publications
* equal contribution
2026
- Textual Supervision Enhances Geospatial Representations in Vision-Language ModelsIn Proceedings of the 43rd International Conference on Machine Learning, Seoul, South Korea, Jul 2026
Geospatial understanding is a critical yet underexplored dimension in the development of machine learning systems for tasks such as image geolocation and spatial reasoning. In this work, we analyze the geospatial representations acquired by three model families: vision-only architectures (e.g., ViT), vision-language models (e.g., CLIP), and large-scale multimodal foundation models (e.g., LLaVA, Qwen, and Gemma). By evaluating across image clusters, including people, landmarks, and everyday objects, grouped based on the degree of localizability, we reveal systematic gaps in spatial accuracy and show that textual supervision enhances the learning of geospatial representations. Our findings suggest the role of language as an effective complementary modality for encoding spatial context and multimodal learning as a key direction for advancing geospatial AI.
@inproceedings{locatelli2026textual, author = {Locatelli, Marcelo Sartori and Tonucci, Fernando and Kwon, Jea and Vecchietti, Luiz Felipe and Wijaya, Bryan Nathanael and Low, Cheng Yaw and Almeida, Virgilio and Cha, Meeyoung}, title = {Textual Supervision Enhances Geospatial Representations in Vision-Language Models}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, location = {Seoul, South Korea}, year = {2026}, month = jul, publisher = {PMLR}, series = {ICML '26}, } - Interpretable Machine Learning for Protein Science: Structure, Function, and InteractionsACM Computing Surveys, 2026
Recent advancements in machine learning (ML) are transforming the field of structural biology. For example, AlphaFold, a groundbreaking neural network for protein structure prediction, has been widely adopted by researchers. The availability of easy-to-use interfaces and interpretable outcomes from the neural network architecture, such as the confidence scores used to assess the quality of predicted structures, has made AlphaFold accessible even to non-ML experts. In this article, we present various methods for representing protein 3D structures from low- to high-resolution, and show how interpretable ML methods can support tasks such as predicting protein structures, protein functions, and protein-protein interactions. This survey also emphasizes the significance of interpreting and visualizing ML-based inference for structure-based protein representations that enhance interpretability and knowledge discovery. Developing such interpretable approaches promises to further accelerate fields, including drug development and protein design.
@article{vecchietti2026interpretable, author = {Vecchietti, Luiz Felipe and Lee, Minji and Hangeldiyev, Begench and Wijaya, Bryan Nathanael and Jung, Hyunkyu and Park, Hahnbeom and Kim, Tae-Kyun and Cha, Meeyoung and Kim, Ho Min}, title = {Interpretable Machine Learning for Protein Science: Structure, Function, and Interactions}, journal = {ACM Computing Surveys}, volume = {58}, number = {11}, pages = {289}, numpages = {24}, year = {2026}, publisher = {Association for Computing Machinery}, address = {New York, NY}, doi = {10.1145/3801737}, } - Longitudinal Identification of Critical Factors in Nurse Turnover for Supporting Workforce Well-BeingIn Proceedings of HCI Korea 2026, Hongcheon, South Korea, Jan 2026
Beyond addressing immediate staffing shortages, predictive models and early intervention strategies may help promote workforce well-being and retention. By analyzing retention dynamics through a large-scale longitudinal survey—with responses from over 25,000 nurses in South Korea, refined to a high-quality subset of approximately 5,000—we aim to develop data-driven tools for strategic workforce planning. Rather than profiling individuals, we focus on uncovering systemic patterns and risk factors associated with nurse resignation. This paper presents a transformer-based framework that captures personalized temporal patterns from a longitudinal survey, enabling nuanced modeling of nurse experiences over time. Our model estimates resignation risk and identifies critical factors through perturbation-based interpretability techniques. Our findings reveal actionable insights into the influential drivers of turnover, offering a data-driven foundation to design targeted retention policies and enhance workforce well-being in the healthcare sector.
@inproceedings{wijaya2026longitudinal, author = {Wijaya, Bryan Nathanael and Vecchietti, Luiz Felipe and Cho, Aram and Han, Sungwon and Cha, Meeyoung and Cha, Chiyoung}, title = {Longitudinal Identification of Critical Factors in Nurse Turnover for Supporting Workforce Well-Being}, booktitle = {Proceedings of HCI Korea 2026}, location = {Hongcheon, South Korea}, pages = {404--410}, numpages = {7}, year = {2026}, month = jan, publisher = {The HCI Society of Korea}, address = {Seoul}, } - Exploring the relationship between air quality and happiness in South Korea using artificial neural networksBryan Nathanael Wijaya, Yumi Park, Ju Hee Jeung, and Kyungmin LeeEnvironmental Impact Assessment Review, 2026
This study investigates the relationship between air quality and subjective happiness across South Korean districts using artificial neural network (ANN)-based modeling. By aggregating the Korean National Assembly Futures Institute’s happiness survey (2020–2021) data with the Korean Ministry of Environment’s air quality data, among others, six major air pollutants were examined for their potential associations with the happiness ladder at the minuscule city level throughout South Korea. Complex non-linear patterns were observed. Among the pollutants, PM2.5 exhibited the most consistent negative association with the happiness ladder. The robust modeling and training strategies provide insights into the intricate relationships between air quality factors and the individual happiness ladder. The analysis effectively captures subtle relationships under fixed socioeconomic and happiness-related conditions, highlighting varying confidence intervals across multiple scenarios. These findings underscore the potential of ANN-based modeling in assessing the environmental factors of subjective happiness. Despite limitations related to the spatiotemporal scale of the annual happiness survey, this study contributes to the methods by applying deep learning techniques to infer the relationship between air quality and happiness, providing evidence that may inform environmental policymaking and urban sustainability strategies.
@article{wijaya2026exploring, author = {Wijaya, Bryan Nathanael and Park, Yumi and Jeung, Ju Hee and Lee, Kyungmin}, title = {Exploring the relationship between air quality and happiness in South Korea using artificial neural networks}, journal = {Environmental Impact Assessment Review}, volume = {117}, pages = {108135}, year = {2026}, publisher = {Elsevier}, address = {Amsterdam}, doi = {10.1016/j.eiar.2025.108135}, }
2025
- Artificial intelligence-driven computational methods for antibody design and optimizationLuiz Felipe Vecchietti*, Bryan Nathanael Wijaya*, Azamat Armanuly, Begench Hangeldiyev, and 4 more authorsmAbs, 2025
Antibodies play a crucial role in our immune system. Their ability to bind to and neutralize pathogens opens opportunities to develop antibodies for therapeutic and diagnostic use. Computational methods capable of designing antibodies for a target antigen can revolutionize drug discovery, reducing the time and cost required for drug development. Artificial intelligence (AI) methods have recently achieved remarkable advancements in the design of protein sequences and structures, including the ability to generate scaffolds for a given motif and binders for a specific target. These generative methods have been applied to antigen-conditioned antibody design, with experimental binding confirmed for de novo-designed antibodies. This review surveys current AI methods used in antibody development, focusing on those for antigen-conditioned antibody design. The results obtained by AI-based methodologies in antibody and protein research suggest a promising direction for generating de novo binders for various target antigens.
@article{vecchietti2025artificial, author = {Vecchietti, Luiz Felipe and Wijaya, Bryan Nathanael and Armanuly, Azamat and Hangeldiyev, Begench and Jung, Hyunkyu and Lee, Sooyeon and Cha, Meeyoung and Kim, Ho Min}, title = {Artificial intelligence-driven computational methods for antibody design and optimization}, journal = {mAbs}, volume = {17}, number = {1}, pages = {2528902}, year = {2025}, publisher = {Taylor \& Francis}, address = {London}, doi = {10.1080/19420862.2025.2528902}, }
2024
- Leveraging Synthetic Data for Data-Free Knowledge DistillationBiniyam Aschalew Tolera*, Bryan Nathanael Wijaya*, and Minhajur Rahman Chowdhury MahimIn Proceedings of KIISE Korea Software Congress 2024, Yeosu, South Korea, Dec 2024
Knowledge distillation has emerged as an effective technique for transferring knowledge from large, complex teacher models to smaller, efficient student models, enabling the deployment of high-performing models on resource-constrained devices. Traditional knowledge distillation methods, however, rely on access to the original training data, which may not be available due to privacy concerns, proprietary restrictions, or data size. Existing data-free knowledge distillation approaches often depend on adversarial generative models, introducing additional complexity and computational overhead. In this paper, we propose a simple yet effective method for data-free knowledge distillation by generating synthetic data through activation maximization of the teacher model, thereby eliminating the need for the original training data or complex generative models. We evaluate our method on MNIST, SVHN, CIFAR-10, and ImageNet32 datasets, demonstrating that significant knowledge can be transferred to student models using only the pre-trained teacher model and zero amount of labeled or unlabeled data. Furthermore, our method requires significantly less computational resources compared to related approaches. Our implementation can be accessed at https://github.com/BiniyamAschalew/KD570.git.
@inproceedings{tolera2024leveraging, author = {Tolera, Biniyam Aschalew and Wijaya, Bryan Nathanael and Mahim, Minhajur Rahman Chowdhury}, title = {Leveraging Synthetic Data for Data-Free Knowledge Distillation}, booktitle = {Proceedings of KIISE Korea Software Congress 2024}, location = {Yeosu, South Korea}, pages = {692--694}, numpages = {3}, year = {2024}, month = dec, publisher = {Korean Institute of Information Scientists and Engineers}, address = {Seoul}, }
2023
- Evaluating Antibody Structure Reconstruction with an SE(3)-Equivariant Graph Neural NetworkIn Proceedings of KIISE Korea Software Congress 2023, Busan, South Korea, Dec 2023
Received the Excellence Award for Undergraduate Division from KIISE in Feb 2024.
Graph and diffusion methods have achieved breakthroughs in the generation of protein and antibody structures, which can be represented as a 3D heterogeneous atom point cloud. However, given the complexity of these 3D structures, for computational efficiency, a coarse-grained (simplified) representation is usually used for generation, while the full atom point cloud is recovered after generation. Therefore, it is crucial to have reconstruction methods that can recover the full atom point cloud with minimum loss. Recent graph methods have been proposed for protein structure reconstruction. Among them, cg2all is an SE(3)-equivariant graph method that allows the recovery of all-atom protein structures from various coarse-grained representations. This paper presents an evaluation of cg2all in reconstructing antibody structures and identifies the factors influencing the performance. We show that reducing antibody structures to a coarse-grained representation via cg2all can significantly improve computational efficiency in its structural studies, and having an effective reconstruction model can accelerate the development of new methods in antibody design. In addition, the antibody reconstruction performance is influenced by the chosen coarse-grained representation complexity but is not affected by the variability of the antibody regions.
@inproceedings{wijaya2023evaluating, author = {Wijaya, Bryan Nathanael and Vecchietti, Luiz Felipe and Cha, Meeyoung and Kim, Ho Min}, title = {Evaluating Antibody Structure Reconstruction with an SE(3)-Equivariant Graph Neural Network}, booktitle = {Proceedings of KIISE Korea Software Congress 2023}, location = {Busan, South Korea}, pages = {1681--1683}, numpages = {3}, year = {2023}, month = dec, publisher = {Korean Institute of Information Scientists and Engineers}, address = {Seoul}, }
2020
- Synthesis and characterization of novel chelation-free Zn(II)-azole complexes: Evaluation of antibacterial, antioxidant, and DNA binding activitySondavid K. Nandanwar, Shweta B. Borkar, Bryan Nathanael Wijaya, Enkhnaran Bayarra, and 5 more authorsIndian Journal of Chemistry - Section A Inorganic, Physical, Theoretical and Analytical Chemistry, 2020
Here, we synthesized novel chelation-free Zn(II)-complexes (1-3) [ZnCl2L2] of monodentate ligands with L = 2-isopropylimidazole (L1), 2-methylbenzimidazole (L2), and 2-methylbenzoxazole (L3) and evaluated their antibacterial, antioxidant and DNA binding activities. The chelation-free properties of these coordination complexes were confirmed by UV-visible spectroscopy, H-1 NMR spectroscopy, single X-ray crystallography and elemental analysis. Complexes 1-3 exhibited substantial antibacterial activity against all antibiotic susceptible bacteria within a concentration range of 100-200 mu g/ml while free ligands L1 and L2 exhibited weak antibacterial activity considerably concentration above 200 mu g/ml. Also, both complexes 2 and 3 were twice more active against methicillin-resistant Staphylococcus aureus (MRSA) than complex 1. Furthermore, we found that complexes 1-3 showed DNA binding activity with E. coli plasmid DNA and calf thymus DNA, which may be a plausible mechanism for their antibacterial activity. We also investigated the antioxidant activity of complexes 1-3 and found that complex 2 exhibited potential antioxidant activity compared to complexes 1 and 3. All these results suggest that the chelation-free Zn(II)-complexes can be the future candidates for more advance biological studies.
@article{nandanwar2020synthesis, author = {Nandanwar, Sondavid K. and Borkar, Shweta B. and Wijaya, Bryan Nathanael and Bayarra, Enkhnaran and Piad, Lei Lottice Anne and Jin, QiuLi and Tiara, Celine Sheila and Kim, Hak Jun and Tarte, Naresh H.}, title = {Synthesis and characterization of novel chelation-free Zn(II)-azole complexes: Evaluation of antibacterial, antioxidant, and DNA binding activity}, journal = {Indian Journal of Chemistry - Section A Inorganic, Physical, Theoretical and Analytical Chemistry}, volume = {59}, number = {5}, pages = {589--597}, year = {2020}, publisher = {CSIR-National Institute of Science Communication and Policy Research}, address = {New Delhi}, doi = {10.56042/ijca.v59i5.24897} } - Cobalt(II) Benzazole Derivative Complexes: Synthesis, Characterization, Antibacterial and Synergistic ActivitySondavid K. Nandanwar, Shweta B. Borkar, Bryan Nathanael Wijaya, Joung Hyung Cho, and 2 more authorsChemistrySelect, 2020
In this study, we synthesized the cobalt(II) chelation free complexes (1–3) with the general formula of [CoCl2L2] containing monodentate ligands (L=2-methylbenzimidazole (L1), 2-methylbenzoxazole (L2), and 2-methylbenzothiazole (L3) and fully characterized by UV-vis, FT-IR spectroscopy, and the crystal structures of 1–3 were determined by X-ray crystallography. Complexes 1–3 showed substantial antibacterial activity against tested bacteria. Complexes 2 and 3 were more active against S. aureus with the Minimum Inhibitory concentration (MIC) value of 12.5 μg/mL compared to other bacteria. Complexes 1–3 exhibited strong synergy with ampicillin (AMP) against Methicillin Resistant S. aureus strains (MRSA). Interestingly, complexes 2 and 3 in combination with AMP showed the membrane permeabilization of MRSA. This indicates that complexes 2 and 3 in combination with AMP had bacterial membrane disruption as a possible mechanism of action.
@article{nandanwar2020cobalt, author = {Nandanwar, Sondavid K. and Borkar, Shweta B. and Wijaya, Bryan Nathanael and Cho, Joung Hyung and Tarte, Naresh H. and Kim, Hak Jun}, title = {Cobalt(II) Benzazole Derivative Complexes: Synthesis, Characterization, Antibacterial and Synergistic Activity}, journal = {ChemistrySelect}, volume = {5}, number = {11}, pages = {3471--3476}, year = {2020}, publisher = {Wiley-VCH}, address = {Weinheim}, doi = {10.1002/slct.202000222}, }