Biology subjects

Furui, K.

Publications and source records attributed to Furui, K..

3 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

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

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↗