The discovery of novel compounds for protein targets is an important step in the drug discovery process. Target-specific molecules can be designed based on protein structures, but molecule prediction based on protein sequences alone remains a challenge. We hypothesize that including information on noninteracting pairs can improve the specificity and generalizability of molecule prediction models. To this end, we introduce ConGen, a sequence-conditioned framework for targeted molecule generation from protein sequence alone. To our knowledge, it is the first model that combines contrastive protein-molecule representation learning with interacting and noninteracting molecules and latent optimization (LO) in a shared pretrained space, addressing the limitations of interacting-only training. ConGen follows two stages: (i) contrastive learning (CL) maps proteins and molecules into a joint space where interacting pairs are pulled together and noninteracting pairs are pushed apart and (ii) LO starts from random molecule latents, moves them toward the target protein's embedding, and decodes a local neighborhood to propose candidates. This approach enables the generation of molecules specifically designed for a given target by efficiently exploiting the structure of the shared latent space. Our results show that ConGen achieves performance comparable to EncDecLM, a state-of-the-art sequence-based model for targeted drug generation. An ablation study further demonstrates the contributions of both the CL and LO stages.

Efficient exploration of chemical space is an essential component of modern generative drug design. Herein, we introduce ChemBang, a computational engine that grows small molecules based on chemical transformations extracted by matched molecular pair analysis of all structures available in catalogues of synthesized molecules. Each chemical transformation is mapped onto its associated atomic environment defined as the substructure within a three-atom radius from the transformation site. Unsupervised chemical evolution is then performed in cycles by systematically applying chemical transformations to all exposed atomic environments present in a seed structure. Multiple physicochemical properties and substructural alerts are incorporated to effectively guide the generation of drug-like synthetically accessible molecules. As a use case, the generation of the Erdafitinib structure from any of its three ring systems (pyrazole, benzene and quinoxaline), and the evolution of the property distributions from all molecules generated in each cycle, are discussed in detail. The ability to explore the chemical space of pharmaceutical relevance is shown by successfully generating the exact chemical structure of 95.3% of all 2,809 small-molecule ATC drugs from their constituting fragments.

Accurately predicting the fraction unbound in plasma (fup) from chemical structures is essential for understanding pharmacokinetic characteristics during the early stage of drug discovery. This prediction serves as a valuable tool for minimizing late-stage setbacks and refining subsequent screening processes. Conventional approaches often rely on complex computational methodologies that may require extensive descriptor sets, resulting in opaque models with limited interpretability. In this study, we applied the read-across strategy in combination with traditional quantitative structure-property relationship to predict fup while minimizing descriptor complexity. Our method employs interpretable models (regression and classification), facilitating insight into the underlying structure-property relationships governing plasma protein binding. Through comprehensive validation and comparison with different machine learning methods, we demonstrated the superior predictive performance of quantitative read-across structure-property relationship multiple linear regression and classification-based read-across structure-property relatonship respectively. support vector classifier models across diverse chemical compounds. This approach offers a valuable tool for predicting fup in the process of drug discovery. Overall, this study aims to advance the field of pharmacokinetic modeling by applying the read-across strategy that improves predictive power with interpretability. By elucidating the complex relationship between chemical structures and fup, our best models have the potential to formulate more rational drug design approaches, ultimately contributing to the development of more effective therapeutics.

Chemical language models (CLMs), particularly encoder‐decoder transformers, have advanced generative molecular design. Transformer CLMs are able to learn a variety of molecular mappings for compound design that can be conditioned using context‐dependent rules. However, their black‐box nature complicates the interpretation of predictions. Current analysis methods mostly focus on attention weights of token relationships or attention flow in encoder and decoder modules and cannot explain predictions at the molecular level. Sequence‐based compound design was used as a model system to investigate transformer learning characteristics through systematic control calculations involving modifications of protein sequences and sequence‐compound pairs. The analysis revealed that compound reproducibility depended on similarity relationships between training and test data and on compound memorization, while specific sequence information was not learned. These findings indicate that predictions of transformer CLMs are driven by memorization effects and statistical correlations rather than by learning specific chemical or biological information. Understanding this learning behavior aids in avoiding over‐interpretation of model outputs and informs the appropriate application of transformer‐based CLMs in molecular design.

Chemical space [1, 2], the abstract universe of all possible molecules, remains a central concept in cheminformatics [3], drug discovery, and several other research areas. With estimates extending far beyond 1060 compounds, its exploration represents both an opportunity and an outstanding challenge. The increasing availability of computational power, innovative algorithms, and constantly increasing, massive compound libraries are opening new avenues for navigating this vast space. However, the complexity of the task—from dataset curation and representation to the reliable prediction of bioactivity—requires a careful balance between scale, diversity, and feasibility. Ultralarge chemical spaces demand not only efficient algorithms for screening but also rigorous strategies to ensure that the generated or selected molecules are synthetically accessible and hold the desired property, for example, that are biologically relevant [4]. In this context, this special issue of Molecular Informatics brings together six contributions that showcase the current progress in the exploration of (ultra)big chemical spaces. The works range from comparative analyses of natural and synthetic compound collections to the definition and practical advances of ultralarge virtual screening. They further illustrate methodological innovations, including new approaches to address activity cliffs in large datasets, benchmarking of dimensionality reduction techniques for chemical space visualization, and generative algorithms for exploring unprecedented peptide–peptoid libraries comprising up to 1060 members. Importantly, the special issue also introduces novel chemical spaces explicitly designed to expand the reach of existing resources, such as the Freedom Space, a synthetically feasible compound set complementary to widely used commercial libraries. Together, these contributions reflect the diversity of approaches currently driving the exploration of chemical space, from fundamental methodological work to practical applications in drug discovery. Taken as a whole, the contributions in this special issue underscore both the vastness of chemical space and the ingenuity required to navigate it. They emphasize the importance of uniting theoretical frameworks with experimental feasibility, integrating data-driven machine learning with human-interpretable visualizations, and expanding the boundaries of accessible molecular libraries while maintaining synthetic plausibility. We anticipate that this special issue will serve as a timely snapshot of a dynamic and rapidly evolving field, and we hope it will inspire continued efforts to explore, map, and harness the immense potential of chemical space in the discovery of novel bioactive molecules. I would like to thank all the authors who contributed to this special issue, as well as the anonymous peer reviewers for their insightful and constructive comments on the submitted manuscripts. On behalf of all contributors, we hope that this special issue will advance the ongoing exploration and understanding of the ever-expanding chemical space.

Docking is a structure‐based cheminformatics tool broadly employed in early drug discovery. Based on the tridimensional structure of the protein target, docking is used to predict the binding interactions between the protein and a ligand, estimate the corresponding binding affinity, or perform virtual screenings (VSs) to identify new active compounds. This study introduces the ligand B‐factor index (LBI), a novel computational metric for prioritizing protein‐ligand complexes for docking. Unlike other metrics, LBI directly compares atomic displacements in the ligand and binding site. LBI is defined as the ratio of the median atomic B‐factor of the binding site to that of the bound ligand. Using the comparative assessment of scoring functions (CASF‐2016) dataset, we evaluated the effectiveness of LBI in guiding the selection of protein‐ligand complexes to enhance docking performance. Our results show a moderate correlation (Spearman ρ ~ 0.48) between LBI and the experimental binding affinities, outperforming several docking scoring functions. Additionally, LBI correlates with improved redocking success (root mean square deviation < 2 Å), underlying the significance of a ligand‐focused metric. While LBI outperforms other metrics such as the protein B‐factor index and resolution, its utility in VS docking remains to be further investigated. LBI is easy to compute, interpretable, applicable in structure‐based cheminformatics, and freely available for calculation at https://chembioinf.ro/tool‐bi‐computing.html.

The use of omics data, including gene expression profiles, has recently gained increasing attention in drug discovery. Omics‐based drug searches and designs are often based on the correlations between chemically induced and disease‐induced gene expression profiles; however, the cell specificity has not been considered. In this study, we designed a novel computational method for drug search and design using cell‐specific correlations between drugs and diseases. A data completion technique allowed the characterization of cell‐specific gene expression patterns in diseased cells. This proposed method was applied to search for drug candidates and generate new chemical structures for gastric cancer and atopic dermatitis. The results of drug search demonstrated that compounds with diverse chemical structures were detected and were associated with target diseases at the molecular pathway levels. The results of drug design also demonstrated that newly generated compounds were reasonable in terms of the reproducibility of registered drugs. The proposed method is expected to be useful for omics‐based drug discovery.