Dongsup Kim

Motivation Recent advances in biomolecular structure prediction, exemplified by AlphaFold3, have opened new opportunities for the prediction of TCR-pMHC binding specificity. Although individual AlphaFold3 confidence metrics provide informative binding signals, their predictive performance varies across datasets, highlighting the need to combine complementary signals rather than rely on any single metric. Results We present TAPAS, a tabular learning framework that integrates AlphaFold3-derived interface confidence and structural geometry with sequence embeddings. Although no single zero-shot metric was the strongest across all benchmarks, TAPAS was consistently the top-ranked method on VDJdb and two external benchmarks, matching or exceeding the strongest zero-shot AlphaFold3 metric. Feature group ablation showed that the contributions of sequence, confidence, and geometric features varied across evaluation settings, and that combining them resulted in the best overall performance. These results highlight the value of integrating complementary structural and sequence features within a unified framework for robust TCR-pMHC binding prediction.

Motivation: Understanding temporal gene expression is fundamental in the study of cellular development and differentiation. In practice, temporal single-cell datasets tend to contain only a limited number of measured time points, which are often unevenly spaced, resulting in irregular intervals between observations due to experimental constraints. Existing methods typically address these intervals by sequentially predicting one time point after another, yet lack mechanisms to explicitly model time intervals, leading to error accumulation. Results: In this work, we introduce scMix, a language-model-based framework for predicting single-cell gene expression, which enables prediction from multiple historical time points. We build scMix on the Receptance Weighted Key Value architecture and use its time decay mechanism to model temporal dependencies over time. Moreover, scMix proposes a delta-time mechanism that allows the model to bypass unmeasured time points, reducing error accumulation and improving robustness. In addition, we incorporate a trend regularization strategy to enhance the temporal coherence of predicted gene expression trajectories. scMix demonstrates state-of-the-art performance in predicting gene expression at unmeasured time points, surpassing existing methods, and also achieves outstanding results on downstream tasks. Availability and implementation: The code used for this study is available at https://doi.org/10.5281/zenodo.18287184.

Knowledge of B cell epitopes is critical to vaccine design, diagnostics, and therapeutics. As experimental validation for epitopes is time-consuming and costly, many in silico tools have been developed to computationally predict the B cell epitopes. While most methods show poor performance, deep learning methods in recent years have shown promising results. We developed a method called EpiGraph that outperformed previous methods, including those that showed a significant improvement in performance in recent years. Our model’s performance can be attributed to the following factors: (1) a combination of structure and sequence feature embeddings obtained from pretrained ESM-IF1 and ESM-2 models could capture the structural and evolutionary features of B cell epitopes, (2) a graph attention network could learn the spatial proximity of B cell epitopes with high graph homophily, and (3) residual connections in the model framework mitigate the over-smoothing problem in the graph neural network. Our model achieved the highest performance on an independent benchmark dataset. The results were also consistent on a different dataset. The datasets and source codes are available at https://github.com/sj584/EpiGraph. A user-friendly web server is freely available at http://epigraph.kaist.ac.kr.

Minimum inhibitory concentration (MIC) denotes the in vitro benchmark indicating the quantity of antibiotic required to inhibit proliferation of specific bacterial strains. Determining MIC values corresponding to the infecting bacterial strain is paramount for tailoring appropriate antibiotic therapy. In the interim between specimen collection and laboratory-derived MIC outcomes, clinicians frequently resort to empirical therapy informed by retrospective analyses. Here introduces two deep learning approaches, a Convolutional Neural Network (CNN)-based model and an Enformer-based model, integrating genomic data of Klebsiella Pneumoniae and molecular structural data of 20 antibiotics to anticipate the MIC value of the bacterium for each antibiotic under consideration. These models demonstrate enhanced raw accuracy over the existing state-of-the-art model, which rely exclusively on genomic data. The CNN-based model achieves a notable 20% increase in raw accuracy while further mirroring the 1-tier accuracy of the state-of-the-art model. Although the Enformer-based model does not quite reach the performance levels of the CNN-based model, it offers an advantage by eliminating the need for arbitrary data processing steps. This streamlining of the data processing pipeline facilitates fast updates and improves the model interpretability. It is expected that these deep learning paradigms can significantly inform and bolster clinician decision-making during the empirical treatment phase.

Motivation: Prediction of T-cell receptor (TCR)–epitope interactions is important for many applications in biomedical research, such as cancer immunotherapy and vaccine design. The prediction of TCR–epitope interactions remains challenging especially for novel epitopes, due to the scarcity of available data. Results: We propose TSpred, a new deep learning approach for the pan-specific prediction of TCR binding specificity based on paired chain TCR data. We develop a robust model that generalizes well to unseen epitopes by combining the predictive power of CNN and the attention mechanism. In particular, we design a reciprocal attention mechanism which focuses on extracting the patterns underlying TCR–epitope interactions. Upon a comprehensive evaluation of our model, we find that TSpred achieves state-of-the-art performances in both seen and unseen epitope specificity prediction tasks. Also, compared to other predictors, TSpred is more robust to bias related to peptide imbalance in the dataset. In addition, the reciprocal attention component of our model allows for model interpretability by capturing structurally important binding regions. Results indicate that TSpred is a robust and reliable method for the task of TCR–epitope binding prediction. Availability and implementation: Source code is available at https://github.com/ha01994/TSpred.

Introduction: Generative AI models have been introduced to find novel chemical compounds for drug discovery. To evaluate different generative AI models quantitatively, computational metrics such as rule of five, validity, novelty, uniqueness and FCD are needed. However, conventional metrics still have limitations in terms of comprehensive evaluation including drug-likeness, diversity, and novelty. Method: In this paper, we propose a metric, novelty, and coverage (NC), which measures the structural similarity between the generated set and the actual ligand set. Novelty and coverage filters the generated set by comparing molecular properties with known drugs, and then calculates structural similarity and diversity, allowing a comprehensive evaluation based on the trade-off relationship between them. For the comparison of conventional metrics and NC, we used models from the MOSES platform in addition to the latest deep learning models. We built the two representative sets for the comparison of molecular properties (ChEMBL) and structural similarity (ZINC). Result: The results show that conventional metrics can be grouped into three clusters (drug-likeness, external diversity, internal diversity) and point out their limitations. We compared the results of conventional metrics and NC, and then verified that NC explained conventional metrics. NC explained comprehensive evaluation in drug discovery with harmonic mean between contrasting features in chemical generative models. Availability and implementation: Codes are freely available to non-commercial users at https://github.com/KyoungYeulLee/drug_novelty_and_coverage.

Motivation: Antibodies are proteins that the immune system produces in response to foreign pathogens. Designing antibodies that specifically bind to antigens is a key step in developing antibody therapeutics. The complementarity determining regions (CDRs) of the antibody are mainly responsible for binding to the target antigen, and therefore must be designed to recognize the antigen. Results: We develop an antibody design model, AbFlex, that exhibits state-of-the-art performance in terms of structure prediction accuracy and amino acid recovery rate. Furthermore, >38% of newly designed antibody models are estimated to have better binding energies for their antigens than wild types. The effectiveness of the model is attributed to two different strategies that are developed to overcome the difficulty associated with the scarcity of antibody–antigen complex structure data. One strategy is to use an equivariant graph neural network model that is more data-efficient. More importantly, a new data augmentation strategy based on the flexible definition of CDRs significantly increases the performance of the CDR prediction model. Availability and implementation: The source code and implementation are available at https://github.com/wsjeon92/AbFlex.

Prediction of T-cell receptor (TCR)-epitope interactions is important for many applications such as cancer immunotherapy. However, due to the scarcity of available data, it is known to be a challenging task particularly for novel epitopes. Here, we propose TSpred, a new ensemble deep learning approach for the pan-specific prediction of TCR binding specificity based on paired chain TCR data. This method combines the predictive power of CNN and the attention mechanism to capture the patterns underlying TCR-epitope interactions. In particular, we design a reciprocal attention mechanism which contributes to higher model generalizability to unseen epitopes. We perform a comprehensive evaluation of our model and observe that TSpred achieves state-of-the-art performances in both seen and unseen epitope specificity prediction tasks. Our model performs consistently well across both of the two widely used negative sampling strategies, while avoiding the potential bias associated with each strategy. Also, compared to other predictors, it is more robust to bias related to peptide imbalance in the dataset. In addition, the reciprocal attention component of our model allows for model interpretability by capturing structurally important binding regions. Results indicate that TSpred is a robust and reliable method for the task of TCR-epitope binding prediction.

Knowledge of B cell epitopes is crucial for vaccine design, diagnostics, and therapeutics. Many in silico tools have been developed to computationally predict the B cell epitope. However, most methods have shown inconsistent performance, thereby degrading the reliability of the predictions. To address this challenge, we developed EpiCluster, an end-to-end deep learning model that significantly outperforms existing methods by a large margin. Our model’s performance is consistent with several benchmark datasets, including the most recent one on which all existing methods performed very poorly. EpiCluster achieves this mainly through two ways. First, it effectively combines the structural and evolutionary features of epitopes. Second, it has the model architecture that exploits the clustering property of epitopes. More importantly, we have demonstrated that an end-to-end learning model architecture enforcing the clustering property of epitopes was critically important for building an accurate epitope prediction model. The source code and implementation are available at https://github.com/sj584/EpiCluster.

Protein complex structure prediction is important for many applications in bioengineering. A widely used method for predicting the structure of protein complexes is computational docking. Although many tools for scoring protein-protein docking models have been developed, it is still a challenge to accurately identify near-native models for unknown protein complexes. A recently proposed model called the Geometric Vector Perceptron-Graph Neural Network (GVP-GNN), a subtype of Equivariant Graph Neural Networks (EGNNs), has demonstrated success in various 3D molecular structure modeling tasks. Herein, we present G-RANK, a GVP-GNN-based method for the scoring of protein-protein docking models. When evaluated on two different test datasets, G-RANK achieved a performance competitive with or better than the state-of-the-art scoring functions. We expect G-RANK to be a useful tool for various applications in biological engineering. Source code is available at https://github.com/ha01994/grank. Supplementary data are available at Bioinformatics Advances online.