Molecular representation, as one of the fundamental challenges in artificial intelligence-driven drug discovery, has attracted increasing attention due to its low cost and impressive speed while it is applied in molecular property prediction, drug molecule generation, drug-drug interactions, etc. Numerous models that integrate multi-modal representations have been proposed for molecular representation learning. However, existing methods have not yet considered the IUPAC International Chemical Identifier (InChI) as one of the multi-modal inputs. To address this issue, we propose InChINet, a self-supervised molecular representation learning framework that is pre-trained on 10 million unlabeled molecules. It leverages mutual information across the simplified molecular line input system (SMILES) and InChI. In addition, we present token reordering and token masking for SMILES. Combined with SMILES enumeration, these three strategies introduce domain knowledge and improve the model's stability against syntactic variations in SMILES representations. Benefiting from the introduction of InChI and augmentation strategies, InChINet achieves impressive performance on a wide range of downstream tasks, including molecular property prediction, drug-drug interaction (DDI) prediction, clustering analysis, zero-shot cross-lingual retrieval, and ablation study.

Knowledge Tracing (KT) uses students’ interaction history on the platform to assess students’ mastery of knowledge at the current moment, but often requires experts to label skills. Therefore, we propose a Q-matrix-based Graph Attention Knowledge Tracing to achieve self-labelling of skills. The model uses the graph attention neural network to aggregate the information of past interactions, and realizes the accurate modeling of students’ knowledge state. In addition, it can learn the Q-matrix without expert labeling, enabling self-labelling of skills. Experiments on ASSIST2009 show our model improves AUC by nearly 3.6

As the rapid development of Intelligent Tutoring Systems (ITS) in the past decade, tracing the students' knowledge state has become more and more important in order to provide individualized learning guidance. This is the main idea of Knowledge Tracing (KT), which models students' mastery of knowledge concepts (KCs, skills needed to solve a question) based on their past interactions on platforms. Plenty of KT models have been proposed and have shown remarkable performance recently. However, the majority of these models use concepts to index questions, which means the predefined skill tags for each question are required in advance to indicate the KCs needed to answer that question correctly. This makes it pretty hard to apply on large-scale online education platforms where questions are often not well-organized by skill tags. In this paper, we propose Q-matrix-based Attentive Knowledge Tracing (QAKT), an end-to-end style model that is able to apply the attentive method to scenes where no predefined skill tags are available without sacrificing its performance. With a novel hybrid embedding method based on the q-matrix and Rasch model, QAKT is capable of modeling problems hierarchically and learning the q-matrix efficiently based on students' sequences. Meanwhile, the architecture of QAKT ensures that it is friendly to questions associated with multiple skills and has outstanding interpretability. After conducting experiments on a variety of open datasets, we empirically validated that our model shows similar or even better performance than state-of-the-art KT methods. Results of further experiments suggest that the q-matrix learned by QAKT is highly model-agnostic and more information-sufficient than the one labeled by human experts, which could help with the data mining tasks in existing ITSs.

Gloss-free sign language translation is a key focus in sign language translation research, enabling effective communication between the deaf and the hearing individuals in a broader and more universal manner. In this work, we propose a Progressive Multi-Level Learning model for sign language translation (PML-SLT), which progressively learns sign representations to improve video understanding. Rather than requiring every frame to attend to all other frames during attention computation, our approach introduces a progressive perceptual field expansion mechanism that gradually broadens the attention scope across video frames. This mechanism continuously expands the perceptual field between frames, effectively capturing both local and global information. Besides, to fully exploit multi-granularity information, we employ a multi-level feature integration scheme that transfers the output of each encoder layer to the corresponding decoder layer, enabling comprehensive utilization of hierarchical temporal features. Additionally, we introduce a multi-modal triplet loss to harmonize semantic information across modalities, aligning the text space with the video space so that the video features acquire richer semantic meaning. Experimental results on two public datasets demonstrate the promising translation performance of the proposed PML-SLT model.

Facial expression recognition (FER) remains a challenging task in computer vision. Recent works have shown excellent performance in overall recognition accuracy, but its accuracy significantly decreases when recognizing similar expressions. This is due to interclass homogeneity and intraclass heterogeneity. To address these issues, we propose a novel dual-stage network called DUAL, inspired by contrastive learning. First, we increase the distance between negative samples while reducing the distance between positive ones. This is achieved by dynamically updating pairs of comparison samples. Second, we introduce a two-stage network architecture. The first stage uses two branches to extract image features and facial keypoint features. These branches interact to learn coarse-grained features through mutual guidance. The second stage focuses on fine-grained features using scale-specific residual blocks. This allows the model to identify facial regions that are critical for recognizing expressions. We conducted extensive experiments on multiple datasets. The results show that DUAL surpasses state-of-the-art models in items of performance. Additionally, the model shows high accuracy even in noisy conditions, highlighting its robustness.

In the treatment of complex diseases, patients often need to use multiple types of drugs at the same time. However, drug-drug interactions (DDIs) often lead to side effects and aggravate disease symptoms. Therefore, accurate prediction of drug interactions is essential to reduce medical risks. Currently, artificial intelligence methods have received increasing attention in exploring potential DDIs. However, the existing techniques usually only extract the features inside the drug molecule or in the DDI network, which is difficult to obtain the necessary information comprehensively. To solve this problem, a drug-drug interaction hierarchical graph neural network model called PHGL-DDI is proposed, which considers both the graph-level information of drugs and the network-level information of DDI. Specifically, each node in the DDI network view represents a drug, and the connections between nodes represent the interactions between drugs. Meanwhile, in the drug graph-level, the self-supervised contrast learning method is employed for pre-training the graph neural network, so as to extract the features inside drugs, thereby building a powerful DDI understanding system. Compared with the link prediction method that only extracts DDI network features, PHGL-DDI not only uses the topological structure information between nodes, but also considers the semantic features of entity nodes. Extensive experimental studies show that PHGL-DDI significantly improves the prediction effect of DDI compared with the most advanced methods, and has better generalization ability

Sign language translation converts sign language videos into spoken language text. Traditional gloss-based approaches require costly gloss annotations, driving recent interest in gloss-free methods. In this paper, we propose a novel gloss-free sign language translation method based on fusion attention (SLTFA) that uniquely models the intrinsic logical structure of sign language. The key innovation is the development of a dual-attention mechanism that mimics the natural hierarchical structure of sign language: intra-gloss attention captures fine-grained relationships within video frame sequences representing individual semantic units, while inter-gloss attention models the broader contextual connections between these units, similar to how words form coherent sentences. Additionally, we introduce a contrastive loss strategy for cross-modal soft alignment that effectively bridges the gap between visual and textual representations. Extensive experiments on the RWTH-PHOENIX-WEATHER-2014T dataset demonstrate SLTFA's superior performance, achieving a BLEU-4 score of 16.99 and a ROUGE score of 40.82. On the CSL-Daily dataset, our approach achieves a BLEU-1 score of 25.56 and a ROUGE score of 27.51, demonstrating strong performance across different sign languages.

The rise of Neural Machine Translation (NMT) models opens doors for translating Chinese text into Braille, improving information access for visually impaired individuals. However, current NMT models, often based on encoder-decoder architectures, utilize sequential rather than parallel processing in the decoder. This autoregressive decoding hinders architectures like the Transformer from fully leveraging their training speed advantages during inference. While the Transformer excels in parallel training, its inference time complexity remains O ( T 2 ), where T represents sequence length. This bottleneck becomes particularly significant when translating Braille, known for its long character sequences. We propose a non-autoregressive Chinese-to- Braille translation model that solely employs the encoder architecture along with Connectionist Temporal Classification (CTC) loss to generate complete Braille sequences simultaneously. This approach significantly improves inference speed, achieving a substantial acceleration compared to autoregressive models during inference with a time complexity of O (1). Remarkably, alongside increased inference speed, translation accuracy also improves. By incorporating a pre-training technique, our method achieves a remarkable BLEU Score of 95.10% with a limited dataset of only 2k Chinese-Braille training pairs.

Deep generative models provide a powerful solution for the de novo design of molecules. However, the majority of existing methods only generate molecules for a single target. Generating molecules with biological activities against multiple specific targets and desired properties remains an extremely difficult challenge. In this study, we propose a novel 3D molecule generation framework based on reinforcement learning and diffusion model to generate molecules with predefined properties for given multiple targets. The proposed framework, MDRL, uses a diffusion model to understand the 3D chemical structure of molecules and employs Kolmogorov-Arnold Networks instead of Multilayer Perceptron to enhance model performance. Through reinforcement learning, the framework is able to generate molecules that simultaneously target two targets and further optimizes multiple molecular properties. Experimental results show that our model exhibits comparable performance to various state-of-the-art molecular generation models, and MDRL can effectively navigate chemical space to design polypharmacological compounds and control multiple molecular properties. In multiple case studies, we verify that the generated molecules can simultaneously target two targets through molecular docking and assess the model's ability to control multiple molecular properties. The results in this study highlight the advantages and practicalities of our model in generating polypharmacological compounds with desired properties.

Porosity is a key parameter for evaluating reservoir performance, but high-precision prediction is highly challenging in complex shale reservoirs due to the strong heterogeneity of the formation and the highly nonlinear relationship between logging parameters and porosity. Traditional prediction methods based on experience or physical models often have low generalization ability and accuracy. This study proposes a hybrid model (MABC-LSSVM) that combines a modified artificial bee colony (MABC) optimization algorithm with a least squares support vector machine (LSSVM) model. Inertia weights and acceleration coefficients are utilized to change the hyperparameters of the optimization model to achieve high-precision prediction of shale reservoir porosity using data-driven methods. The model inputs include compensating neutron log (CNL), density log (DEN), photoelectric absorption cross-section index (PE), and gamma ray log (GR) parameters. The proposed model is compared with the LSSVM, gradient boosting decision tree (GBDT), and ABC-LSSVM. The results show that the MABC-LSSVM model exhibits the best predictive performance. Its prediction results are highly consistent with the true porosity curve. The coefficient of determination (R2) is 0.93, significantly higher than for all comparison models. The findings demonstrate the effectiveness of combining an intelligent optimization algorithm with the LSSVM model. This approach is reliable for predicting the porosity in complex formations and performing reservoir evaluations in oil and gas exploration and development.

The complexity of clouds, particularly in terms of texture detail at high resolutions, has not been well explored by most existing cloud detection networks. We introduce the high-resolution cloud detection network (HR-cloud-Net), which utilizes a hierarchical high-resolution integration approach. HR-cloud-Net integrates a high-resolution representation module, layer-wise cascaded feature fusion module, and multiresolution pyramid pooling module to effectively capture complex cloud features. This architecture preserves detailed cloud texture information while facilitating feature exchange across different resolutions, thereby enhancing the overall performance in cloud detection. Additionally, an approach is introduced wherein a student view, trained on noisy augmented images, is supervised by a teacher view processing normal images. This setup enables the student to learn from cleaner supervisions provided by the teacher, leading to an improved performance. Extensive evaluations on three optical satellite image cloud detection datasets validate the superior performance of HR-cloud-Net compared with existing methods. (c) 2024 SPIE and IS&T

The education of visually impaired children remains a focal topic, and the implementation of Braille–Chinese translation can facilitate improved understanding and learning of Braille for these children. Braille–Chinese translation refers to the conversion of Braille text into Chinese characters. Owing to the scarcity of Braille–Chinese parallel corpora, achieving Braille–Chinese translation with limited data remains a significant research challenge. This study tackles the challenge by utilizing both Braille joint training and pre-trained language models to conduct analogy learning and transfer learning on a substantial volume of readily available data. The method of Braille joint training also establishes a universal language model in the Braille domain, capable of simultaneously performing multiple tasks like Braille–Chinese translation, Chinese–Braille translation, Braille word segmentation, etc. This greatly caters to users' requirements for Braille information processing while also mitigating extensive deployment expenses associated with multiple models. Furthermore, we put forward a Braille phoneme recognition translation approach distinct from conventional Braille–Chinese translation methodologies. This approach aims to attain precise translation by identifying the Chinese characters corresponding to each Braille phoneme. This method achieves a BLEU score of 96.22 on the Braille–Chinese entity annotation dataset and exhibits high translation efficiency.

Online education systems have gained increasing popularity due to their capability to fully preserve users’ learning data. This advantage enables researchers to assess learners’ mastery through their learning trajectories, thereby facilitating personalized education and support. Knowledge tracing, an effective educational aid, simulates students’ implicit knowledge states and predicts their mastery over knowledge based on their historical answer records. However, for newly developed online learning platforms, the lack of sufficient historical answer data may impede accurate prediction of students’ knowledge states, rendering existing knowledge tracing models less effective. This paper introduces the first pre-trained knowledge tracing model that leverages a substantial amount of existing data for pre-training and a smaller dataset for fine-tuning. Validated across several publicly available knowledge tracing datasets, our method demonstrates significant improvement in tracing performance on small datasets, with a maximum AUC increase of 5.07

Traditional medical research is characterised by lengthy duration, significant financial investment, and substantial risk of failure. In response to these challenges, network medicine, combined with medicine and computer technology, has become an important development direction, and computational methods have been proposed to predict potential associations. However, most of the current computational methods focus on single-potential association prediction tasks, which face issues of association sparsity and weak generalisation ability. To address these challenges, we developed a heterogeneous biological network multi-task learning model (HBNMM). Unlike previous methods based on bipartite graphs, HBNMM constructs a complex heterogeneous biological network, including ncRNA-disease-drug association networks and diverse similarity networks. HBNMM applies graph attention networks to aggregate node neighbourhood information and acquire node feature embeddings, and is then trained with a multi-task learning strategy to simultaneously predict potential ncRNA-disease, ncRNA-drug, and drug-disease associations. As a result, the HBNMM achieves an excellent performance that is higher than that of the state-of-the-art models. Furthermore, five case studies supported by experiments showed powerful predictive ability for drug discovery and disease treatment.

The ordered assembly of Tau protein into filaments characterizes Alzheimer's and other neurodegenerative diseases, and thus, stabilization of Tau protein is a promising avenue for tauopathies therapy. To dissect the underlying aggregation mechanisms on Tau, we employ a set of molecular simulations and the Markov state model to determine the kinetics of ensemble of K18. K18 is the microtubule-binding domain of Tau protein and plays a vital role in the microtubule assembly, recycling processes, and amyloid fibril formation. Here, we efficiently explore the conformation of K18 with about 150 mu s lifetimes in silico. Our results observe that all four repeat regions (R1-R4) are very dynamic, featuring frequent conformational conversion and lacking stable conformations, and the R2 region is more flexible than the R1, R3, and R4 regions. Additionally, it is worth noting that residues 300-310 in R2-R3 and residues 319-336 in R3 tend to form sheet structures, indicating that K18 has a broader functional role than individual repeat monomers. Finally, the simulations combined with Markov state models and deep learning reveal 5 key conformational states along the transition pathway and provide the information on the microsecond time scale interstate transition rates. Overall, this study offers significant insights into the molecular mechanism of Tau pathological aggregation and develops novel strategies for both securing tauopathies and advancing drug discovery.

Visually assistive systems for the visually impaired play a pivotal role in enhancing the quality of life for the visually impaired. Assistive technologies for the visually impaired have undergone a remarkable transformation with the advent of deep learning and sophisticated assistive devices. In particular, the paper utilizes the latest machine translation models and techniques to accomplish the Chinese–Braille translation task, providing convenience for visually impaired individuals. The Traditional end-to-end Chinese–Braille translation approach incorporates Braille dots and Braille word segmentation symbols as tokens within the model’s vocabulary. However, our findings reveal that Braille word segmentation is significantly more complex than Braille dot prediction. The paper proposes a novel Two-Part Loss (TPL) method that treats these tasks distinctly, leading to significant accuracy improvements. To enhance translation performance further, we introduce a BERT-Enhanced Segmentation Transformer (BEST) method. BEST leverages knowledge distillation techniques to transfer knowledge from a pre-trained BERT model to the translate model, mitigating its limitations in word segmentation. Additionally, soft label distillation is employed to improve overall efficacy further. The TPL approach achieves an average BLEU score improvement of 1.16 and 5.42 for Transformer and GPT models on four datasets, respectively. In addition, The work presents a two-stage deep learning-based translation approach that outperforms traditional multi-step and end-to-end methods. The proposed two-stage translation method achieves an average BLEU score improvement of 0.85 across four datasets.

This paper introduces ALFKT, a novel knowledge tracing model that combines attention mechanisms and learning features to create a detailed representation of students’ learning trajectories. ALFKT addresses these limitations by synergistically integrating attention mechanisms and recurrent neural networks to capture time series data efficiently. Learning gates are introduced to regulate fluctuations in the student’s knowledge state between interactions, accurately modeling the uptake rate of learning. Individual differences are accounted for by incorporating various learning characteristics, using answer accuracy to measure question difficulty and interval time to measure forgetting. These components, along with the student’s present knowledge state, enable accurate prediction of answer accuracy in subsequent interactions. Extensive experiments on four public datasets validate the superiority of ALFKT, with a remarkable 12.69

Innovations in drug-target interactions (DTIs) prediction accelerate the progression of drug development. The introduction of deep learning models has a dramatic impact on DTIs prediction, with a distinct influence on saving time and money in drug discovery. This study develops an end-to-end deep collaborative learning model for DTIs prediction, called EDC-DTI, to identify new targets for existing drugs based on multiple drug-target-related information including homogeneous information and heterogeneous information by the way of deep learning. Our end-to-end model is composed of a feature builder and a classifier. Feature builder consists of two collaborative feature construction algorithms that extract the molecular properties and the topology property of networks, and the classifier consists of a feature encoder and a feature decoder which are designed for feature integration and DTIs prediction, respectively. The feature encoder, mainly based on the improved graph attention network, incorporates heterogeneous information into drug features and target features separately. The feature decoder is composed of multiple neural networks for predictions. Compared with six popular baseline models, EDC-DTI achieves highest predictive performance in the case of low computational costs. Robustness tests demonstrate that EDC-DTI is able to maintain strong predictive performance on sparse datasets. As well, we use the model to predict the most likely targets to interact with Simvastatin (DB00641), Nifedipine (DB01115) and Afatinib (DB08916) as examples. Results show that most of the predictions can be confirmed by literature with clear evidence.

The technology for converting Chinese to Braille is of great importance. When paired with a Braille display, it can better meet the educational and daily needs of the visually impaired community, especially children and students. Incorporating visual assistance mechanisms can further enhance the user experience and provide comprehensive support for individuals with visual impairments. In recent years, the use of end-to-end neural machine translation models for Chinese–Braille translation has gained traction. However, this task requires large, high-quality, and domain-specific parallel data to train robust models. Unfortunately, the existing Chinese–Braille parallel data is insufficient to achieve satisfactory results. To address this challenge, this paper puts forward a groundbreaking approach that integrates pre-training models into the Chinese Braille translation task. This represents the first-ever application of such technology in this context and it is different from traditional pre-training methods. While previous pre-training method of natural language processing mainly utilized raw text data, we have identified its limitations in improving Chinese–Braille translation. Therefore, we have proposed three novel forms of pre-training datasets, instead of relying solely on raw text data. By utilizing the Transformer model, our approach achieves the highest BLEU score of 94.53 on a 10k parallel corpus, presenting a new direction for Chinese–Braille translation research. Furthermore, we introduce a new form of data that enables Chinese–Braille translation solely using the encoder framework. Leveraging the MacBERT model, this approach achieves a BLEU score of 98.87 on the test set and demonstrates an inference speed 54 times faster than the Transformer model. These findings have significant implications for the field of Chinese–Braille translation, providing insights for future research endeavors.