Small-molecule drug design aims to generate compounds that target specific proteins, playing a crucial role in the early stages of drug discovery. Recently, research has emerged that utilizes the GPT model, which has achieved significant success in various fields to generate molecular compounds. However, due to the persistent challenge of small datasets in the pharmaceutical field, there has been some degradation in the performance of generating target-specific compounds. To address this issue, we propose an enhanced target-specific drug generation model, Adapt-cMolGPT, which modifies molecular representation and optimizes the fine-tuning process. In particular, we introduce a new fine-tuning method that incorporates an adapter module into a pre-trained base model and alternates weight updates by sections. We evaluated the proposed model through multiple experiments and demonstrated performance improvements compared to previous models. In the experimental results, Adapt-cMolGPT generated a greater number of novel and valid compounds compared to other models, with these generated compounds exhibiting properties similar to those of real molecular data. These results indicate that our proposed method is highly effective in designing drugs targeting specific proteins.

We propose a novel power control technique called PC-ECGNN, which uses edge convolution to optimize power allocation in wireless IoT networks. PC-ECGNN leverages interference link distances as edge features and desired link channel gains as initial vertex features, iteratively updating vertex features based on neighbors and edge features. PC-ECGNN is the first technique to incorporate edge convolution into power control and has been customized for the considered scenario, optimizing the neural network structure to provide fast convergence and high performance simultaneously. Experimental results show that PC-ECGNN outperformed the state-of-the-art PC-MPGNN, achieving a 4% increase in average spectral efficiency and a 4dBm reduction in average transmit power compared to PC-MPGNN. Furthermore, our technique demonstrates advantages over existing methods in dynamic environmental changes. The proposed model, trained in a fixed environment, showed minimal performance degradation across various test environments different from the training setting, outperforming traditional models trained in individual environments. When applying meta-learning, the proposed model achieved better performance in each test environment after additional fine-tuning with only 1% of the pre-training epochs, compared to models trained with the full number of epochs in each individual test environment.

Drug-phospholipid complexing is a promising formulation technology for improving the low bioavailability of active pharmaceutical ingredients (APIs). However, identifying whether phospholipid and candidate drug can form a complex through in vitro tests can be costly and time-consuming due to the physicochemical properties and experimental environment. In a previous study, the authors developed seven machine learning models to predict drug-phospholipid complex formation, and the lightGBM model demonstrated the best performance. However, the previous study was unable to sufficiently address the degradation of test performance caused by the small size of the training data with class imbalance, and it had the limitation of considering only machine learning techniques. To overcome these limitations, we propose a new deep learning-based prediction model that employs variational autoencoder (VAE) and principal component analysis (PCA) techniques to improve prediction performance. The model uses a multi-layer one-dimensional convolutional neural network (CNN) with a skip connection to effectively capture the complex relationship between drugs and lipid molecules. The computer simulation results demonstrate that our proposed model performs better than the previous model in all performance metrics.

In recent research on the formulation prediction of oral dissolving drugs, deep learning models with significantly improved performance compared to machine learning models were proposed. However, the performance degradation due to limitations of an imbalanced dataset with a small size and inefficient neural network structure has still not been resolved. Therefore, we propose new deep learning-based prediction models that maximize the prediction performance for disintegration time of oral fast disintegrating films (OFDF) and cumulative dissolution profiles of sustained-release matrix tablets (SRMT). In the case of OFDF, we use principal component analysis (PCA) to reduce the dimensionality of the dataset, thereby improving the prediction performance and reducing the training time. In the case of SRMT, the Wasserstein generative adversarial network (WGAN), a neural network-based generative model, is used to overcome the limitation of performance improvement due to the lack of experimental data. To the best of our knowledge, this is the first work that utilizes WGAN for pharmaceutical formulation prediction. Experimental results show that the proposed methods have superior performance than the existing methods for all the performance metrics considered.

Small-molecule drug design aims to generate compounds that target specific proteins, playing a crucial role in the early stages of drug discovery. Recently, research has emerged that utilizes the GPT model, which has achieved significant success in various fields to generate molecular compounds. However, due to the persistent challenge of small datasets in the pharmaceutical field, there has been some degradation in the performance of generating target-specific compounds. To address this issue, we propose an enhanced target-specific drug generation model, Adapt-cMolGPT, which modifies molecular representation and optimizes the fine-tuning process. In particular, we introduce a new fine-tuning method that incorporates an adapter module into a pre-trained base model and alternates weight updates by sections. We evaluated the proposed model through multiple experiments and demonstrated performance improvements compared to previous models. In the experimental results, Adapt-cMolGPT generated a greater number of novel and valid compounds compared to other models, with these generated compounds exhibiting properties similar to those of real molecular data. These results indicate that our proposed method is highly effective in designing drugs targeting specific proteins.

We propose a novel power control technique called PC-ECGNN, which uses edge convolution to optimize power allocation in wireless IoT networks. PC-ECGNN leverages interference link distances as edge features and desired link channel gains as initial vertex features, iteratively updating vertex features based on neighbors and edge features. PC-ECGNN is the first technique to incorporate edge convolution into power control and has been customized for the considered scenario, optimizing the neural network structure to provide fast convergence and high performance simultaneously. Experimental results show that PC-ECGNN outperformed the state-of-the-art PC-MPGNN, achieving a 4% increase in average spectral efficiency and a 4dBm reduction in average transmit power compared to PC-MPGNN. Furthermore, our technique demonstrates advantages over existing methods in dynamic environmental changes. The proposed model, trained in a fixed environment, showed minimal performance degradation across various test environments different from the training setting, outperforming traditional models trained in individual environments. When applying meta-learning, the proposed model achieved better performance in each test environment after additional fine-tuning with only 1% of the pre-training epochs, compared to models trained with the full number of epochs in each individual test environment.

Drug-phospholipid complexing is a promising formulation technology for improving the low bioavailability of active pharmaceutical ingredients (APIs). However, identifying whether phospholipid and candidate drug can form a complex through in vitro tests can be costly and time-consuming due to the physicochemical properties and experimental environment. In a previous study, the authors developed seven machine learning models to predict drug-phospholipid complex formation, and the lightGBM model demonstrated the best performance. However, the previous study was unable to sufficiently address the degradation of test performance caused by the small size of the training data with class imbalance, and it had the limitation of considering only machine learning techniques. To overcome these limitations, we propose a new deep learning-based prediction model that employs variational autoencoder (VAE) and principal component analysis (PCA) techniques to improve prediction performance. The model uses a multi-layer one-dimensional convolutional neural network (CNN) with a skip connection to effectively capture the complex relationship between drugs and lipid molecules. The computer simulation results demonstrate that our proposed model performs better than the previous model in all performance metrics.

In recent research on the formulation prediction of oral dissolving drugs, deep learning models with significantly improved performance compared to machine learning models were proposed. However, the performance degradation due to limitations of an imbalanced dataset with a small size and inefficient neural network structure has still not been resolved. Therefore, we propose new deep learning-based prediction models that maximize the prediction performance for disintegration time of oral fast disintegrating films (OFDF) and cumulative dissolution profiles of sustained-release matrix tablets (SRMT). In the case of OFDF, we use principal component analysis (PCA) to reduce the dimensionality of the dataset, thereby improving the prediction performance and reducing the training time. In the case of SRMT, the Wasserstein generative adversarial network (WGAN), a neural network-based generative model, is used to overcome the limitation of performance improvement due to the lack of experimental data. To the best of our knowledge, this is the first work that utilizes WGAN for pharmaceutical formulation prediction. Experimental results show that the proposed methods have superior performance than the existing methods for all the performance metrics considered.

This research is aimed at predicting the physical stability for amorphous solid dispersion by utilizing deep learning methods. We propose a prediction model that effectively learns from a small dataset that is imbalanced in terms of class. In order to overcome the imbalance problem, our model performs a hybrid sampling which combines synthetic minority oversampling technique (SMOTE) algorithm with edited nearest neighbor (ENN) algorithm and reduces the dimensionality of the dataset using principal component analysis (PCA) algorithm during data preprocessing. After the preprocessing, it performs the learning process using a carefully designed neural network of simple but effective structure. Experimental results show that the proposed model has faster training convergence speed and better test performance compared to the existing DNN model. Furthermore, it significantly reduces the computational complexity of both training and test processes.

Abstract Millimeter wave (mmWave) communication systems use beamforming and large antenna arrays to achieve high data rates by directing signals through narrow beams, reducing interference and enhancing transmission efficiency. Efficient beamforming requires real‐time beam adjustments to adapt to user positions and environmental changes, but traditional methods relying on frequent measurements can lead to significant overhead in dynamic environments. AI/ML approaches leveraging sensor data and historical information can improve beam prediction and tracking efficiency. Building on this, we propose a multimodal beam tracking model for UAV communication, integrating image and GPS data to predict UAV movement and optimize beam tracking. The model employs ResNet‐SE blocks for feature extraction, CAformer blocks for multimodal data fusion, and LSTM for capturing sequential historical features. Experimental results show that the proposed model outperforms single‐modal methods, achieving a 24.4% improvement in Top‐1 beam accuracy.

ABSTRACT This paper proposes DeepSC‐S‐MECA, a semantic communication model for improving speech semantic reconstruction. The model combines a Mamba Transformer (MambaTrans) block, which integrates a State Space Model (SSM) into a Transformer to capture temporal and contextual dependencies in sequential signals, and an ECA‐ResNet block, which applies efficient channel attention (ECA) to emphasize informative features. Experimental results under AWGN and Rician channels show that DeepSC‐S‐MECA achieves higher SDR and PESQ than DeepSC‐S‐CBAMR and DeepSC‐S‐ECAR, providing robust speech reconstruction while reducing parameters and FLOPs by about 39% compared with DeepSC‐S‐CBAMR.

Abstract Accurate feedback of downlink channel state information (CSI) plays a vital role in enabling beamforming in frequency‐division duplex, massive multiple‐input multiple‐output (MIMO) systems. However, the feedback overhead scales with the number of antennas, severely impacting uplink efficiency. While transformer‐based methods, such as TransNet, have exhibited high‐reconstruction accuracy, their computational complexity limits the practical deployment on user equipment. STNet reduces this complexity via spatially separable attention but still underperforms TransNet in terms of accuracy. In this letter, we propose SPINNet, a lightweight CSI feedback network that combines spectral and inception modules to extract efficiently frequency‐domain features. Experimental results on the COST2100 dataset show that the proposed SPINNet outperforms STNet by only half its computational cost and surpasses TransNet in certain scenarios.

This paper proposes semantic interference cancellation (SemantIC)-Mamba, a novel framework that enhances semantic reconstruction accuracy while stably maintaining channel decoding performance. The proposed model follows a turbo structure in which information is complementarily exchanged between the semantic domain and the signal domain, progressively refining the reconstruction quality through iterative processing. To effectively perform this iterative information refinement, the proposed framework adopts a semantic autoencoder composed of three key components: a Conv block that extracts local features, a Mamba block that efficiently models long-range dependencies to integrate global contextual information, and an UpConv block that restores low-resolution features to the original resolution. Experimental results demonstrate that SemantIC-Mamba consistently achieves improved PSNR and classification accuracy compared to conventional SemantIC and SemantIC++ while maintaining channel decoding performance at a level comparable to existing models.