Biology subjects

Ohue, M.

Publications and source records attributed to Ohue, M..

14 recordsLinked to original sources

bioRxiv2025

ALLM-Ab: Active learning-driven antibody optimization using fine-tuned protein language models

Antibody engineering requires a delicate balance between enhancing binding affinity and maintaining developability properties. In this study, we present ALLM-Ab (Active Learning with Language Models for Antibodies), a novel active learning framework that leverages fine-tuned protein language models to accelerate antibody sequence optimization. By employing parameter-efficient fine-tuning via low-rank adaptation, coupled with a learning-to-rank strategy, ALLM-Ab accurately assesses mutant fitness while efficiently generating candidate sequences through direct sampling from the models probability distribution. Furthermore, by integrating a multi-objective optimization scheme incorporating antibody developability metrics, the framework ensures that optimized sequences retain therapeutic antibody-like properties alongside improved binding affinity. We validate ALLM-Ab in both offline experiments using deep mutational scanning (DMS) data from the BindingGYM dataset and online active learning trials targeting Flex ddG energy minimization across three antigens. Results demonstrate that ALLM-Ab not only expedites the discovery of high-affinity variants compared to baseline Gaussian process regression and genetic algorithm-based approaches, but also preserves critical antibody developability metrics. This work lays the foundation for more efficient and reliable antibody design strategies, with the potential to significantly reduce therapeutic development costs.

bioinformatics↗

bioRxiv2025

Protein-ligand affinity prediction via Jensen-Shannon divergence of molecular dynamics simulation trajectories

Predicting the binding affinity between proteins and ligands is a critical task in drug discovery. Although various computational methods have been proposed to estimate ligand target affinity, the method of Yasuda et al. (2022) ranks affinities based on the dynamic behavior obtained from molecular dynamics (MD) simulations without requiring structural similarity among ligand substituents. Thus, its applicability is broader than that of relative binding free energy calculations. However, their approach suffers from high computational costs due to the extensive simulation time and the deep learning computations needed for each ligand pair. Moreover, in the absence of experimental {Delta}G values (oracle), the sign of the correlation can be misinterpreted. In this study, we present improvements to Yasuda et al.s method. Our contributions are threefold: (1) By introducing the Jensen-Shannon (JS) divergence, we eliminate the need for deep learning-based similarity estimation, thereby significantly reducing computation time; (2) We demonstrate that production run simulation times can be halved while maintaining comparable accuracy; and (3) We propose a method to predict the sign of the correlation between the first principal component (PC1) and {Delta}G by using coarse {Delta}G estimations obtained via AutoDock Vina.

bioinformatics↗

bioRxiv2025

Generation of appropriate protein structures for virtual screening using AlphaFold3 predicted protein-ligand complexes

In early drug discovery, virtual screening--a computational method for selecting candidate compounds--helps reduce development costs. Traditionally, structure-based virtual screening required experimental protein structures, but advances like AlphaFold2 have begun to overcome this limitation. However, AlphaFold2 does not capture ligand-induced conformational changes (transition from apo to holo forms), limiting its utility for protein-ligand docking. In this study, we evaluate AlphaFold3, which predicts protein-ligand complex structures when both protein and ligand inputs are provided. Using the DUD-E dataset and Uni-Dock, we show that holo structures predicted with ligand inclusion yield higher screening performance than apo structures generated without ligand input. Notably, incorporating active ligands enhances screening performance, whereas inactive (decoy) ligands produce results similar to apo predictions. The use of template structures further improves outcomes. We also analyze the impact of ligand molecular weight, binding pocket location, and AlphaFold3 ranking scores on screening performance. Our findings indicate that lower molecular weight ligands tend to generate predicted structures that more closely resemble experimental holo structures, thus improving screening efficacy. Conversely, larger ligands (700-800) can induce open binding pockets that favor screening for some targets. These results suggest that employing AlphaFold3 with appropriate ligand inputs is a promising strategy for virtual screening, particularly for proteins lacking experimental structural data.

bioinformatics↗

bioRxiv2025

Leveraging AlphaFold2 structural space exploration for generating drug target structures in structure-based virtual screening

In early drug discovery, computational virtual screening (VS) is vital for selecting candidate compounds and reducing costs. However, the lack of experimentally determined 3D structures has limited the application of structure-based VS. Advances in protein structure prediction, notably AlphaFold2, have begun to address this gap. Yet, studies indicate that direct use of AlphaFold2-predicted structures often leads to suboptimal VS performance--likely because these structures fail to capture ligand-induced conformational changes (apo-to-holo transitions). To overcome this, we propose an approach that explores and modifies the structural space of AlphaFold2 predictions to generate conformations more amenable to VS. Our method deliberately alters the multiple sequence alignment (MSA) by introducing alanine mutations at key residues in the ligand-binding site, thereby inducing significant conformational shifts. The exploration process is guided by iterative ligand docking simulations, with mutation strategies optimized either by a genetic algorithm or via random search. Our evaluation shows that when sufficient active compounds are available, the genetic algorithm significantly enhances VS accuracy. In contrast, with limited active compound data, a random search strategy proves more effective. Moreover, our approach is particularly promising for targets that yield poor screening results when using experimentally determined structures from the PDB. Overall, these findings underscore the practical utility of modified AlphaFold2-derived structures in VS and expand the potential of computationally predicted protein models in drug discovery.

bioinformatics↗

bioRxiv2024

SpatialPPI 2.0: Enhancing Protein-Protein Interaction Prediction through Inter-Residue Analysis in Graph Attention Networks

Protein-protein interactions (PPIs) are fundamental to cellular functions, and accurate prediction of these interactions is crucial to understanding biological mechanisms and facilitating drug discovery. SpatialPPI 2.0 is an advanced graph neural network-based model that predicts PPIs by utilizing interresidue contact maps derived from both structural and sequence data. By leveraging the comprehensive PINDER dataset, which includes interaction data from the RCSB PDB and the AlphaFold database, SpatialPPI 2.0 improves the specificity and robustness of the prediction of PPI. Unlike the original SpatialPPI, the updated version employs interaction interface prediction as an intermediate step, allowing for a more effective assessment of interactions between isolated proteins. The model utilizes Graph Attention Networks (GAT) and Graph Convolutional Networks (GCN) to capture both local and global structural features. SpatialPPI 2.0 outperforms several state-of-the-art PPI and interface predictors, demonstrating superior accuracy and reliability. Furthermore, the model shows robustness when using structures predicted by AlphaFold, indicating its potential to predict interactions for proteins without experimentally determined structures. SpatialPPI 2.0 offers a promising solution for the accurate prediction of PPIs, providing insight into protein function and supporting advances in drug discovery and synthetic biology. SpatialPPI 2.0 is available at https://github.com/ohuelab/SpatialPPI2.0

bioinformatics↗

bioRxiv2024

Benchmarking HelixFold3-Predicted Holo Structures for Relative Free Energy Perturbation Calculations

AlphaFold2 demonstrated remarkable capabilities for protein structure prediction. However, it is limited to downstream tasks, such as ligand docking and free energy calculations, as it cannot predict holo structures with bound ligands. AlphaFold3, a state-of-the-art protein structure prediction model, can predict the binding structures of complexes with proteins, nucleic acids, small molecules, ions, and modified residues with cutting-edge performance. However, AlphaFold3 does not currently provide access to some functions, such as the prediction of protein-ligand complex structures. To reduce the enormous costs in early small molecule drug discovery, verifying the utility of protein-ligand complex prediction methods, such as AlphaFold3, in downstream tasks like free energy perturbation calculations is crucial. In this study, we evaluated Helix-Fold3, designed to emulate AlphaFold3, in predicting holo and apo structures complex formations and examined its utility in free energy perturbation calculations. Regarding the complex structure prediction performance of the 8 targets from Wang et al.s FEP benchmark, HelixFold3 showed superior performance to AlphaFold2 and existing methods. Predicting a holo structure rather than an apo structure resulted in higher binding site prediction accuracy. Furthermore, using HelixFold3 predicted structures in practical situations, where binding free energies of all derivatives were estimated, both structures achieved accuracies comparable to crystal structures. Additionally, novel derivatives not included in the training data were accurately predicted, demonstrating that free energy calculations using these novel structures are sufficiently usable.

bioinformatics↗

bioRxiv2023

SpatialPPI: three-dimensional space protein-protein interaction prediction with AlphaFold Multimer

The rapid advancement of protein sequencing technology has resulted in a gap between proteins with identified sequences and those with mapped structures. Although sequence-based predictions offer insights, they can be incomplete due to the absence of structural details. On the other hand, structure-based methods face challenges with newly sequenced proteins. AlphaFold emerges as a potential solution, especially in predicting protein-protein interactions. This research delves into using deep neural networks, specifically analyzing protein complex structures as predicted by AlphaFold Multimer. By transforming atomic coordinates and utilizing sophisticated image processing techniques, the study extracts vital 3D structural details from protein complexes. Recognizing the significance of evaluating residue distances in protein interactions, the study leverages image recognition approaches, notably integrating DenseNet and ResNet within 3D convolutional networks for protein 3D structure analysis. When benchmarked against leading protein-protein interaction prediction methods like SpeedPPI, D-script, DeepTrio, and PEPPI, our proposed method, named SpatialPPI, exhibits notable efficacy, emphasizing the promising role of 3D spatial processing in advancing the realm of structural biology. O_FIG O_LINKSMALLFIG WIDTH=200 HEIGHT=109 SRC="FIGDIR/small/571766v1_ufig1.gif" ALT="Figure 1"> View larger version (32K): org.highwire.dtl.DTLVardef@14fa761org.highwire.dtl.DTLVardef@761293org.highwire.dtl.DTLVardef@1bd9508org.highwire.dtl.DTLVardef@1837a8e_HPS_FORMAT_FIGEXP M_FIG C_FIG

bioinformatics↗

bioRxiv2023

Design of Cyclic Peptides Targeting Protein-Protein Interactions using AlphaFold

More than 930,000 protein-protein interactions (PPIs) have been identified in recent years, but their physicochemical properties differ from conventional drug targets, complicating the use of conventional small molecules as modalities. Cyclic peptides are a promising modality for targeting protein-protein interactions (PPIs), but it is difficult to predict the structure of a target protein-cyclic peptide complex or to design a cyclic peptide sequence that binds to the target protein using computational methods. Recently, AlphaFold with a cyclic offset has enabled predicting the structure of cyclic peptides, thereby enabling de novo cyclic peptide designs. We developed a cyclic peptide complex offset to enable the structural prediction of target proteins and cyclic peptide complexes and found AlphaFold2 with a cyclic peptide complex offset can predict structures with high accuracy. We also applied the cyclic peptide complex offset to the binder hallucination protocol of AfDesign, a de novo protein design method using AlphaFold, and we could design a high predicted local-distance difference test and lower separated binding energy per unit interface area than the native MDM2/p53 structure. Furthermore, the method was applied to 12 other protein-peptide complexes and one protein-protein complex. Our approach shows that it is possible to design putative cyclic peptide sequences targeting PPI.

bioinformatics↗

bioRxiv2023

Antibody Complementarity-Determining Region Sequence Design using AlphaFold2 and Binding Affinity Prediction Model

Affinity maturation in the immune response is limited in terms of affinity gain, and natural antibodies often do not have the binding affinity required for therapeutic applications. Therefore, improving the binding affinity of antibodies is essential for developing antibody-based therapeutics. Designing antibodies using experimental methods is expensive in terms of cost and time owing to the large range of complementarity-determining regions to be explored. Recently, computational methods have been developed as low-cost and fast means of designing and redesigning antibodies. This study evaluated the design performance of AlphaFold2 and the binder hallucination, which can predict protein 3D structures with high accuracy even without experimental antibodies, by redesigning antibody sequences to improve the binding affinity of existing antigen-antibody complexes. Therefore, antibody sequences with higher affinity can be designed for antigen-antibody complexes not included in the training data of AlphaFold2, indicating that the proposed method may be effective as an antibody design method.

bioinformatics↗

bioRxiv2022

Solubility-aware protein binding peptide design using AlphaFold

New protein-protein interactions (PPIs) are being identified, but PPIs have different physicochemical properties compared with conventional targets, making it difficult to use small molecules. Peptides offer a new modality to target PPIs, but designing appropriate peptide sequences by computation is challenging. Recently, AlphaFold and RosettaFold have made it possible to predict protein structures from amino acid sequences with ultra-high accuracy, enabling de novo protein design. We designed peptides likely to have PPI as the target protein using the "binder hallucination" protocol of AfDesign, a de novo protein design method using AlphaFold. However, the solubility of the peptides tended to be low. Therefore, we designed a solubility loss function using solubility indices for amino acids and developed a solubility-aware AfDesign binder hallucination protocol. The peptide solubility in sequences designed using the new protocol increased with the weight of the solubility loss function; moreover, they captured the characteristics of the solubility indices. Moreover, the new protocol sequences tended to have higher affinity than random or single residue substitution sequences when evaluated by docking binding affinity. Our approach shows that it is possible to design peptide sequences that can bind to the interface of PPI while controlling solubility.

bioinformatics↗

bioRxiv2021

Distinct and interchangeable growing patterns in colorectal cancer stem-like cells are regulated by Musashi-1

The dynamic and heterogeneous features of cancer stem-like cells (CSCs) have been widely recognized, but their nongenetic cellular plasticity mechanisms remain elusive. By using colorectal cancer organoids, we phenotypically tracked their spheroid formation and growth capacity to a single-cell resolution, and we discovered that the spheroid-forming cells exhibit a heterogeneous growth pattern, consisting of slow- and fast-growing spheroids. The isolated fast-growing spheroids seem to preserve a dual-growing pattern through multiple passages, whereas the isolated slow-growing spheroids are restricted to a slow-growing pattern. Notably, the spheroids of both patterns were tumorigenic. Moreover, the expression of CSC markers varied among the subpopulations with different growth patterns. The isolated slow-growing spheroids adopted the dual-growing pattern by various extrinsic triggers, in which Musashi-1 plays a key role. The slow-growing fraction was resistant to chemotherapy, and its successful isolation can provide an in vitro platform allowing us to elucidate their role in drug resistance.

cancer biology↗

bioRxiv2020

Structural basis for binding mechanism of human serum albumin complexed with cyclic peptide dalbavancin

Cyclic peptides, with unique structural features, have emerged as new candidates for drug discovery; their association with human serum albumin (HSA; long blood half-life), is crucial to improve drug delivery and avoid renal clearance. Here, we present the crystal structure of HSA complexed with dalbavancin, a clinically used cyclic peptide. SAXS and ITC experiments showed that the HSA-dalbavancin complex exists in a monomeric state; dalbavancin is only bound to the subdomain IA of HSA in solution. Structural analysis and MD simulation revealed that the swing of Phe70 and movement of the helix near dalbavancin were necessary for binding. The flip of Leu251 promoted the formation of the binding pocket with an induced-fit mechanism; moreover, the movement of the loop region including Glu60 increased the number of non-covalent interactions with HSA. These findings may support the development of new cyclic peptides for clinical use, particularly the elucidation of their binding mechanism to HSA.

biochemistry↗

bioRxiv2020

Re-ranking of computational protein-peptide docking solutions with amino acid profiles of rigid-body docking results

Protein-peptide interactions, in which one partner is a globular protein and the other is a flexible linear peptide, are important for understanding cellular processes and regulatory pathways, and are therefore targets for drug discovery. In this study, I combined rigid-body protein-protein docking software (MEGADOCK) and global flexible protein-peptide docking software (CABS-dock) to establish a re-ranking method with amino acid contact profiles using rigid-body sampling decoys. I demonstrate that the correct complex structure cannot be predicted (< 10 [A] peptide RMSD) using the current version of CABS-dock alone. However, my newly proposed re-ranking method based on the amino acid contact profile using rigid-body search results (designated the decoy profile) demonstrated the possibility of improvement of predictions. Adoption of my proposed method along with continuous efforts for effective computational modeling of protein-peptide interactions can provide useful information to understand complex biological processes in molecular detail and modulate protein-protein interactions in disease treatment.

bioinformatics↗

bioRxiv2020

Rescoring ensembles of protein-protein docking poses using consensus approaches

Scoring is a challenging step in protein-protein docking, where typically thousands of solutions are generated. Successful scoring is more often based on physicochemical evaluation of the generated interfaces and/or statistical potentials that reproduce known interface properties. Another route is offered by consensus-based rescoring, where the set of solutions is used to build statistics in order to identify recurrent solutions. We explore several ways to perform consensus-based rescoring on the ZDOCK decoy set for Benchmark 4. We show that the information of the interface size is critical for successful rescoring. We combine consensus-based rescoring with the ZDOCK native scoring function and show that this improves the initial results.

bioinformatics↗