Juyong Lee

Protein generative models design therapeutic backbones and sequences that are then filtered by cross-validation metrics. We review these methods, their therapeutic applications, and how far these metrics actually predict experimental success.

Motivation: Accurate peptide–MHC class II (pMHC-II) binding prediction is complicated by MHC polymorphism and context-dependent peptide recognition. Protein representations are usually evaluated in different pipelines, obscuring their contribution to predictive performance. Results: We present PREpiBind, a joint-attention framework comparing ten protein representations under a fixed architecture and identical data splits across qualitative binding, mass spectrometry, and thresholded IC50 datasets. PLMs generally achieved the strongest pooled performance; ESM3 Large reached a ROC-AUC of 0.927 ± 0.002 on the qualitative dataset. PREpiBind PLM variants exceeded evaluated reference methods on qualitative and mass-spectrometry benchmarks, whereas NetMHCIIpan-4.3 was higher on thresholded IC50 using its binding-affinity head. PLM advantages narrowed in allele-wise and held-out MHC-molecule evaluations, and H2-out rankings depended on aggregation, indicating scenario-dependent representation performance. Availability and implementation: Source code and the training and evaluation datasets are available at https://github.com/daylight-00/PREpiBind and archived at https://doi.org/10.5281/zenodo.22934281 ; trained model checkpoints are at https://huggingface.co/daylight-00/prepibind . Implemented in Python and released under the MIT License. Contact: nicole23@snu.ac.kr Supplementary information: Supplementary data are available online.

A bstract: Biomolecular structures provide a common physical basis for understanding molecular function and interactions. However, computational models typically specialize in individual molecular classes or tasks, hindering a unified representation of structural information across biological systems. Here we introduce OmniScore, a universal geometric encoder that learns shared representations of proteins, nucleic acids, small molecules and their complexes. Pretraining integrates geometric graph encoding, sequential modeling and tokenization with structural and physicochemical supervision. The frozen pretrained encoder supports complex-level model quality assessment and interaction prediction, as well as residue-level prediction, through task-specific heads. OmniScore outperforms the evaluated antibody–antigen quality-assessment baselines and also improves RNA structure ranking over baselines. For peptide–major histocompatibility complex interaction prediction, the OmniScore embeddings complement structure-prediction confidence metrics to improve predictive performance. Together, these findings support OmniScore as a shared geometric foundation for biomolecular representation and scoring across molecular classes and prediction scales, which may facilitate multi-component therapeutics development.

Motivation: Ligand binding-site similarity search is a crucial step in drug discovery that reduces the conformational search space for docking and other downstream tasks by comparing a target protein against experimentally identified binding sites. Existing methods rely on either direct structural alignment or lossy compression of structural information, producing a trade-off between scalability and precision. Results: We propose LEN-Seek, a ligand binding-site search method based on a graph neural network (GNN)-driven variational autoencoder (VAE) that encodes the 3D structural and physicochemical context of a binding site into a probabilistic latent space, enabling similarity search within a low-dimensional vector space. A binding site is modeled as a graph of amino acid residues, with node features adopted from the protein language model, Ankh, and edges encoded as SE(3)-invariant (roto-translational invariant) geometric relationships, thereby avoiding expensive data augmentation or SE(3)-equivariant models. Compared to ProBiS, the purely geometric graph-clique based method, LEN-Seek successfully retrieves a substantial portion of similar binding sites with a roughly 3,400-fold lower per-comparison cost, demonstrating its potential as a scalable approach to template-based ligand binding-site search in large-scale protein structure databases. Supplementary information: Supplementary data are available at Bioinformatics online.

Despite the recent advancement of proteolysis-targeting chimera (PROTAC) development, they remain predominantly dependent on two E3 ligases, CRBN and VHL, which are ubiquitously expressed in all types of cells. Recently, efforts to discover tissue-specific E3 ligase ligands get attention as a promising strategy to enable tissue-specific protein degradation and avoid off-target tissue effects. Advancing this line of research, we discover a ligand of KLHL41, a muscle-specific E3 ligase, through virtual screening. Building on the KLHL41 ligand, we develop KBD-1, a muscle-specific BRD4-targeting PROTAC with micromolar activity. To enhance degradation efficiency, we employ a two-body kinetic strategy, resulting in the covalent PROTAC cKBD-1, which achieves sub-nanomolar activity. cKBD-1 demonstrates muscle-specific BRD4 degradation through KLHL41 recruitment both in vitro and in vivo. Moreover, the KLHL41 ligand enables AR-targeting PROTAC development, demonstrating its broad applicability. These findings highlight the potential of KLHL41 as a platform for tissue-specific protein degradation and its applicability in therapeutic development.

Artificial intelligence (AI) is reshaping biomedical research by providing scalable computational frameworks suited to the complexity of biological systems. Central to this revolution are bio/chemical language models, including large language models, which are reconceptualizing molecular structures as a form of ‘language’ amenable to advanced computational techniques. Here we critically examine the role of these models in biology and chemistry, tracing their evolution from molecular representation to molecular generation and optimization. This review covers key molecular representation strategies for both biological macromolecules and small organic compounds—ranging from protein and nucleotide sequences to single-cell data, string-based chemical formats, graph-based encodings and three-dimensional point clouds—highlighting their respective advantages and inherent limitations in AI applications. The discussion further explores core model architectures, such as bidirectional encoder representations from transformers-like encoders, generative pretrained transformer-like decoders and encoder–decoder transformers, alongside their sophisticated pretraining strategies such as self-supervised learning, multitask learning and retrieval-augmented generation. Key biomedical applications, spanning protein structure and function prediction, de novo protein design, genomic analysis, molecular property prediction, de novo molecular design, reaction prediction and retrosynthesis, are explored through representative studies and emerging trends. Finally, the review considers the emerging landscape of agentic and interactive AI systems, showcasing briefly their potential to automate and accelerate scientific discovery while addressing critical technical, ethical and regulatory considerations that will shape the future trajectory of AI in biomedicine.

Accurate prediction of protein–ligand bound poses and their affinities is essential in structure-based drug discovery. Here, we present an integrated deep-learning framework that disentangles the two core tasks─affinity estimation and pose evaluation─within complementary graph neural network architectures. BA-Pred is for predicting binding affinity and RMSD-Pred is for binding pose assessment, predicting the root-mean-squared deviation of ligand poses from crystal structures. Both models employ a Gated Graph Convolutional Network with Learnable Structural Positional Encoding (GatedGCN-LSPE) architecture to capture spatial and chemical dependencies across protein–ligand graphs. BA-Pred achieved state-of-the-art scoring power on the CASF-2016 benchmark with a root-mean-squared error of 1.10 pK_d, while RMSD-Pred exhibited strong docking power with a top-1 success rate of 96%, comparable to the best reported deep-learning scoring functions. The robust generalization capability of RMSD-Pred was further validated on the external Astex diverse set and PoseBusters benchmarks, where it significantly improved the pose selection success rates of AutoDock-GPU by up to 33.1%. The accuracy of our methodology was demonstrated on pharmaceutical targets in the 16th Critical Assessment of Structure Prediction, where our approach ranked second in the ligand binding affinity prediction category. By using our models, an integrated pipeline was developed for virtual screening, where pose selection was performed with RMSD-Pred and binding affinities were predicted with BA-Pred. This combined approach demonstrated robust screening performance, achieving an enrichment factor (EF) 1% of 21.1 on the CASF-2016 benchmark. Furthermore, on the LIT-PCBA benchmark, rescoring poses docked by AutoDock-GPU with our pipeline significantly improved the EF 1% from 2.18 to 3.19. These various benchmark results demonstrate that our graph-neural network models show good and balanced performance in diverse protein–ligand interaction prediction tasks. Thus, we expect that our models will serve as a promising framework to accelerate the drug discovery process.

Virtual screening is a critical step in computer-aided drug discovery. As synthetic databases and reaction schemes expand, the number of synthesizable compounds grows, making efficient screening essential. The success of ultra-large-scale virtual screening hinges on two factors: accuracy in binding affinity prediction and speed. We previously developed AK-Score2 for accurate protein-ligand binding prediction and V-Dock for estimating docking scores. Here, we present Docking of Millions (DoM), an ultra-large-scale virtual screening system integrating both methods. Iterative learning of V-Dock to approximate AK-Score2-based affinities accelerates screening by eliminating the need to dock all compounds, saving time and resources. In benchmarking, screening 5 million compounds against DDR1, c-kit, ASK1, NSD1, CREBBP, and PDE5, DoM consumed only 319 h on average, 12% of full library screening time. Also, it achieved an average retrieval rate of 89% in the top 100 compounds. Inhibition assays identified that 1 out of 27 and 4 out of 31 molecules suppressed ASK1 and DDR1 activity by more than 50% at 10 μM. Further validation of selected compounds identified that the lowest IC50 values are 1.96 μM and 788 nM for ASK1 and DDR1, respectively. These findings demonstrate DoM as a practical and efficient platform for real-world drug discovery.

Accurate identification of protein-ligand binding sites is an essential step in structure-based drug discovery. Herein, we present SwinSite, a deep learning framework that leverages a hybrid architecture combining 3D convolutional neural networks and hierarchical vision transformer modules to predict ligand binding sites based on a 3D structure of a target protein. SwinSite encodes spatial information by voxelizing a protein structure into 3D grids centered around surface residues, allowing for a detailed spatial representation of the protein's surface environment. By combining local feature extraction with hierarchical self-attention via shifted windows, SwinSite effectively captures both fine-grained geometric features and long-range dependencies. Evaluations on multiple benchmark data sets demonstrate that SwinSite outperforms existing CNN- and GNN-based ligand binding site detection methods consistently, highlighting its robustness and generalization ability.

Motivation: Predicting immunoglobulin–antigen (Ig–Ag) binding remains a significant challenge due to the paucity of experimentally resolved complexes and the limited accuracy of de novo Ig structure prediction. Results: We introduce IgPose, a generalizable framework for Ig–Ag pose identification and scoring, built on a generative data-augmentation pipeline. To mitigate data scarcity, we constructed the Structural Immunoglobulin Decoy Database (SIDD), a comprehensive repository of high-fidelity synthetic decoys. IgPose integrates equivariant graph neural networks, ESM-2 embeddings, and gated recurrent units to synergistically capture both geometric and evolutionary features. We implemented interface-focused k-hop sampling with biologically guided pooling to enhance generalization across diverse interfaces. The framework comprises two sub-networks—IgPoseClassifier for binding pose discrimination and IgPoseScore for DockQ score estimation—and achieves robust performance on curated internal test sets and the CASP-16 benchmark compared to physics and deep learning baselines. IgPose serves as a versatile computational tool for high-throughput antibody discovery pipelines by providing accurate pose filtering and ranking. Availability and implementation: IgPose is available on GitHub (https://github.com/arontier/igpose).