JOURNALS
2026
2.
Yunju Song; Hwan Choi; Sunyong Yoo
Abstract | Links | BibTeX | Dimensions | Tags: Attention mechanism, Bioinformatics, Carcinogenicity, Deep learning, in silico, Systems biology
@article{Song2026,
title = {Carcinogenicity prediction via multi-task learning of cross-organ representations with attention mechanisms},
author = {Yunju Song and Hwan Choi and Sunyong Yoo},
url = {https://academic.oup.com/bib/article/27/3/bbag296/8702471?login=true},
doi = {10.1093/bib/bbag296},
issn = {1477-4054},
year = {2026},
date = {2026-06-04},
urldate = {2026-06-04},
journal = {Briefings in Bioinformatics},
volume = {27},
issue = {3},
pages = {bbag296},
abstract = {Cancer is caused by the uncontrolled growth and division of abnormal cells. In industrialized societies, chemical exposure is a leading cause of cancer. Since certain compounds induce cancer by damaging genes or affecting cellular metabolism, studying carcinogens is essential. However, previous studies used separate models for each organ and failed to capture carcinogenic features shared across organs, limiting generalization. Thus, this study developed a multi-task learning framework to predict organ-specific carcinogenicity in the liver, lung, stomach, and breast. This framework consisted of a shared layer and task-specific layers. The shared layer contains a graph attention network layer to make atom-level representations, along with parallel fully connected layers designed for each task combination. The resulting shared representations are passed to task-specific layers to predict organ-specific carcinogenicity. The training process followed stepwise learning, whereby the model was first trained using partially labeled data to capture cross-organ representations and determine initial weights. In the second step, fully labeled data for all organs were used for final training. The proposed multi-task model achieved superior performance in the liver, lung, and stomach tasks. Notably, it recorded the highest area under the receiver operating characteristic curve in the stomach task (0.7636), outperforming the single-task model (0.7055) and all comparative models (0.5527–0.7418). The highest area under the precision–recall curve was observed in the liver task (0.9646), surpassing the single-task model (0.9505) and all comparative models (0.9373–0.9621). We further analyzed molecules with high predicted carcinogenicity and identified critical substructures using an attention mechanism. This research can contribute to predicting organ-specific carcinogenicity of candidate chemicals in the early stages of drug development.
},
note = {Correspondence to Sunyong Yoo},
keywords = {Attention mechanism, Bioinformatics, Carcinogenicity, Deep learning, in silico, Systems biology},
pubstate = {published},
tppubtype = {article}
}
Cancer is caused by the uncontrolled growth and division of abnormal cells. In industrialized societies, chemical exposure is a leading cause of cancer. Since certain compounds induce cancer by damaging genes or affecting cellular metabolism, studying carcinogens is essential. However, previous studies used separate models for each organ and failed to capture carcinogenic features shared across organs, limiting generalization. Thus, this study developed a multi-task learning framework to predict organ-specific carcinogenicity in the liver, lung, stomach, and breast. This framework consisted of a shared layer and task-specific layers. The shared layer contains a graph attention network layer to make atom-level representations, along with parallel fully connected layers designed for each task combination. The resulting shared representations are passed to task-specific layers to predict organ-specific carcinogenicity. The training process followed stepwise learning, whereby the model was first trained using partially labeled data to capture cross-organ representations and determine initial weights. In the second step, fully labeled data for all organs were used for final training. The proposed multi-task model achieved superior performance in the liver, lung, and stomach tasks. Notably, it recorded the highest area under the receiver operating characteristic curve in the stomach task (0.7636), outperforming the single-task model (0.7055) and all comparative models (0.5527–0.7418). The highest area under the precision–recall curve was observed in the liver task (0.9646), surpassing the single-task model (0.9505) and all comparative models (0.9373–0.9621). We further analyzed molecules with high predicted carcinogenicity and identified critical substructures using an attention mechanism. This research can contribute to predicting organ-specific carcinogenicity of candidate chemicals in the early stages of drug development.
2025
1.
Md Sanzid Bin Hossain; Hwan Choi; Zhishan Guo; Sunyong Yoo; Min-Keun Song; Hyunjun Shin; Dexter Hadley
Abstract | Links | BibTeX | Dimensions | Tags: Systems biology
@article{Hossain2025,
title = {Knowledge transfer-driven estimation of knee moments and ground reaction forces from smartphone videos via temporal-spatial modeling of augmented joint kinematics},
author = {Md Sanzid Bin Hossain and Hwan Choi and Zhishan Guo and Sunyong Yoo and Min-Keun Song and Hyunjun Shin and Dexter Hadley},
url = {https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0335257},
doi = {10.1371/journal.pone.0335257},
issn = {1932-6203},
year = {2025},
date = {2025-11-07},
urldate = {2025-11-07},
journal = {PLOS One},
volume = {20},
number = {11},
pages = {e0335257},
abstract = {The knee adduction and flexion moment provides critical information about knee joint health, while 3D ground reaction forces (GRFs) help identify force and energy characteristics for maneuvering the entire human body. Existing methods of acquiring joint moments and GRFs require expensive equipment, time-consuming pre-processing, and limited accessibility. This study proposes to tackle these limitations by utilizing only smartphone videos to estimate joint moments and 3D GRFs accurately. We also propose the augmentation of joint kinematics by generating additional modalities of 2D joint center velocity and acceleration from 2D joint center position acquired from the videos. This augmented joint kinematics helps to apply a multi-modal fusion module to learn the importance of inter-modal interactions. Additionally, we utilize recurrent neural networks and graph convolutional networks to perform temporal-spatial modeling of joint center dynamics for enhanced accuracy. To overcome another challenge of video-based estimation, particularly the lack of inertial information related to body segments, we propose multi-modal knowledge transfer to train the video-only student model from a teacher model that integrates both video and inertial measurement unit (IMU) data. The student model significantly reduces the normalized root mean square error (NRMSE) from 5.71 to 4.68 and increases the Pearson correlation coefficient (PCC) from 0.929 to 0.951. These results demonstrate that knowledge transfer, augmentation of joint kinematics for multi-modal fusion, and temporal-spatial modeling significantly enhance smartphone video-based estimation, offering a potential cost-effective alternative to traditional motion capture for clinical assessments, rehabilitation, and sports applications.},
note = {Correspondence to Hwan Choi},
keywords = {Systems biology},
pubstate = {published},
tppubtype = {article}
}
The knee adduction and flexion moment provides critical information about knee joint health, while 3D ground reaction forces (GRFs) help identify force and energy characteristics for maneuvering the entire human body. Existing methods of acquiring joint moments and GRFs require expensive equipment, time-consuming pre-processing, and limited accessibility. This study proposes to tackle these limitations by utilizing only smartphone videos to estimate joint moments and 3D GRFs accurately. We also propose the augmentation of joint kinematics by generating additional modalities of 2D joint center velocity and acceleration from 2D joint center position acquired from the videos. This augmented joint kinematics helps to apply a multi-modal fusion module to learn the importance of inter-modal interactions. Additionally, we utilize recurrent neural networks and graph convolutional networks to perform temporal-spatial modeling of joint center dynamics for enhanced accuracy. To overcome another challenge of video-based estimation, particularly the lack of inertial information related to body segments, we propose multi-modal knowledge transfer to train the video-only student model from a teacher model that integrates both video and inertial measurement unit (IMU) data. The student model significantly reduces the normalized root mean square error (NRMSE) from 5.71 to 4.68 and increases the Pearson correlation coefficient (PCC) from 0.929 to 0.951. These results demonstrate that knowledge transfer, augmentation of joint kinematics for multi-modal fusion, and temporal-spatial modeling significantly enhance smartphone video-based estimation, offering a potential cost-effective alternative to traditional motion capture for clinical assessments, rehabilitation, and sports applications.