The structural diversity of chemical libraries, which are systematic collections of compounds that have potential to bind to biomolecules, can be represented by chemical latent space. A chemical latent space is a projection of a compound structure into a mathematical space based on several molecular features, and it can express structural diversity within a compound library in order to explore a broader chemical space and generate novel compound structures for drug candidates. In this study, we developed a deep-learning method, called NP-VAE (Natural Product-oriented Variational Autoencoder), based on variational autoencoder for managing hard-to-analyze datasets from DrugBank and large molecular structures such as natural compounds with chirality, an essential factor in the 3D complexity of compounds. NP-VAE was successful in constructing the chemical latent space from large-sized compounds that were unable to be handled in existing methods, achieving higher reconstruction accuracy, and demonstrating stable performance as a generative model across various indices. Furthermore, by exploring the acquired latent space, we succeeded in comprehensively analyzing a compound library containing natural compounds and generating novel compound structures with optimized functions.

Free energy perturbation (FEP) calculations are a powerful tool for predicting binding affinities in drug discovery, but their accuracy heavily depends on accurate protein-ligand complex structures. While AlphaFold2 revolutionized protein structure prediction, its inability to predict holo structures limits its application in structure-based drug design. AlphaFold3 and its reproduction HelixFold3 demonstrated the ability to predict protein complexes with various binding partners, including small molecules. In this study, we evaluated HelixFold3's ability to predict protein-ligand complexes using eight targets from Wang et al.'s FEP benchmark set. Our analysis revealed that HelixFold3 outperformed the existing methods, including AlphaFold2, in predicting binding site conformations. Notably, the prediction of holo structures yielded a higher binding site accuracy compared to apo structures. FEP calculations using both HelixFold3-predicted holo and apo structures achieved accuracy comparable to that of calculations using crystal structures. Furthermore, HelixFold3 successfully predicted complex structures for novel derivatives not present in its training data, and FEP calculations using these predicted structures maintained reliable accuracy. These results suggest that HelixFold3-predicted structures can effectively substitute for crystal structures in early stage drug discovery.

Protein language models (pLMs) are rapidly emerging as revolutionary artificial intelligence technologies that bring transformative changes to drug discovery and therapeutic research. pLMs acquire rich representational capabilities from large-scale sequence datasets, enabling the solution of various biological problems that were difficult with conventional methods. In this review, we provide a comprehensive overview of various pLMs and their implementations, exploring their potential utility in drug discovery and therapeutic research. First, we systematically classify pLMs based on their architectures and information sources while discussing their development to the present. We also explain recent trends in multimodal approaches that integrate co-evolutionary information, structural information, and functional information, as well as domain-specific models specialized for particular domains such as antibodies and T-cell receptors. We then provide a comprehensive overview of various therapeutic applications of pLMs, including mutation effect prediction, function prediction, and structure prediction. Finally, we discuss future prospects of pLMs toward therapeutic applications and challenges for transforming them into technologies that contribute to actual diseases.

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 model's probability distribution. Furthermore, by integrating a multiobjective 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 data set and online active learning trials targeting Flex ddG energy minimization across 15 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.

In recent years, free energy perturbation calculations have garnered increasing attention as tools to support drug discovery. The lead optimization mapper (Lomap) was proposed as an algorithm to calculate the relative free energy between ligands efficiently. However, Lomap requires checking whether each edge in the FEP graph is removable, which necessitates checking the constraints for all edges. Consequently, conventional Lomap requires significant computation time, at least several hours for cases involving hundreds of compounds, and is impractical for cases with more than tens of thousands of edges. In this study, we aimed to reduce the computational cost of Lomap to enable the construction of FEP graphs for hundreds of compounds. We can reduce the overall number of constraint checks required from an amount dependent on the number of edges to one dependent on the number of nodes by using the chunk check process to check the constraints for as many edges as possible simultaneously. Based on the analysis of the execution profiles, we also improved the speed of cycle constraint and diameter constraint checks. Moreover, the output graph is the same as that obtained using the conventional Lomap, enabling direct replacement of the original one with our method. With our improvement, the execution was hundreds of times faster than that of the original Lomap.

Natural products are an important source of medicines, yet their discovery can be a slow and laborious process.