Front Oncol. 2025 Jan 24;15:1516264. doi: 10.3389/fonc.2025.1516264. eCollection 2025.

ABSTRACT

The integrated application of artificial intelligence (AI) and digital pathology (DP) technology has opened new avenues for advancements in oncology and molecular pathology. Consequently, studies in renal cell carcinoma (RCC) have emerged, highlighting potential in histological subtype classification, molecular aberration identification, and outcome prediction by extracting high-throughput features. However, reviews of these studies are still rare. To address this gap, we conducted a thorough literature review on DP and AI applications in RCC through database searches. Notably, we found that AI models based on deep learning achieved area under the curve (AUC) of over 0.93 in subtype classification, 0.89-0.96 in grading of clear cell RCC, 0.70-0,89 in molecular prediction, and over 0.78 in survival prediction. This review finally discussed the current state of researches and potential future directions.

PMID:39926279 | PMC:PMC11802434 | DOI:10.3389/fonc.2025.1516264

RSC Adv. 2025 Feb 7;15(6):4262-4273. doi: 10.1039/d4ra07951e. eCollection 2025 Feb 6.

ABSTRACT

Virtual screening (VS) methodologies have become key in the drug discovery process but are also applicable to other fields including catalysis, material design, and, more recently, insecticide solutions. Indeed, the search for effective pest control agents is a critical industrial objective, driven by the need to meet stringent regulations and address public health concerns. Cockroaches, known vectors of numerous diseases, represent a major challenge due to the toxicity of existing control measures to humans. In this article, we leverage an Artificial Intelligence (AI)-based screening of the Drug Bank (DB) database to identify novel acetylcholinesterase (AChE) inhibitors, a previously uncharacterized target in the American cockroach (Periplaneta americana). Our AI-based VS pipeline starts with the deep-learning-based AlphaFold to predict the previously unknown 3D structure of AChE based on its amino acid sequence. This first step enables the subsequent ligand-receptor VS of potential inhibitors, the development of which is performed using a consensus VS protocol based on two different tools: Glide, an industry-leading solution, and METADOCK 2, a metaheuristic-based tool that takes advantage of GPU acceleration. The proposed VS pipeline is further refined through rescoring to pinpoint the most promising biocide compounds against cockroaches. We show the search space explored by different metaheuristics generated by METADOCK 2 and how this search is more exhaustive, but complementary, than the one offered by Glide. Finally, we applied Molecular Mechanics Generalized Born Surface Area (MMGBSA) to list the most promising compounds to inhibit the AChE enzyme.

PMID:39926230 | PMC:PMC11804414 | DOI:10.1039/d4ra07951e

World J Gastroenterol. 2025 Feb 7;31(5):102692. doi: 10.3748/wjg.v31.i5.102692.

ABSTRACT

In this letter, we comment on a recent article published in the World Journal of Gastroenterology by Xiao et al, where the authors aimed to use a deep learning model to automatically detect gastrointestinal lesions during capsule endoscopy (CE). CE was first presented in 2000 and was approved by the Food and Drug Administration in 2001. The indications of CE overlap with those of regular diagnostic endoscopy. However, in clinical practice, CE is usually used to detect lesions in areas inaccessible to standard endoscopies or in cases of bleeding that might be missed during conventional endoscopy. Since the emergence of CE, many physiological and technical challenges have been faced and addressed. In this letter, we summarize the current challenges and briefly mention the proposed methods to overcome these challenges to answer a central question: Do we still need CE?

PMID:39926220 | PMC:PMC11718605 | DOI:10.3748/wjg.v31.i5.102692

Sci Rep. 2025 Feb 9;15(1):4815. doi: 10.1038/s41598-025-88753-3.

ABSTRACT

Skin cancer is a prevalent health concern, and accurate segmentation of skin lesions is crucial for early diagnosis. Existing methods for skin lesion segmentation often face trade-offs between efficiency and feature extraction capabilities. This paper proposes Dual Skin Segmentation (DuaSkinSeg), a deep-learning model, to address this gap by utilizing dual encoders for improved performance. DuaSkinSeg leverages a pre-trained MobileNetV2 for efficient local feature extraction. Subsequently, a Vision Transformer-Convolutional Neural Network (ViT-CNN) encoder-decoder architecture extracts higher-level features focusing on long-range dependencies. This approach aims to combine the efficiency of MobileNetV2 with the feature extraction capabilities of the ViT encoder for improved segmentation performance. To evaluate DuaSkinSeg's effectiveness, we conducted experiments on three publicly available benchmark datasets: ISIC 2016, ISIC 2017, and ISIC 2018. The results demonstrate that DuaSkinSeg achieves competitive performance compared to existing methods, highlighting the potential of the dual encoder architecture for accurate skin lesion segmentation.

PMID:39924555 | DOI:10.1038/s41598-025-88753-3

Sci Rep. 2025 Feb 9;15(1):4825. doi: 10.1038/s41598-025-88579-z.

ABSTRACT

To accelerate the clinical adoption of quantitative magnetic resonance imaging (qMRI), frameworks are needed that not only allow for rapid acquisition, but also flexibility, cost efficiency, and high accuracy in parameter mapping. In this study, feed-forward deep neural network (DNN)- and iterative fitting-based frameworks are compared for multi-parametric (MP) relaxometry based on phase-cycled balanced steady-state free precession (pc-bSSFP) imaging. The performance of supervised DNNs (SVNN), self-supervised physics-informed DNNs (PINN), and an iterative fitting framework termed motion-insensitive rapid configuration relaxometry (MIRACLE) was evaluated in silico and in vivo in brain tissue of healthy subjects, including Monte Carlo sampling to simulate noise. DNNs were trained on three distinct in silico parameter distributions and at different signal-to-noise-ratios. The PINN framework, which incorporates physical knowledge into the training process, ensured more consistent inference and increased robustness to training data distribution compared to the SVNN. Furthermore, DNNs utilizing the full information of the underlying complex-valued MR data demonstrated ability to accelerate the data acquisition by a factor of 3. Whole-brain relaxometry using DNNs proved to be effective and adaptive, suggesting the potential for low-cost DNN retraining. This work emphasizes the advantages of in silico DNN MP-qMRI pipelines for rapid data generation and DNN training without extensive dictionary generation, long parameter inference times, or prolonged data acquisition, highlighting the flexible and rapid nature of lightweight machine learning applications for MP-qMRI.

PMID:39924554 | DOI:10.1038/s41598-025-88579-z

Sci Rep. 2025 Feb 9;15(1):4835. doi: 10.1038/s41598-025-85301-x.

ABSTRACT

Accurate segmentation of skin lesions is crucial for reliable clinical diagnosis and effective treatment planning. Automated techniques for skin lesion segmentation assist dermatologists in early detection and ongoing monitoring of various skin diseases, ultimately improving patient outcomes and reducing healthcare costs. To address limitations in existing approaches, we introduce a novel U-shaped segmentation architecture based on our Residual Space State Block. This efficient model, termed 'SSR-UNet,' leverages bidirectional scanning to capture both global and local features in image data, achieving strong performance with low computational complexity. Traditional CNNs struggle with long-range dependencies, while Transformers, though excellent at global feature extraction, are computationally intensive and require large amounts of data. Our SSR-UNet model overcomes these challenges by efficiently balancing computational load and feature extraction capabilities. Additionally, we introduce a spatially-constrained loss function that mitigates gradient stability issues by considering the distance between label and prediction boundaries. We rigorously evaluated SSR-UNet on the ISIC2017 and ISIC2018 skin lesion segmentation benchmarks. The results showed that the accuracy of Mean Intersection Over Union, Classification Accuracy and Specificity indexes in ISIC2017 datasets reached 80.98, 96.50 and 98.04, respectively, exceeding the best indexes of other models by 0.83, 0.99 and 0.38, respectively. The accuracy of Mean Intersection Over Union, Dice Coefficient, Classification Accuracy and Sensitivity on ISIC2018 datasets reached 82.17, 90.21, 95.34 and 88.49, respectively, exceeding the best indicators of other models by 1.71, 0.27, 0.65 and 0.04, respectively. It can be seen that SSR-UNet model has excellent performance in most aspects.

PMID:39924544 | DOI:10.1038/s41598-025-85301-x

Sci Rep. 2025 Feb 9;15(1):4826. doi: 10.1038/s41598-024-83597-9.

ABSTRACT

Mammography is the recommended imaging modality for breast cancer screening. Expressions of human epidermal growth factor receptor 2 (HER2), estrogen receptor (ER), and progesterone receptor (PR) are critical to the development of therapeutic strategies for breast cancer. In this study, a deep learning model (CBAM ResNet-18) was developed to predict the expression of these three receptors on mammography without manual segmentation of masses. Mammography of patients with pathologically proven breast cancer was obtained from two centers. A deep learning-based model (CBAM ResNet-18) for predicting HER2, ER, and PR expressions was trained and validated using five-fold cross-validation on a training dataset. The performance of the model was further tested using an external test dataset. Area under receiver operating characteristic curve (AUC), accuracy (ACC), and F1-score were calculated to assess the ability of the model to predict each receptor. For comparison we also developed original ResNet-18 without attention module and VGG-19 with and without attention module. The AUC (95% CI), ACC, and F1-score were 0.708 (0.609, 0.808), 0.651, 0.528, respectively, in the HER2 test dataset; 0.785 (0.673, 0.897), 0.845, 0.905, respectively, in the ER test dataset; and 0.706 (0.603, 0.809), 0.678, 0.773, respectively, in the PR test dataset. The proposed model demonstrates superior performance compared to the original ResNet-18 without attention module and VGG-19 with and without attention module. The model has the potential to predict HER2, PR, and especially ER expressions, and thus serve as an adjunctive diagnostic tool for breast cancer.

PMID:39924532 | DOI:10.1038/s41598-024-83597-9

Sci Rep. 2025 Feb 10;15(1):4857. doi: 10.1038/s41598-025-88688-9.

ABSTRACT

Raindrops on the windscreen significantly impact a driver's visibility during driving, affecting safe driving. Maintaining a clear windscreen is crucial for drivers to mitigate accident risks in rainy conditions. A real-time rain detection system and an innovative wiper control method are introduced based on machine vision and deep learning. An all-weather raindrop detection model is constructed using a convolutional neural network (CNN) architecture, utilising an improved YOLOv8 model. The all-weather model achieved a precision rate of 0.89, a recall rate of 0.83, and a detection speed of 63 fps, meeting the system's real-time requirements. The raindrop area ratio is computed through target detection, which facilitates the assessment of rainfall begins and ends, as well as intensity variations. When the raindrop area ratio exceeds the wiper activation threshold, the wiper starts, and when the area ratio approaches zero, the wiper stops. The wiper control method can automatically adjust the detection frequency and the wiper operating speed according to changes in rainfall intensity. The wiper activation threshold can be adjusted to make the wiper operation more in line with the driver's habits.

PMID:39924520 | DOI:10.1038/s41598-025-88688-9

Biomed Eng Online. 2025 Feb 9;24(1):16. doi: 10.1186/s12938-025-01348-x.

ABSTRACT

PURPOSE: The aim of this study is to convert low-dose PET (L-PET) images to full-dose PET (F-PET) images based on our Diffused Multi-scale Generative Adversarial Network (DMGAN) to offer a potential balance between reducing radiation exposure and maintaining diagnostic performance.

METHODS: The proposed method includes two modules: the diffusion generator and the u-net discriminator. The goal of the first module is to get different information from different levels, enhancing the generalization ability of the generator to the image and improving the stability of the training. Generated images are inputted into the u-net discriminator, extracting details from both overall and specific perspectives to enhance the quality of the generated F-PET images. We conducted evaluations encompassing both qualitative assessments and quantitative measures. In terms of quantitative comparisons, we employed two metrics, structure similarity index measure (SSIM) and peak signal-to-noise ratio (PSNR) to evaluate the performance of diverse methods.

RESULTS: Our proposed method achieved the highest PSNR and SSIM scores among the compared methods, which improved PSNR by at least 6.2% compared to the other methods. Compared to other methods, the synthesized full-dose PET image generated by our method exhibits a more accurate voxel-wise metabolic intensity distribution, resulting in a clearer depiction of the epilepsy focus.

CONCLUSIONS: The proposed method demonstrates improved restoration of original details from low-dose PET images compared to other models trained on the same datasets. This method offers a potential balance between minimizing radiation exposure and preserving diagnostic performance.

PMID:39924498 | DOI:10.1186/s12938-025-01348-x

Light Sci Appl. 2025 Feb 10;14(1):75. doi: 10.1038/s41377-024-01713-w.

ABSTRACT

Free of noble-metal and high in unit internal quantum efficiency of electroluminescence, organic molecules with thermally activated delayed fluorescence (TADF) features pose the potential to substitute metal-based phosphorescence materials and serve as the new-generation emitters for the mass production of organic light emitting diodes (OLEDs) display. Predicting the function of TADF emitters beyond classic chemical synthesis and material characterization experiments remains a great challenge. The advances in deep learning (DL) based artificial intelligence (AI) offer an exciting opportunity for screening high-performance TADF materials through efficiency evaluation. However, data-driven material screening approaches with the capacity to access the excited state properties of TADF emitters remain extremely difficult and largely unaddressed. Inspired by the fundamental principle that the excited state properties of TADF molecules are strongly dependent on their D-A geometric and electronic structures, we developed the Electronic Structure-Infused Network (ESIN) for TADF emitter screening. Designed with capacities of accurate prediction of the photoluminescence quantum yields (PLQYs) of TADF molecules based on elemental molecular geometry and orbital information and integrated with frontier molecular orbitals (FMOs) weight-based representation and modeling features, ESIN is a promising interpretable tool for emission efficiency evaluation and molecular design of TADF emitters.

PMID:39924488 | DOI:10.1038/s41377-024-01713-w

BMC Bioinformatics. 2025 Feb 9;26(1):45. doi: 10.1186/s12859-025-06069-5.

ABSTRACT

BACKGROUND: Single-cell technologies enable comprehensive profiling of diverse immune cell-types through the measurement of multiple genes or proteins per individual cell. In order to translate immune signatures assayed from blood or tissue into powerful diagnostics, machine learning approaches are often employed to compute immunological summaries or per-sample featurizations, which can be used as inputs to models for outcomes of interest. Current supervised learning approaches for computing per-sample representations are trained only to accurately predict a single outcome and do not take into account relevant additional clinical features or covariates that are likely to also be measured for each sample.

RESULTS: Here, we introduce a novel approach for incorporating measured covariates in optimizing model parameters to ultimately specify per-sample encodings that accurately affect both immune signatures and additional clinical information. Our introduced method CytoCoSet is a set-based encoding method for learning per-sample featurizations, which formulates a loss function with an additional triplet term penalizing samples with similar covariates from having disparate embedding results in per-sample representations.

CONCLUSIONS: Overall, incorporating clinical covariates enables the learning of encodings for each individual sample that ultimately improve prediction of clinical outcome. This integration of information disparate more robust predictions of clinical phenotypes and holds significant potential for enhancing diagnostic and treatment strategies.

PMID:39924480 | DOI:10.1186/s12859-025-06069-5

J Prosthet Dent. 2025 Feb 8:S0022-3913(25)00049-6. doi: 10.1016/j.prosdent.2024.12.029. Online ahead of print.

NO ABSTRACT

PMID:39924432 | DOI:10.1016/j.prosdent.2024.12.029

JMIR Med Inform. 2025 Feb 7;13:e55825. doi: 10.2196/55825.

ABSTRACT

BACKGROUND: Chronic kidney disease (CKD) is a prevalent condition with significant global health implications. Early detection and management are critical to prevent disease progression and complications. Deep learning (DL) models using retinal images have emerged as potential noninvasive screening tools for CKD, though their performance may be limited, especially in identifying individuals with proteinuria and in specific subgroups.

OBJECTIVE: We aim to evaluate the efficacy of integrating retinal images and urine dipstick data into DL models for enhanced CKD diagnosis.

METHODS: The 3 models were developed and validated: eGFR-RIDL (estimated glomerular filtration rate-retinal image deep learning), eGFR-UDLR (logistic regression using urine dipstick data), and eGFR-MMDL (multimodal deep learning combining retinal images and urine dipstick data). All models were trained to predict an eGFR<60 mL/min/1.73 m², a key indicator of CKD, calculated using the 2009 CKD-EPI (Chronic Kidney Disease Epidemiology Collaboration) equation. This study used a multicenter dataset of participants aged 20-79 years, including a development set (65,082 people) and an external validation set (58,284 people). Wide Residual Networks were used for DL, and saliency maps were used to visualize model attention. Sensitivity analyses assessed the impact of numerical variables.

RESULTS: eGFR-MMDL outperformed eGFR-RIDL in both the test and external validation sets, with area under the curves of 0.94 versus 0.90 and 0.88 versus 0.77 (P<.001 for both, DeLong test). eGFR-UDLR outperformed eGFR-RIDL and was comparable to eGFR-MMDL, particularly in the external validation. However, in the subgroup analysis, eGFR-MMDL showed improvement across all subgroups, while eGFR-UDLR demonstrated no such gains. This suggested that the enhanced performance of eGFR-MMDL was not due to urine data alone, but rather from the synergistic integration of both retinal images and urine data. The eGFR-MMDL model demonstrated the best performance in individuals younger than 65 years or those with proteinuria. Age and proteinuria were identified as critical factors influencing model performance. Saliency maps indicated that urine data and retinal images provide complementary information, with urine offering insights into retinal abnormalities and retinal images, particularly the arcade vessels, being key for predicting kidney function.

CONCLUSIONS: The MMDL model integrating retinal images and urine dipstick data show significant promise for noninvasive CKD screening, outperforming the retinal image-only model. However, routine blood tests are still recommended for individuals aged 65 years and older due to the model's limited performance in this age group.

PMID:39924305 | DOI:10.2196/55825

J Cancer Res Clin Oncol. 2025 Feb 9;151(2):72. doi: 10.1007/s00432-025-06105-0.

ABSTRACT

INTRODUCTION: For patients with breast cancer, the amplification of Human Epidermal Growth Factor 2 (HER2) is closely related to their prognosis and treatment decisions. This study aimed to further improve the accuracy and efficiency of HER2 amplification status detection with a deep learning model, and apply the model to predict the efficacy of neoadjuvant therapy.

METHODS: We combined Next-Generation Sequencing (NGS) data and IHC staining images of 606 breast cancer patients and developed a Vision Transformer (ViT) deep learning model to identify the amplification of HER2 through these IHC staining images. This model was then applied to predict the efficacy of neoadjuvant therapy in 399 HER2-positive breast cancer patients.

RESULTS: The NGS data of 606 patients were split into training (N = 404), validation (N = 101), and testing (N = 101) sets. The top 3 genes with highest mutation frequency were TP53, ERBB2 and PIK3CA. With the NGS results as deep learning model labels, the accuracy of our ViT model was 93.1% for HER2 amplification recognition. The misidentifications was likely due to the heterogeneity of HER2 expression in cancer tissues. For predicting the efficacy of neoadjuvant therapy, receiver operating characteristic (ROC) curves were plotted, and the combination of image recognition result and clinical pathological features yielded an area under the curve (AUC) value of 0.855 in the training set and 0.841 in the testing set.

CONCLUSIONS: Our study provided a method of HER2 status recognition based on IHC images, improving the efficiency and accuracy of HER2 status assessment, and can be used for predicting the efficacy of anti-HER2 targeted neoadjuvant therapy. We intend our deep learning model to assist pathologists in HER2 amplification recognition.

PMID:39923208 | DOI:10.1007/s00432-025-06105-0

J Neuroimaging. 2025 Jan-Feb;35(1):e70024. doi: 10.1111/jon.70024.

ABSTRACT

BACKGROUND AND PURPOSE: Accurate and consistent lesion segmentation from magnetic resonance imaging is required for longitudinal multiple sclerosis (MS) data analysis. In this work, we propose two new transfer learning-based pipelines to improve segmentation performance for subjects in longitudinal MS datasets.

METHOD: In general, transfer learning is used to improve deep learning model performance for the unseen dataset by fine-tuning a pretrained model with a limited number of labeled scans from the unseen dataset. The proposed methodologies fine-tune the deep learning model for each subject using the first scan and improve segmentation performance for later scans for the same subject. We also investigated the statistical benefits of the proposed methodology by modeling lesion volume over time between progressors according to confirmed disability progression and nonprogressors for a large in-house dataset (937 MS patients, 3210 scans) using a linear mixed effect (LME) model.

RESULTS: The results show statistically significant improvement for the proposed methodology compared with the traditional transfer learning method using Dice (improvement: 2%), sensitivity (6%), and average volumetric difference (16%), as well as visual analysis for public and in-house datasets. The LME result showed that the proposed subject-wise transfer learning method had increased statistical power for the measurement of longitudinal lesion volume.

CONCLUSION: The proposed method improved lesion segmentation performance and can reduce manual effort to correct the automatic segmentations for final data analysis in longitudinal studies.

PMID:39923192 | DOI:10.1111/jon.70024

Comput Med Imaging Graph. 2025 Feb 5;121:102505. doi: 10.1016/j.compmedimag.2025.102505. Online ahead of print.

ABSTRACT

In recent years, with the increasing attention from researchers towards medical imaging, deep learning-based image segmentation techniques have become mainstream in the field, requiring large amounts of manually annotated data. Annotating datasets for Intracranial Hemorrhage(ICH) is particularly tedious and costly. Few-shot segmentation holds significant potential for medical imaging. In this work, we designed a novel segmentation model CGNet to leverage a limited dataset for segmenting ICH regions, we propose a Cross Feature Module (CFM) enhances the understanding of lesion details by facilitating interaction between feature information from the query and support sets and Support Guide Query (SGQ) refines segmentation targets by integrating features from support and query sets at different scales, preserving the integrity of target feature information while further enhancing segmentation detail. We first propose transforming the ICH segmentation task into a few-shot learning problem. We evaluated our model using the publicly available BHSD dataset and the private IHSAH dataset. Our approach outperforms current state-of-the-art few-shot segmentation models, outperforming methods of 3% and 1.8% in Dice coefficient scores, respectively, and also exceeds the performance of fully supervised segmentation models with the same amount of data.

PMID:39921928 | DOI:10.1016/j.compmedimag.2025.102505

Comput Med Imaging Graph. 2025 Feb 1;121:102508. doi: 10.1016/j.compmedimag.2025.102508. Online ahead of print.

ABSTRACT

This study proposes a deep learning framework, DLPVI, which integrates projection, view-by-view backprojection (VVBP), and image domains to improve the quality of high-sparse-view and ultra-sparse-view cone-beam computed tomography (CBCT) images. The DLPVI comprises a projection domain sub-framework, a VVBP domain sub-framework, and a Transformer-based image domain model. First, full-view projections were restored from sparse-view projections via the projection domain sub-framework, then filtered and view-by-view backprojected to generate VVBP raw data. Next, the VVBP raw data was processed by the VVBP domain sub-framework to suppress residual noise and artifacts, and produce CBCT axial images. Finally, the axial images were further refined using the image domain model. The DLPVI was trained, validated, and tested on CBCT data from 163, 30, and 30 real patients respectively. Quantitative metrics including root-mean-square error (RMSE), peak signal-to-noise ratio (PSNR), structural similarity (SSIM), and feature similarity (FSIM) were calculated to evaluate the method performance. The DLPVI was compared with 15 state-of-the-art (SOTA) methods, including 2 projection domain models, 10 image domain models, and 3 projection-image dual-domain frameworks, on 1/8 high-sparse-view and 1/16 ultra-sparse-view reconstruction tasks. Statistical analysis was conducted using the Kruskal-Wallis test, followed by the post-hoc Dunn's test. Experimental results demonstrated that the DLPVI outperformed all 15 SOTA methods for both tasks, with statistically significant improvements (p < 0.05 in Kruskal-Wallis test and p < 0.05/15 in Dunn's test). The proposed DLPVI effectively improves the quality of high- and ultra-sparse-view CBCT images.

PMID:39921927 | DOI:10.1016/j.compmedimag.2025.102508

BMC Med Inform Decis Mak. 2025 Feb 8;25(1):66. doi: 10.1186/s12911-025-02912-0.

ABSTRACT

BACKGROUND: In Japan, reporting of medical device malfunctions and related health problems is mandatory, and efforts are being made to standardize terminology through the Adverse Event Terminology Collection of the Japan Federation of Medical Device Associations (JFMDA). Internationally, the Adverse Event Terminology of the International Medical Device Regulators Forum (IMDRF-AET) provides a standardized terminology collection in English. Mapping between the JFMDA terminology collection and the IMDRF-AET is critical to international harmonization. However, the process of translating the terminology collections from English to Japanese and reconciling them is done manually, resulting in high human workloads and potential inaccuracies.

OBJECTIVE: The purpose of this study is to investigate the optimal machine translation model for the IMDRF-AET into Japanese for the part of a function for the automatic terminology mapping system.

METHODS: English-Japanese parallel data for IMDRF-AET published by the Ministry of Health, Labor and Welfare in Japan was obtained from 50 sentences randomly extracted from the terms and their definitions. These English sentences were fed into the following machine translation models to produce Japanese translations: mBART50, m2m-100, Google Translation, Multilingual T5, GPT-3, ChatGPT, and GPT-4. The evaluations included the quantitative metrics of BiLingual Evaluation Understudy (BLEU), Character Error Rate (CER), Word Error Rate (WER), Metric for Evaluation of Translation with Explicit ORdering (METEOR), and Bidirectional Encoder Representations from Transformers (BERT) score, as well as qualitative evaluations by four experts.

RESULTS: GPT-4 outperformed other models in both the quantitative and qualitative evaluations, with ChatGPT showing the same capability, but with lower quantitative scores, in the qualitative evaluation. Scores of other models, including mBART50 and m2m-100, lagged behind, particularly in the CER and BERT scores.

CONCLUSION: GPT-4's superior performance in translating medical terminology, indicates its potential utility in improving the efficiency of the terminology mapping system.

PMID:39923074 | DOI:10.1186/s12911-025-02912-0

Commun Chem. 2025 Feb 8;8(1):40. doi: 10.1038/s42004-025-01437-x.

ABSTRACT

Designing molecules with desirable properties is a critical endeavor in drug discovery. Because of recent advances in deep learning, molecular generative models have been developed. However, the existing compound exploration models often disregard the important issue of ensuring the feasibility of organic synthesis. To address this issue, we propose TRACER, which is a framework that integrates the optimization of molecular property optimization with synthetic pathway generation. The model can predict the product derived from a given reactant via a conditional transformer under the constraints of a reaction type. The molecular optimization results of an activity prediction model targeting DRD2, AKT1, and CXCR4 revealed that TRACER effectively generated compounds with high scores. The transformer model, which recognizes the entire structures, captures the complexity of the organic synthesis and enables its navigation in a vast chemical space while considering real-world reactivity constraints.

PMID:39922979 | DOI:10.1038/s42004-025-01437-x

Sci Rep. 2025 Feb 8;15(1):4778. doi: 10.1038/s41598-025-89516-w.

ABSTRACT

Artificial intelligence (AI) and deep learning are making groundbreaking strides in protein structure prediction. AlphaFold is remarkable in this arena for its outstanding accuracy in modelling proteins fold based solely on their amino acid sequences. In spite of these remarkable advances, experimental structure determination remains critical. Here we report severe deviations between the experimental structure of a two-domain protein and its equivalent AI-prediction. These observations are particularly relevant to the relative orientation of the domains within the global protein scaffold. We observe positional divergence in equivalent residues beyond 30 Å, and an overall RMSD of 7.7 Å. Significant deviation between experimental structures and AI-predicted models echoes the presence of unusual conformations, insufficient training data and high complexity in protein folding that can ultimately lead to current limitations in protein structure prediction.

PMID:39922965 | DOI:10.1038/s41598-025-89516-w