Hyeongwoo Kim
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Hyeongwoo Kim has a total of 12 co-authors. Here are the top co-authors by number of shared publications:
- Woo Youn Kim
Department of Chemistry, KAIST, Daejeon, Republic of Korea
5 shared publications - Seokhyun Moon
Department of Chemistry, KAIST, Daejeon, Republic of Korea
4 shared publications - Jaechang Lim
HITS Inc., Seoul, Republic of Korea
2 shared publications - Shinwoo Kim
Korea Advanced Institute of Science and Technology, Daejeon, Republic of Korea
2 shared publications - Wonho Zhung
Department of Chemistry, KAIST, Daejeon, Republic of Korea
2 shared publications Chansu Kim
Department of Chemistry, KAIST, 291 Daehak-ro, Yuseong-gu, Daejeon 34141, Republic of Korea
1 shared publication
Latest Publications
- Open Access
How Do Co-folding Models Organize Structural Information?
Co-folding models accurately predict biomolecular complexes, but how their internal representations support joint structure prediction remains unclear. We analyze the Boltz-1 trunk by decomposing its pair representation into intra-chain and inter-chain blocks and applying ablation, layer-wise representation geometry, and activation patching. Our results show that structural information is distributed across three streams: the single representation carries token-internal atomic detail, the intra-chain pair representation encodes chain-level geometry, and the inter-chain pair representation controls relative molecular arrangement. Across the trunk depth, representations follow a Mix–Compress–Refine trajectory, while linear probes reveal that intra-chain distance content is largely supplied by MSA preconditioning and inter-chain distance content is constructed progressively across Pairformer layers at each recycle. Apo–holo patching further shows that ligand-induced backbone changes are not fully resolved before diffusion; instead, the diffusion module reconciles partially inconsistent intra- and inter-chain constraints. These findings provide a mechanistic view of how co-folding trunks organize structural information and suggest more partition-aware designs for future biomolecular structure prediction models.
Open AccessHow Do Co-folding Models Organize Structural Information?
Co-folding models accurately predict biomolecular complexes, but how their internal representations support joint structure prediction remains unclear. We analyze the Boltz-1 trunk by decomposing its pair representation into intra-chain and inter-chain blocks and applying ablation, layer-wise representation geometry, and activation patching. Our results show that structural information is distributed across three streams: the single representation carries token-internal atomic detail, the intra-chain pair representation encodes chain-level geometry, and the inter-chain pair representation controls relative molecular arrangement. Across the trunk depth, representations follow a Mix–Compress–Refine trajectory, while linear probes reveal that intra-chain distance content is largely supplied by MSA preconditioning and inter-chain distance content is constructed progressively across Pairformer layers at each recycle. Apo–holo patching further shows that ligand-induced backbone changes are not fully resolved before diffusion; instead, the diffusion module reconciles partially inconsistent intra- and inter-chain constraints. These findings provide a mechanistic view of how co-folding trunks organize structural information and suggest more partition-aware designs for future biomolecular structure prediction models.
- Open Access
AtlasFold: Protein structure prediction with metagenomic-scale language models
Protein language models (PLMs) trained on evolutionary sequences learn representations that encode protein structure, enabling direct structure prediction without multiple-sequence alignments (MSAs). Here we present the Atlas model family, an open and trainable system spanning protein language modeling, monomer folding, and protein-complex prediction. AtlasLM-3B is a 3B-scale language model trained with masked language modeling on approximately 1.56 billion sequences, including metagenomic data, and outperforms the similarly sized ESM2-3B in unsupervised contact prediction. Building on these representations, AtlasFold predicts all-atom protein structures and achieves state-of-the-art accuracy among PLM-based folding models. Fine-tuning AtlasFold for protein-complex prediction produces AtlasFold-Multimer (AtlasFold-M), whose antibody–antigen prediction performance is comparable to that of AlphaFold3 and ESMFold2. This protein-specific folding architecture enables fast, memory-efficient inference with AtlasFold and AtlasFold-M. By releasing the training code and data, stage checkpoints, and model weights under the MIT License, we provide a foundation for advancing PLM-based protein structure prediction. 1
Open AccessAtlasFold: Protein structure prediction with metagenomic-scale language models
Protein language models (PLMs) trained on evolutionary sequences learn representations that encode protein structure, enabling direct structure prediction without multiple-sequence alignments (MSAs). Here we present the Atlas model family, an open and trainable system spanning protein language modeling, monomer folding, and protein-complex prediction. AtlasLM-3B is a 3B-scale language model trained with masked language modeling on approximately 1.56 billion sequences, including metagenomic data, and outperforms the similarly sized ESM2-3B in unsupervised contact prediction. Building on these representations, AtlasFold predicts all-atom protein structures and achieves state-of-the-art accuracy among PLM-based folding models. Fine-tuning AtlasFold for protein-complex prediction produces AtlasFold-Multimer (AtlasFold-M), whose antibody–antigen prediction performance is comparable to that of AlphaFold3 and ESMFold2. This protein-specific folding architecture enables fast, memory-efficient inference with AtlasFold and AtlasFold-M. By releasing the training code and data, stage checkpoints, and model weights under the MIT License, we provide a foundation for advancing PLM-based protein structure prediction. 1
- Open Access
- 27 July 2026
- Published by Springer Nature in Nature Communications
- Vol. 17 (1) , 9113
Autonomous bioisosteric replacement for multi-property optimization in drug design
Optimizing molecular properties while preserving biological activity is a central challenge in drug design. Bioisosteric replacement, which substitutes a molecular fragment with a chemically or biologically analogous moiety, offers a powerful strategy for fine-tuning properties without disrupting target binding. However, existing in silico approaches often rely on expert-defined modification sites or struggle to modulate multiple molecular properties simultaneously. Here, we present DeepBioisostere, a deep generative model that performs end-to-end bioisosteric replacement by autonomously selecting and substituting molecular fragments to satisfy multiple target properties. The model captures complex relationships across the molecular graph, enabling the optimization of sophisticated properties such as drug-likeness and synthetic accessibility. By learning from experimental bioassay data, DeepBioisostere proposes replacements that maintain biological activities, even generating potential bioisosteres beyond the training data. We demonstrate the effectiveness of the model in computational hit-to-lead optimization scenarios, highlighting its potential to accelerate rational molecular design without relying on expert heuristics or pre-established substitution rules.
- 27 July 2026
- Published by Springer Nature in Nature Communications
- Vol. 17 (1) , 9113
Open AccessAutonomous bioisosteric replacement for multi-property optimization in drug design
Optimizing molecular properties while preserving biological activity is a central challenge in drug design. Bioisosteric replacement, which substitutes a molecular fragment with a chemically or biologically analogous moiety, offers a powerful strategy for fine-tuning properties without disrupting target binding. However, existing in silico approaches often rely on expert-defined modification sites or struggle to modulate multiple molecular properties simultaneously. Here, we present DeepBioisostere, a deep generative model that performs end-to-end bioisosteric replacement by autonomously selecting and substituting molecular fragments to satisfy multiple target properties. The model captures complex relationships across the molecular graph, enabling the optimization of sophisticated properties such as drug-likeness and synthetic accessibility. By learning from experimental bioassay data, DeepBioisostere proposes replacements that maintain biological activities, even generating potential bioisosteres beyond the training data. We demonstrate the effectiveness of the model in computational hit-to-lead optimization scenarios, highlighting its potential to accelerate rational molecular design without relying on expert heuristics or pre-established substitution rules.
- Open Access
- 9 July 2024
- Published by Springer Nature in Research Square
DeepBioisostere: Deep Learning-based Bioisosteric Replacements for Optimization of Multiple Molecular Properties
Optimizing molecules to improve their properties is a fundamental challenge in drug design. For a fine-tuning of molecular properties without losing bio-activity validated in advance, the concept of bioisosterism has emerged. Many in silico methods have been proposed for discovering bioisosteres, but they require expert knowledge for their applications or are restricted to known databases. Here, we introduce DeepBioisostere, a deep generative model to design suitable bioisosteric replacements. Our model allows a fully automated bioisosteric replacement by intelligently selecting fragments for both removal and insertion. Through various scenarios of multiple property control, we showcase the model's capability to delicately tune specific properties, addressing the challenge in molecular optimization. Our model's innovation lies in its capacity to design a bioisosteric replacement reflecting the compatibility with the surroundings of the modification site, facilitating the control of sophisticated properties that depend on the overall molecular structures such as drug-likeness. DeepBioisostere can also provide previously unseen bioisosteric replacements, highlighting its capability for exploring diverse chemical modifications rather than just mining them from known databases. Lastly, we used DeepBioisostere to improve the sensitivity of a known SARS-CoV-2 main protease inhibitor to the E166V mutant that exhibits drug resistance to the inhibitor, demonstrating its potential application in drug discovery.
- 9 July 2024
- Published by Springer Nature in Research Square
Open AccessDeepBioisostere: Deep Learning-based Bioisosteric Replacements for Optimization of Multiple Molecular Properties
Optimizing molecules to improve their properties is a fundamental challenge in drug design. For a fine-tuning of molecular properties without losing bio-activity validated in advance, the concept of bioisosterism has emerged. Many in silico methods have been proposed for discovering bioisosteres, but they require expert knowledge for their applications or are restricted to known databases. Here, we introduce DeepBioisostere, a deep generative model to design suitable bioisosteric replacements. Our model allows a fully automated bioisosteric replacement by intelligently selecting fragments for both removal and insertion. Through various scenarios of multiple property control, we showcase the model's capability to delicately tune specific properties, addressing the challenge in molecular optimization. Our model's innovation lies in its capacity to design a bioisosteric replacement reflecting the compatibility with the surroundings of the modification site, facilitating the control of sophisticated properties that depend on the overall molecular structures such as drug-likeness. DeepBioisostere can also provide previously unseen bioisosteric replacements, highlighting its capability for exploring diverse chemical modifications rather than just mining them from known databases. Lastly, we used DeepBioisostere to improve the sensitivity of a known SARS-CoV-2 main protease inhibitor to the E166V mutant that exhibits drug resistance to the inhibitor, demonstrating its potential application in drug discovery.
- 31 August 2023
- Published by American Chemical Society (ACS) in Journal of Chemical Information and Modeling
- Vol. 64 (7) , 2432-2444
DFRscore: Deep Learning-Based Scoring of Synthetic Complexity with Drug-Focused Retrosynthetic Analysis for High-Throughput Virtual Screening
Recently emerging generative AI models enable us to produce a vast number of compounds for potential applications. While they can provide novel molecular structures, the synthetic feasibility of the generated molecules is often questioned. To address this issue, a few recent studies have attempted to use deep learning models to estimate the synthetic accessibility of many molecules rapidly. However, retrosynthetic analysis tools used to train the models rely on reaction templates automatically extracted from a large reaction database that are not domain-specific and may exhibit low chemical correctness. To overcome this limitation, we introduce DFRscore (Drug-Focused Retrosynthetic score), a deep learning-based approach for a more practical assessment of synthetic accessibility in drug discovery. The DFRscore model is trained exclusively on drug-focused reactions, providing a predicted number of minimally required synthetic steps for each compound. This approach enables practitioners to filter out compounds that do not meet their desired level of synthetic accessibility at an early stage of high-throughput virtual screening for accelerated drug discovery. The proposed strategy can be easily adapted to other domains by adjusting the synthesis planning setup of the reaction templates and starting materials.
- 31 August 2023
- Published by American Chemical Society (ACS) in Journal of Chemical Information and Modeling
- Vol. 64 (7) , 2432-2444
DFRscore: Deep Learning-Based Scoring of Synthetic Complexity with Drug-Focused Retrosynthetic Analysis for High-Throughput Virtual Screening
Recently emerging generative AI models enable us to produce a vast number of compounds for potential applications. While they can provide novel molecular structures, the synthetic feasibility of the generated molecules is often questioned. To address this issue, a few recent studies have attempted to use deep learning models to estimate the synthetic accessibility of many molecules rapidly. However, retrosynthetic analysis tools used to train the models rely on reaction templates automatically extracted from a large reaction database that are not domain-specific and may exhibit low chemical correctness. To overcome this limitation, we introduce DFRscore (Drug-Focused Retrosynthetic score), a deep learning-based approach for a more practical assessment of synthetic accessibility in drug discovery. The DFRscore model is trained exclusively on drug-focused reactions, providing a predicted number of minimally required synthetic steps for each compound. This approach enables practitioners to filter out compounds that do not meet their desired level of synthetic accessibility at an early stage of high-throughput virtual screening for accelerated drug discovery. The proposed strategy can be easily adapted to other domains by adjusting the synthesis planning setup of the reaction templates and starting materials.