelation work for better feature extraction from historical data and also make the model more interpretable. Finally, a deep deterministic policy gradient actor–critic RL is applied to explore optimal policy in portfolio management. We conduct our approach in a challenging non‐short‐selling market, and the experiment results show that our method outperforms the state‐of‐the‐art methods in both profit and risk criteria. Specifically, with 6.72% improvement on an annualized rate of return, 7.72% decrease in maximum drawdown, and a better annualized Sharpe ratio of 0.112. Also, the loss penalty and WCU provide new aspects for future work in risk control.

Generating high-quality and drug-like molecules from scratch within the expansive chemical space presents a significant challenge in the field of drug discovery. In prior research, value-based reinforcement learning algorithms have been employed to generate molecules with multiple desired properties iteratively. The immediate reward was defined as the evaluation of intermediate-state molecules at each step, and the learning objective would be maximizing the expected cumulative evaluation scores for all molecules along the generative path. However, this definition of the reward was misleading, as in reality, the optimization target should be the evaluation score of only the final generated molecule. Furthermore, in previous works, randomness was introduced into the decision-making process, enabling the generation of diverse molecules but no longer pursuing the maximum future rewards. In this paper, immediate reward is defined as the improvement achieved through the modification of the molecule to maximize the evaluation score of the final generated molecule exclusively. Originating from the A search, path consistency (PC), i.e., values on one optimal path should be identical, is employed as the objective function in the update of the value estimator to train a multi-objective de novo drug designer. By incorporating the value into the decision-making process of beam search, the DrugBA algorithm is proposed to enable the large-scale generation of molecules that exhibit both high quality and diversity. Experimental results demonstrate a substantial enhancement over the state-of-the-art algorithm QADD in multiple molecular properties of the generated molecules.

Encoding sketches as Gaussian mixture model (GMM)-distributed latent codes is an effective way to control sketch synthesis. Each Gaussian component represents a specific sketch pattern, and a code randomly sampled from the Gaussian can be decoded to synthesize a sketch with the target pattern. However, existing methods treat the Gaussians as individual clusters, which neglects the relationships between them. For example, the giraffe and horse sketches heading left are related to each other by their face orientation. The relationships between sketch patterns are important messages to reveal cognitive knowledge in sketch data. Thus, it is promising to learn accurate sketch representations by modeling the pattern relationships into a latent structure. In this article, we construct a tree-structured taxonomic hierarchy over the clusters of sketch codes. The clusters with the more specific descriptions of sketch patterns are placed at the lower levels, while the ones with the more general patterns are ranked at the higher levels. The clusters at the same rank relate to each other through the inheritance of features from common ancestors. We propose a hierarchical expectation-maximization (EM)-like algorithm to explicitly learn the hierarchy, jointly with the training of encoder-decoder network. Moreover, the learned latent hierarchy is utilized to regularize sketch codes with structural constraints. Experimental results show that our method significantly improves controllable synthesis performance and obtains effective sketch analogy results.

Deep learning methods have demonstrated promising performance on the NP-hard Graph Matching (GM) problems. However, the state-of-the-art methods usually require the ground-truth labels, which may take extensive human efforts or be impractical to collect. In this paper, we present a robust self-supervised bidirectional learning method (IA-SSGM) to tackle GM in an unsupervised manner. It involves an affinity learning component and a classic GM solver. Specifically, we adopt the Hungarian solver to generate pseudo correspondence labels for the simple probabilistic relaxation of the affinity matrix. In addition, a bidirectional recycling consistency module is proposed to generate pseudo samples by recycling the pseudo correspondence back to permute the input. It imposes a consistency constraint between the pseudo affinity and the original one, which is theoretically supported to help reduce the matching error. Our method further develops a graph contrastive learning jointly with the affinity learning to enhance its robustness against the noise and outliers in real applications. Experiments deliver superior performance over the previous state-of-the-arts on five real-world benchmarks, especially under the more difficult outlier scenarios, demon- strating the effectiveness of our method.

Graphic sketch representations are effective for representing sketches. Existing methods take the patches cropped from sketches as the graph nodes, and construct the edges based on sketch's drawing order or Euclidean distances on the canvas. However, the drawing order of a sketch may not be unique, while the patches from semantically related parts of a sketch may be far away from each other on the canvas. In this paper, we propose an order-invariant, semantics-aware method for graphic sketch representations. The cropped sketch patches are linked according to their global semantics or local geometric shapes, namely the synonymous proximity, by computing the cosine similarity between the captured patch embeddings. Such constructed edges are learnable to adapt to the variation of sketch drawings, which enable the message passing among synonymous patches. Aggregating the messages from synonymous patches by graph convolutional networks plays a role of denoising, which is beneficial to produce robust patch embeddings and accurate sketch representations. Furthermore, we enforce a clustering constraint over the embeddings jointly with the network learning. The synonymous patches are self-organized as compact clusters, and their embeddings are guided to move towards their assigned cluster centroids. It raises the accuracy of the computed synonymous proximity. Experimental results show that our method significantly improves the performance on both controllable sketch synthesis and sketch healing.

Finding known homologs of newly predicted proteins is essential for understanding their functions and mechanisms. It is a highly complex task because proteins undergo various changes during evolution. Traditional methods based on sequence or structure alignment either have low accuracy or take a long time. Recent deep learning-based methods primarily focus on structural information, yet they can’t fully exploiting protein information. To solve this problem, in this paper, we propose a novel general-purpose protein embedding model that can be used for homology search. It first employs a protein language pre-trained model to extract protein sequence embeddings, capturing intricate biological patterns. Subsequently, a Transformer integrating protein structural information generates the high-level representations. By combining protein sequence and structural features, the model can effectively exploit the rich contextual and spatial information inherent in proteins. We applied the model to the SCOP dataset for protein superfamily classification, achieving a classification accuracy of 86.97%, outperforming state-of-the-art method by 7.91%. The source code has been published on GitHub (https://github.com/CMACH508/PEST).

Epilepsy is a prevalent chronic neurological disorder in childhood, imposing a heavy burden on patients and their families. The development of deep learning and the accumulation of clinical medical data has led to a surge in neural network algorithms proposed for the automatic detection of childhood epilepsy using electroencephalogram (EEG) signals. However, for Benign Childhood Epilepsy with Centro-Temporal Spikes (BECT), the most common type of childhood epilepsy, there is no large-scale public dataset available for its detection. Moreover, although there were a few studies of BECT using private data, they focused only on identifying the presence of abnormal discharges, ignoring the sleep stage of the discharges, which is actually a critical factor for the doctor’s diagnosis. To tackle these challenges, we create a BECT dataset containing more than 38,000 real samples from 100 subjects, meticulously annotated by neurology experts. Building on this dataset, we propose a multi-view attention and multi-branch graph convolutional neural network (MAMB) to differentiate abnormal discharges and classify sleep stages in the samples. Experimental results demonstrate that the model benefits from three-dimensional attention mechanisms in the spatial, temporal, and spectral domains, allowing better exploration of intrinsic relationships within EEG signals. Additionally, the sleep and BECT branches enhance the model’s ability to detect abnormal discharges and differentiate various sleep stages. Our model achieves 88.07% and 90.91% accuracy for BECT classification and sleep stage staging, and 81.93% for the four-classification task of simultaneously judging BECT and sleep stage. In addition, the introduction of the multi-view attention mechanism makes the model interpretable and raises hopes of further assisting experts in disease and medication analysis. The implementation code is shown at github1.

Sleep staging using electroencephalogram (EEG) is of great significance for diagnosing sleep disorders. Recently, due to the high cost of manually annotating EEG signals and the problem of domain shift across different datasets, many unsupervised domain adaptation methods have been applied to sleep staging. However, all these methods are single-source unsupervised domain adaptation (SUDA) methods. When applied to the multi-source unsupervised domain adaptation (MUDA) problem, most SUDA methods treat all source domains equally, without considering the domain shift between them. As a result, achiving feature alignment across all domains becomes a challenge for these SUDA methods. In our proposed model, we address the challenge of domain shift in multi-source unsupervised domain adaptation (MUDA) for automatic sleep staging. Specially, we create a domain-specific branch for each pair of source and target domains, which focuses on aligning the features extracted from the respective domains using an adversarial-based domain adaptation approach. Additionally, we incorporate a domain-invariant branch into our model for learning the domain-invariant representation. We also introduce an adaptive-based mixing strategy to assigns weights to each branches, which can dynamically adjust the importance of each branch depending on their relevance to the specific prediction task. The experiments conducted on three public datasets show the superior performance of our model. Compared to other state-of-the-art MUDA models, our model’s average classification accuracy improves by 2% to 9.9%. The source code is available at https://github.com/CMACH508/MUDAEEG.

Modern very-large-scale integrated (VLSI) circuit placement with huge state space is a critical task for achieving layouts with high performance. Recently, reinforcement learning (RL) algorithms have made a promising breakthrough to dramatically save design time than human effort. However, the previous RL-based works either require a large dataset of chip placements for pre-training or produce illegal final placement solutions. In this paper, DeepTH, a three-head policy gradient placer, is proposed to learn from scratch without the need of pre-training, and generate superior chip floorplans. Graph neural network is initially adopted to extract the features from nodes and nets of chips for estimating the policy and value. To efficiently improve the quality of floorplans, a reconstruction head is employed in the RL network to recover the visual representation of the current placement, by enriching the extracted features of placement embedding. Besides, the reconstruction error is used as a bonus during training to encourage exploration while alleviating the sparse reward problem. Furthermore, the expert knowledge of floorplanning preference is embedded into the decision process to narrow down the potential action space. Experiment results on the ISPD 2005 benchmark have shown that our method achieves 19.02% HPWL improvement than the analytic placer DREAMPlace and 19.89% improvement at least than the state-of-the-art RL algorithms.

Existing machine learning methods are effective in portfolio optimization on a small pool of assets. This is still not optimal because a larger number of assets in markets offers more opportunities for investors. However, existing methods are usually not scalable to large amount of assets which brings new challenges of high dimensionality and computing complexity. In this paper, we present a neural network temporal factor analysis (NN-TFA) model for dimensionality reduction and it enables us to build a scalable deep reinforcement learning method for large-scale portfolio management. Traditional TFA models the relation between asset prices and real economic activities via a small set of independent hidden factors. NN-TFA is developed from the traditional TFA by replacing the linear autoregressive model over the hidden factors with a neural network function, which well captures the complicated temporal patterns. The hidden factors are then sent to a policy network to generate portfolio weights. A calibration module to extract information from other assets features and a ratio module to catch the trend of the selected assets pool are proposed to enhance the performance of the policy network. Extensive tests demonstrate that our methods are capable of handling large-scale datasets and achieving promising results.

Protein binding site prediction is an important prerequisite for the discovery of new drugs. Usually, natural 3D U-Net is adopted as the standard site prediction framework to do per-voxel binary mask classification. However, this scheme only performs feature extraction for single-scale samples, which may bring the loss of global or local information, resulting in incomplete, artifacted or even missed predictions. To tackle this issue, we propose a network called GLPocket, which is based on the Lmser (Least mean square error reconstruction) network and utilizes multi-scale representation to predict binding sites. Firstly, GLPocket uses Target Cropping Block (TCB) for targeted prediction. TCB selects the local interested feature from the global representations to perform concentrated prediction, and reduces the volume of feature maps to be calculated by 82% without adding additional parameters. It integrates global distribution information into local regions, making prediction more concentrated on decoding stage. Secondly, GLPocket establishes long-range relationship of patches within the local region with Transformer Block (TB), to enrich local context semantic information. Experiments show that GLPocket improves by 0.5%-4% on DCA Top-n prediction compared with previous state-of-the-art methods on four datasets. Our code has been released in https://github.com/CMACH508/GLPocket.

Reinforcement learning with the help of neural-guided search consumes huge computational resources to achieve remarkable performance. Path consistency (PC), i.e., $f$ values on one optimal path should be identical, was previously imposed on MCTS by PCZero to improve the learning efficiency of AlphaZero. Not only PCZero still lacks a theoretical support but also considers merely board games. In this paper, PCZero is generalized into GW-PCZero for real applications with non-zero immediate reward. A weighting mechanism is introduced to reduce the variance caused by scouting's uncertainty on the $f$ value estimation. For the first time, it is theoretically proved that neural-guided MCTS is guaranteed to find the optimal solution under the constraint of PC. Experiments are conducted on the Atari $100$k benchmark with $26$ games and GW-PCZero achieves $198\%$ mean human performance, higher than the state-of-the-art EfficientZero's $194\\%$, while consuming only $25\\%$ of the computational resources consumed by EfficientZero.

Accurate prediction of synergistic effects of drug combinations can reduce the experimental costs for drug development and facilitate the discovery of novel efficacious combination therapies for clinical studies. The drug combinations with high synergy scores are regarded as synergistic ones, while those with moderate or low synergy scores are additive or antagonistic ones. The existing methods usually exploit the synergy data from the aspect of synergistic drug combinations, paying little attention to the additive or antagonistic ones. Also, they usually do not leverage the common patterns of drug combinations across different cell lines. In this paper, we propose a multi-channel graph autoencoder (MGAE)-based method for predicting the synergistic effects of drug combinations (DC), and shortly denote it as MGAE-DC. A MGAE model is built to learn the drug embeddings by considering not only synergistic combinations but also additive and antagonistic ones as three input channels. The later two channels guide the model to explicitly characterize the features of non-synergistic combinations through an encoder-decoder learning process, and thus the drug embeddings become more discriminative between synergistic and non-synergistic combinations. In addition, an attention mechanism is incorporated to fuse each cell-line's drug embeddings across various cell lines, and a common drug embedding is extracted to capture the invariant patterns by developing a set of cell-line shared decoders. The generalization performance of our model is further improved with the invariant patterns. With the cell-line specific and common drug embeddings, our method is extended to predict the synergy scores of drug combinations by a neural network module. Experiments on four benchmark datasets demonstrate that MGAE-DC consistently outperforms the state-of-the-art methods. In-depth literature survey is conducted to find that many drug combinations predicted by MGAE-DC are supported by previous experimental studies. The source code and data are available at https://github.com/yushenshashen/MGAE-DC.

Protein binding site prediction is an important prerequisite task of drug discovery and design. While binding sites are very small, irregular and varied in shape, making the prediction very challenging. Standard 3D U-Net has been adopted to predict binding sites but got stuck with unsatisfactory prediction results, incomplete, out-of-bounds, or even failed. The reason is that this scheme is less capable of extracting the chemical interactions of the entire region and hardly takes into account the difficulty of segmenting complex shapes. In this paper, we propose a refined U-Net architecture, called RefinePocket, consisting of an attention-enhanced encoder and a mask-guided decoder. During encoding, taking binding site proposal as input, we employ Dual Attention Block (DAB) hierarchically to capture rich global information, exploring residue relationship and chemical correlations in spatial and channel dimensions respectively. Then, based on the enhanced representation extracted by the encoder, we devise Refine Block (RB) in the decoder to enable self-guided refinement of uncertain regions gradually, resulting in more precise segmentation. Experiments show that DAB and RB complement and promote each other, making RefinePocket has an average improvement of 10.02% on DCC and 4.26% on DVO compared with the state-of-the-art method on four test sets.

It is an essential step to locate the binding sites or pockets of drug molecules on protein structure in drug design. This is challenging because the 3D protein structures are usually in complicated, irregular shape and the pockets are relatively small. Existing deep learning methods for this task are U-Net models, and they have forward skip connections to efficiently transfer features of different levels of 3D structure from encoder to decoder for improving pocket prediction. However, there is still room to improve prediction accuracy. In this paper, we propose RecurPocket, a recurrent Lmser (Least mean square error reconstruction) network for pocket detection. A gated recurrent refinement is devised in RecurPocket to enhance the representation learning on the 3D protein structures. This is fulfilled by feedback connections in RecurPocket network from decoder to encoder, recurrently and progressively improving the feature embedding for accurate prediction. Moreover, a 3D gate mechanism filters out irrelevant information through the feedback links that interfere with detection, making the prediction more precise and clear. Experiments show that RecurPocket improves by 3%-9% on top-n prediction compared with previous state-of-the-art on five benchmark data sets. The source code and trained model are available at https://github.con CMACH508/RecurPocket.

A number of existing causal effect estimation approaches for time-to-event outcomes in observational studies employ the propensity score-based strategies to balance covariates in treatment groups and reduce confounding effects. The difficulty of causal inference using propensity score lies in the issue that correctly specified propensity score model is required to obtain causal effects, otherwise, substantial bias may be introduced. In this paper, we develop a nonparametric weighting approach, which balances the covariates between treatment groups by Kernel Mean Matching (KMM) with Mahalanobis distance kernels. The proposed estimator adapts KMM procedure to causal effect estimation for time-to-event outcomes, rather than its original applications on empirical risk for classification or regression. Given universal kernels, KMM procedure leads to the balance of covariate distributions. For time-to-event outcome which has highly nonlinear relations with covariates, balancing covariates at distribution level is necessary rather than finite order moments. Furthermore, in consideration of practical applications, unlike the Euclidean distance, the Mahalanobis distance is essentially suitable for the anisotropic real-world data with extremely different scales and highly correlated dimensions, so that we can make the best use of the optimization power. We theoretically proved the consistency of the proposed estimator. Simulations and real-world applications illustrate the effectiveness of the proposed method.

Timely recognition of extubation readiness is critical, because prolonged and premature intubation will lead to sever complications and costs. Clinical assessment is time consuming and challenging and it has attracted increasing attention of machine learning in recent years. However, the data used for extubation predictions have the following flaws: 1) Manual recording errors and missing data; 2) Unreliable ventilation labels due to inadequate judgement from clinicians. Both may possibly lead to wrong ventilation labels, but existing machine learning methods for extubation prediction largely ignored this critical issue. In this paper, we proposed a self-rectifying deep learning method for extubation readiness prediction, called VentSR. It improves the prediction performance by a self-rectifying strategy, and the rectification is achieved through model training without clinical experience. To be detailed, VentSR firstly identifies possibly wrong samples by two components: Inconsistency between K-means and Labels (IKL) and Inconsistency between Model Predictions and Labels (IPL). IKL partitions a rough subset, and IPL iteratively refines this subset through training. Additionally, we designed Adjustment Operation to enhance IPL ability for refinement. Samples identified in this subset are rectified and used to train the model. The unrectified test set is directly fed into the trained model to obtain prediction results. Experiments demonstrate that VentSR outperforms other baselines. Further comparisons on high-confidence test set indicate that VentSR achieves 79.4 AUPRC, increasing by 26.0%. Feature importance analysis and case study illustration again reveals that VentSR are of potential practical usage of informing clinicians with accurate extubation readiness.

Predicting the synergistic effects of drug combinations can accelerate the identification process of novel potential combination therapies for clinical studies. Although extensive efforts have been made in the field, the problem is still challenging due to the high sparsity of drug combinations’ synergy data and the existence of false positive combinations resulted from the noise in experiments. In this paper, we develop a Knowledge Graph Embedding-based method for predicting the synergistic effects of Drug Combinations, namely KGE-DC, which fully extracts the features of drug combinations. Firstly, a largescale knowledge graph including drugs, targets, enzymes and transporters is constructed, therefore, the sparsity of the drug combinations’ data is reduced and the reliability of the data is increased. Then, knowledge graph embedding, which are capable of capturing complex semantic information of various entities in the knowledge graph, is adopted for learning low-dimensional representations for the drugs and cell lines. Finally, the synergy scores of drug combinations are predicted based on the drug and cell line embeddings of the drug combinations’ synergy data. Extensive experiments on benchmark dataset with four different synergy types demonstrate that KGE-DC outperforms state-of the-art methods on both the regression and classification tasks, namely predicting the synergy scores of drug combinations and predicting whether the drug combinations are synergistic combinations. Our results indicate that KGE-DC is a valuable tool to facilitate the discovery of novel combination therapies for cancer treatment. The implemented code and experimental dataset are available online at https://github.com/yushenshashen/KGE-DC.

In recent years, deep reinforcement learning have made great breakthroughs on board games. Still, most of the works require huge computational resources for a large scale of environmental interactions or self-play for the games. This paper aims at building powerful models under a limited amount of self-plays which can be utilized by a human throughout the lifetime. We proposes a learning algorithm built on AlphaZero, with its path searching regularised by a path consistency (PC) optimality, i.e., values on one optimal search path should be identical. Thus, the algorithm is shortly named PCZero. In implementation, historical trajectory and scouted search paths by MCTS makes a good balance between exploration and exploitation, which enhances the generalization ability effectively. PCZero obtains $94.1%$ winning rate against the champion of Hex Computer Olympiad in 2015 on $13\times 13$ Hex, much higher than $84.3%$ by AlphaZero. The models consume only $900K$ self-play games, about the amount humans can study in a lifetime. The improvements by PCZero have been also generalized to Othello and Gomoku. Experiments also demonstrate the efficiency of PCZero under offline learning setting.

While the quality of face manipulation has been improved tremendously, the ability to control face components, e.g., eyebrows, is still limited. Although existing methods have realized component editing with user-provided geometry guidance, such as masks or sketches, their performance is largely dependent on the user's painting efforts. To address these issues, we propose Box-FaceS, a bidirectional method that can edit face components by simply translating and zooming the bounding boxes. This framework learns representations for every face component, independently, as well as a high-dimensional tensor capturing face outlines. To enable box-guided face editing, we develop a novel Box Adaptive Modulation (BAM) module for the generator, which first transforms component embeddings to style parameters and then modulates visual features inside a given box-like region on the face outlines. A cooperative learning scheme is proposed to impose independence between face outlines and component embeddings. As a result, it is flexible to determine the component style by its embedding, and to control its position and size by the provided bounding box. Box-FaceS also learns to transfer components between two faces while maintaining the consistency of image content. In particular, Box-FaceS can generate creative faces with reasonable exaggerations, requiring neither supervision nor complex spatial morphing operations. Through the comparisons with state-of-the-art methods, Box-FaceS shows its superiority in component editing, both qualitatively and quantitatively. To the best of our knowledge, Box-FaceS is the first approach that can freely edit the position and shape of the face components without editing the face masks or sketches. Our implementation is available at https://github.com/CMACH508/Box-FaceS.

Identifying synergistic drug combinations (SDCs) is a great challenge due to the combinatorial complexity and the fact that SDC is cell line specific. The existing computational methods either did not consider the cell line specificity of SDC, or did not perform well by building model for each cell line independently. In this paper, we present a novel encoder-decoder network named SDCNet for predicting cell line-specific SDCs. SDCNet learns common patterns across different cell lines as well as cell line-specific features in one model for drug combinations. This is realized by considering the SDC graphs of different cell lines as a relational graph, and constructing a relational graph convolutional network (R-GCN) as the encoder to learn and fuse the deep representations of drugs for different cell lines. An attention mechanism is devised to integrate the drug features from different layers of the R-GCN according to their relative importance so that representation learning is further enhanced. The common patterns are exploited through partial parameter sharing in cell line-specific decoders, which not only reconstruct the known SDCs but also predict new ones for each cell line. Experiments on various datasets demonstrate that SDCNet is superior to state-of-the-art methods and is also robust when generalized to new cell lines that are different from the training ones. Finally, the case study again confirms the effectiveness of our method in predicting novel reliable cell line-specific SDCs.

Hyperspectral imaging (HSI) greatly improves the capacity to identify and monitor ground objects due to the high spectral resolution. As the real-time remote sensing monitoring and warning tasks are getting more attention, new algorithms for low-power on-board classification are required to reduce the transmission time of satellite downlink. In this paper, we propose the Multi-Scale Local Maximum Random Forest (MSLM-RF) to significantly reduce the energy consumption while retaining high classification accuracy. The proposed MSLM-RF uses multi-scale maximum filters for spatial feature extraction and Random Forest for classification after spectral and spatial features fusion. The spatial features are efficiently extracted with low computational complexity by regarding the maximum light intensity values in different ranges of pixels as anchor points. MSLM-RF only consists of integer comparisons and a few additions, thereby eliminating the energy-hungry operations such as multiplication and exponentiation. According to experimental results on the HSI benchmark datasets, MSLM-RF delivers a better trade-off in accuracy and computational complexity than the state-of-the-art classification algorithms. Besides, MSLM-RF gets higher average classification accuracy and lower energy consumption than the previous on-board algorithms. The obtained results show the suitability of the proposed algorithm to accomplish practical real-time classification tasks on-board with low energy consumption.

Most existing reinforcement learning (RL)-based portfolio management models do not take into account the market conditions, which limits their performance in risk-return balancing. In this paper, we propose DeepTrader, a deep RL method to optimize the investment policy. In particular, to tackle the risk-return balancing problem, our model embeds macro market conditions as an indicator to dynamically adjust the proportion between long and short funds, to lower the risk of market fluctuations, with the negative maximum drawdown as the reward function. Additionally, the model involves a unit to evaluate individual assets, which learns dynamic patterns from historical data with the price rising rate as the reward function. Both temporal and spatial dependencies between assets are captured hierarchically by a specific type of graph structure. Particularly, we find that the estimated causal structure best captures the interrelationships between assets, compared to industry classification and correlation. The two units are complementary and integrated to generate a suitable portfolio which fits the market trend well and strikes a balance between return and risk effectively. Experiments on three well-known stock indexes demonstrate the superiority of DeepTrader in terms of risk-gain criteria.

Existing deep learning methods for graph matching(GM) problems usually considered affinity learningto assist combinatorial optimization in a feedforward pipeline, and parameter learning is executed by back-propagating the gradients of the matching loss. Such a pipeline pays little attention to the possible complementary benefit from the optimization layer to the learning component. In this paper, we overcome the above limitation under a deep bidirectional learning framework.Our method circulates the output of the GM optimization layer to fuse with the input for affinity learning. Such direct feedback enhances the input by a feature enrichment and fusion technique, which exploits andintegrates the global matching patterns from the deviation of the similarity permuted by the current matching estimate. As a result, the circulation enables the learning component to benefit from the optimization process, taking advantage of both global feature and the embedding result which is calculated by local propagationthrough node-neighbors. Moreover, circulation consistency induces an unsupervised loss that can be implemented individually or jointly to regularize the supervised loss. Experiments on challenging datasets demonstrate the effectiveness of our methods for both supervised learning and unsupervised learning.

The skip connections in U-Net pass features from the levels of encoder to the ones of decoder in a symmetrical way, which makes U-Net and its variants become state-of-the-art approaches for biomedical image segmentation. However, the U-Net skip connections are unidirectional without considering feedback from the decoder, while this paper exploits the feedback information to refine the segmentation. We develop a deep bidirectional network based on the least mean square error reconstruction (Lmser) self-organizing network, an early model that folds an autoencoder along the central hidden layer such that the neurons on the paired layers between encoder and decoder merge into one, equivalently forming bidirectional skip connections between encoder and decoder. We find that the feedback links indeed increase the segmentation accuracy, but may also bring certain noise into the segmentation. To tackle this problem, we present a gating and masking mechanism on the feedback connections to filter the irrelevant information. Experimental results on MoNuSeg, TNBC, and EM membrane datasets demonstrate that our method are robust and outperforms state-of-the-art methods.

It is challenging to build a machine learning model for automatic arrhythmia diagnosis from Electrocardiograph (ECG) signals, because the variation in ECG signals is big between different patients or over time, and the available training datasets usually contain limited, unbalanced number of data for multiple disease types. Most existing methods relied on labeled data from a single dataset, and the performance is poor when generalizing to unseen heart disease types, limited labels, or distribution shifts. In this paper, we propose a multi-source unsupervised domain adaption (MUDA) neural network for ECG classification, to make effective use of data of multiple sources and improve the model’s generalization ability. Our model is featured by a two-branch domain adaption and a sample-imbalance aware mixing strategy to fuse the information across domains. Specifically, one branch is devised to learn domain-invariant representation, while the other is to extract domain-specific features. The two branches align the ECG in the target domain to individual source domain in an exclusive and complementary manner, leading to enhanced discriminative features for domain invariant/specific classifiers. The final prediction, which is a linear combination of the domain classification decisions, is very robust and accurate, by making use of the prior distribution of sample size across domains to place confidence scores over each classifier. Experiments on five ECG datasets indicate superior performance of our method over the existing ones.

Developing an efficient system for automated detection and segmentation of intracranial aneurysms (IAs) became an active research topic recently. However, existing methods are poor in detecting small IA with high false positives. In this paper, we present a feature enrichment (FE) based deep learning method for robust IA detection and segmentation. The FE technique is featured by reconstructing a 3D model from all computed tomography angiography (CTA) images and then projecting 3D information into the so-called projection images of different slicing levels along various directions. The appearances of aneurysms in the projection images are enhanced in morphology and 3D neighborhood features, and thus are easy to be detected and segmented by the widely-used faster RCNN and V-Net. To evaluate our method, we collect CTA images from 145 patients (including 148 IAs) for training and testing. The proposed method achieves 96.0 % sensitivity for all aneurysms, and 80.0% sensitivity for aneurysms smaller than 4mm, better than the state-of-the-art methods. Also, the segmentation performance is improved on the detected IAs.

S1eep staging via electroencephalogram (EEG) is the fundamental step to sleep quality assessment and disease diagnose. Deep learning methods have been demonstrated to be promising for automatic sleep staging, but the performance is still not satisfied, because learning representations over the EEG signals is very challenging. In this paper, we propose RecSleepNet, an automatic sleep staging model based on a hybrid structure of Convolutional Neural Networks (CNN) and Long Short Term Memory (LSTM) networks. To enhance the representation learning, we devise a Convolutional Reconstruction Block (CRB), which enforces the extracted features to reconstruct the previous low-level input well. Moreover, we introduce a Global Max and Average Pooling Block (GMAPB) to squeeze parameter numbers and extract shift-invariant features. Experiments on four public datasets of single-channel EEG signals indicate that RecSleepNet achieves better or at least comparable performance to the state-of-the-art methods. CRB and GMAPB enable the training to be more efficient with fewer parameters b ut faster convergence.

To fully understand the detailed regulation mechanism of genomes and their functions, increasing computational methods have been developed to predict the TF binding site and intensity mainly based on DNA sequences or epigenomic data but ignoring the TF binding data across cell types. To address this problem, we proposed a multimodal deep learning framework, DeepBSI, to predict TF binding site and intensity in target cell type by leveraging the corresponding TF binding data across cell types. The framework can not only detect associations between sequence context features but also incorporate the correlations between TF binding signal values within and across cell types to make the prediction. In addition, the front modules of the framework employ the same convolutional neural network (CNN) and recurrent neural network (RNN) hybrid architecture model providing valuable information of TF motifs and their interactions, which make the framework interpretable. Applying DeepBSI to ten representative TFs across five cell types proved that models contain the TF binding information across cell types can significantly improve the performance of models in both TF binding site and intensity predicting tasks. The implemented code and experimental dataset are available online at https://github.com/yushenshashen/DeepBSI.

The coronavirus disease 2019 (COVID-19) epidemic continues to spread rapidly around the world and nearly 20 millions people are infected. This paper utilises both single-locus analysis and joint-SNPs analysis for detection of significant single nucleotide polymorphisms (SNPs) in the phenotypes of symptomatic versus asymptomatic, the early collection time versus the late collection time, the old versus the young, and the male versus the female. Also, this paper analyses the relationship between any two SNPs via linkage disequilibrium analysis, and visualises the patterns of cumulative mutations of SNPs over collection time. The results are in three folds. First, the SNP which locates at the nucleotide position 4321 is found to be an independent significant locus associated with all the first three phenotypes. Moreover, 12 significant SNPs are found in the first two studies. Second, gene orf1ab containing SNP-4321 is detected to be significantly associated with the first three phenotypes, and the three genes S, ORF3a, and N, are detected to be significant in the first two phenotypes. Third, some of the detected genes or SNPs are related to the SARS-COV-2 as supported by literature survey, which indicates that the results here may be helpful for further investigation.

For the problem of early detection of atrial fibrillation (AF) from electrocardiogram (ECG), it is difficult to capture subject-invariant discriminative features from ECG signals, due to the high variation in ECG morphology across subjects and the noise in ECG. In this paper, we propose an Discrete Biorthogonal Wavelet Transform (DBWT) Based Convolutional Neural Network (CNN) for AF detection, shortly called DBWT-AFNet. In DBWT-AFNet, rather than directly feeding ECG into CNN, DBWT is used to separate sub-signals in frequency band of heart beat from ECG, whose output is fed to CNN for AF diagnosis. Such sub-signals are better than the raw ECG for subject-invariant CNN representation learning because noisy information irrelevant to human beat has been largely filtered out. To strengthen the generalization ability of CNN to discover subject-invariant pattern in ECG, skip connection is exploited to propagate information well in neural network and channel attention is designed to adaptively highlight informative channel-wise features. Experiments show that the proposed DBWT-AFNet outperforms the state-of- the-art methods, especially for ECG segments classification across different subjects, where no data from testing subjects have been used in training.

Traditional pipeline for the task of detecting phenotypic biomarkers is a two-stage implementation, i.e., differentially expressed candidates are identified by NT tests, and then a subset of the candidates are further detected by phenotype-targeted tests (PT test) for significant phenotypic features, where N is short for Normal data and T is for Treatment/Trouble data. Such a two-stage procedure has low detection power as they do not make full use of the information contained in the (T, N). In this paper, we apply the two-variate PT test which jointly considers tumor-adjacent data and tumor data for improving the detection power. We investigate its performance by experiments on real-world datasets of breast cancer, considering phenotypes including BMI, overall survival time, pathologic stage, and tumor size. The results show that the method has high detection power and is more reliable, and the tumor-adjacent normal data plays an important role in the detection of phenotypic biomarkers. Finally, we obtain a new finding that the gene TCTEX1D2 is significantly related to tumor size in breast cancer.

Recently, one earliest skip connected networks named Lmser was revisited and its convolutional layer based version named CLmser was proposed. This paper studies CLmser for segmentation (shortly CLmser-S) of Electron Microscopy (EM) images and also one further development. First, we experimentally show that CLmser-S outperforms the popular U-Net and save many free parameters. Second, we combine one newest formulation named Flexible Lmser (F-Lmser) and CLmser-S into a version called F-CLmser-S, together with learned masks replacing the similarity based one used in F-Lmser for implementing fast-lane skip connections. Experimental results on the ISBI 2012 EM dataset show that F-CLmser-S improves CLmser and achieves competitive performance with state-of-the-art results.

We present a GAN-Lmser network for the problem of transforming an image from one domain A to another B. The proposed network is based on CNN-Lmser, a recent further extension to deep convolutional layers from least mean square error reconstruction (Lmser) network, which was originally proposed in 1991. Specifically, in GAN-Lmser, the two directions, A-to-B and B-to-A, share the same architecture and symmetrically the same weights, by following the duality in bidirectional architecture (DBA) and duality connection weights (DCW) of Lmser, and an adversarial loss from GAN(generative adversarial network) was added to Lmser. Compared with the famous image-to-image translation model CycleGAN, the GAN-Lmser is compact with a significantly reduced number of parameters and is able to transfer learning through weight sharing between the two directions. Experiments demonstrate that GAN-Lmser is at least comparable to CycleGAN in benchmark datasets, and is robust when the training sample size is small.

Existing single image super-resolution (SISR) methods usually focus on Low-Resolution (LR) images which are artificially generated from High-Resolution (HR) images by a down-sampling process, but are not robust for unmatched training set and testing set. This paper proposes a GAN Flexible Lmser (GFLmser) network that bidirectionally learns the High-to-Low (H2L) process that degrades HR images to LR images and the Low-to-High (L2H) process that recovers the LR images back to HR images. The two directions share the same architecture, added with the gated skip connections from the H2L-net to the L2H-net in order to enhance information transferring for super-resolution. In comparison with several related state-of-the-art methods, experiments demonstrate that not only GFLmser is the most robust method on images of unmatched training set and testing set, but also its performance on real-world face LR images is best in PSNR and reasonably good in FID.

Predicting drug side effects is a critical task in the drug discovery, which attracts great attentions in both academy and industry. Although lots of machine learning methods have been proposed, great challenges arise with boom of precision medicine. On one hand, many methods are based on the assumption that similar drugs may share same side effects, but measuring the drug-drug similarity appropriately is challenging. One the other hand, multi-source data provide diverse information for the analysis of side effects, and should be integrated for the high-accuracy prediction. In this paper, we tackle the side effect prediction problem through linear neighborhoods and multi-source data integration. In the feature space, linear neighborhoods are constructed to extract the drug-drug similarity, namely “linear neighborhood similarity”. By transferring the similarity into the side effect space, known side effect information is propagated through the similarity-based graph. Thus, we propose the linear neighborhood similarity method (LNSM), which utilizes single-source data for the side effect prediction. Further, we extend LNSM to deal with multi-source data, and propose two data integration methods: similarity matrix integration method (LNSM-SMI) and cost minimization integration method (LNSM-CMI), which integrate drug substructure data, drug target data, drug transporter data, drug enzyme data, drug pathway data and drug indication data to improve the prediction accuracy. The proposed methods are evaluated on the benchmark datasets. The linear neighborhood similarity method (LNSM) can produce satisfying results on the single-source data. Data integration methods (LNSM-SMI and LNSM-CMI) can effectively integrate multi-source data, and outperform other state-of-the-art side effect prediction methods in the cross validation and independent test. The proposed methods are promising for the drug side effect prediction.

Transcription factor activities (TFAs), rather than expression levels, control gene expression and provide valuable information for investigating TF-gene regulations. Network Component Analysis (NCA) is a model based method to deduce TFAs and TF-gene control strengths from microarray data and a priori TF-gene connectivity data. We modify NCA to model gene expression regulation by non-Gaussian Factor Analysis (NFA), which assumes TFAs independently comes from Gaussian mixture densities. We properly incorporate a priori connectivity and/or sparsity on the mixing matrix of NFA, and derive, under Bayesian Ying-Yang (BYY) learning framework, a BYY-NFA algorithm that can not only uncover the latent TFA profile similar to NCA, but also is capable of automatically shutting off unnecessary connections. Simulation study demonstrates the effectiveness of BYY-NFA, and a preliminary application to two real world data sets shows that BYY-NFA improves NCA for the case when TF-gene connectivity is not available or not reliable, and may provide a preliminary set of candidate TF-gene interactions or double check unreliable connections for experimental verification.

We propose a clustering algorithm based on a structural prior based Local Factor Analysis (spLFA) model under the Bayesian Ying-Yang harmony learning, which automatically determines the hidden dimensionalities during parameter learning, reduces the number of free parameters by projecting the mean vectors onto a low dimensional manifold, imposes the sparseness by a Normal-Jeffreys prior. Experiments on the diagnostic research dataset show that BYY-spLFA outperforms the k-means clustering and single-link hierarchical clustering. The experiments on a lymphoma cancer datset further indicate the BYY-spLFA is able to uncover the number of phenotypes correctly and cluster the phenotypes more accurately. In addition, we modify BYY-spLFA to implement supervised learning and preliminarily demonstrate its effectiveness on a Leukemia data for classification.

Addressing the problem of detecting the number of source signals as selecting the hidden dimensionality of Factor Analysis (FA) model, we investigate several model selection criteria via a new empirical analyzing tool that examines the joint effect of signal-noise ratio (SNR) and sample size N on the model selection performance. The contours of the model selection accuracies visualize a three-region partition on the space of SNR andN, and a diminishing marginal effect which trades off SNR and N on the performance. Moreover, the newly derived Variational Bayes algorithm and three variants of Bayesian Ying-Yang (BYY) algorithms are more robust against reducing SNR and N, where the BYY with priors' hyperparameters updated is the best in general.