Last week, I attended the 2026 RDKit UGM in person. It was really nice to meet other participants and have lots of discussions with them. All the oral presentations and poster presentations were great, and I was able to learn about a lot of cutting-edge work in cheminformatics. I learned about a new conformer generation method, implemented in the latest version of RDKit, during the "What's New" session.

As most of readers know that OSS is contribute not only bioinformatics but also cheminformatics area and recently lots of AI tools use OSS as a parts of them. RDKit, OpenBabel and CDK are famous cheminformatics TK for handing chemical structure. I've started my journey of cheminformatics in Drug Discovery field with python and RDKit for almost 15 years ago.

Extending searching in synthon spaces to use Gaussian shape similarity

Strava connects millions of runners, cyclists, hikers, walkers and other active people through the sports they love – all on our mobile app and website.

2026 RDKit User Group Meeting (in person and online)

I'm on summer vacation right now. Humidity and temperatures are high at this time... So I sifted running time from lunch time to morning. The kids seem to be busy with club activities even during summer vacation, so I have plenty of time to run, read papers and coding 🙂 Now there are lots of molecular filters are developed and available for filtering unwanted molecules from given dataset.

Recently several co-folding AI are available, for example AF3, Chai-1, protenix and openfold. This area is hot and growing day by day. Co-Folding AI has potential for taking SBDD approach without determining the target protein-ligand structure. As most of readers know that there are still limitation of using co-folding AI for SBDD but really useful I think. I have an interest in openfold because openfold disclose code and has good open scientific community.

Now there are lots of AI tools for drug discovery. I think it's difficult to expect 8 years before when I moved to cheminformatics team from MedChem 🙂 And one of the hot area is Folding AI such as Boltz, Alpha fold and OpenFold etc. Boltz-2 has not only cofolding prediction module but also binding affinity prediction module. It will be useful if the module works well.

曲・2010年・時間：3:35

曲・2010年・時間：4:46

Collecting sets of closely related compounds

As readers know that RECAP and BRICS are famous fragmentation method of molecules. Both methods use pre-defined rules. Today I found new fragmentaion method named CRUSH in Journal of cheminformatics reported from Edgar's group. The arthors defeined new chemically aware fragmentation rules with SMIRKS and the rule defeined as txt format in CRUSH repository. They compared difference between CRUSH and traditional methods (RECAP/BRICS) and found that CRUSH can generate new fragments from same molecules.

As most of readers know that Open Molecular Software Founcation (OMSF) has lots of attractive projects. Such as Alchemistry, Open Free Energy, Open Fold and Open ADMET etc. Predict ADMET properties for drug design is useful because chemists should optimize molecule not only primary tareget potency but also other profiles such as ADMET. From OMSF site, what is open admet ~~~

A slew of start-ups and academic labs are leaning on AI agents and bots, rather than humans, to speed up their chemistry

Today I would like to introduce transformer based molecular generator named 'ANNalog' which can generate molecules like medchem. The URL of the article is below. ANNalog model is trained with pair of molecules which are paired not only Tanimoto similarity but also biological similarity. It's important concept because medchem often would like to do scaffold hopping which means design novel molecules which have same biological activities but different scaffold.

Optimization of absorption, distribution, metabolism, and excretion (ADME) properties is central to small-molecule drug discovery, yet is often constrained by limited access to large and consistent ex...

Two next-generation open models for small-molecule and protein prediction, now self-serve through the Boltz API.

このブログは、第一三共株式会社 スマートリサーチ第二研究所と QSimulate による共著です。 はじめに […]

Recently, generative AI can generate reasonable molecules under user defined conditions but it's difficult to consider synthetic feasibility of generated molecules correctory. So building block beased approaches are developed for design novel drug like molecules. By using the approach, origin of designed molecules is building blocks and known reactions. It means synthetic feasibility is high compared to fully de novo designed molecules by using AI only.

Targeted Protein Degrader (TPD) is one of the attractive modality of recent drug discovery area. As most of readers know that First TPD, ER degrader gains FDA approval in this year. TPD molecule induce PPI of POI and E3 ligase and degradation of POI. It is unique mechanism because chemical compounds can induce artificial degradation process. But the process is really complex so it's difficult to design TPD molecule rationally such as SBDD.

I'm interested in how to apply quantum chemistry for my drug discovery journy longtime. My former boss teached me details of FMO. FMO is really is really useful tool for understanding protein-ligand or protein-protein interaction form QM point of veiw. I love it 🙂 But GAMESS don't have official python API. And I and kzfm developed psikit which is wrapper of psi4 using rdkit.

Understanding protein-ligand interaction is important for rational drug design. We can use X-ray co crystal structures for SBDD and information of ligand binding site is essential reagion. For example QM/MM approach applies QM level calculation to around binding site and applies MM to far from binding site. QM often offer more accurate energy state but its computational cost is high. So MM is used for not critical reagion for protein-ligand binding.

Accurate prediction of molecular properties requires the refinement of the full conformational ensemble at high levels of theory, which remains computationally demanding. Existing workflows typically rely on ad-hoc scripts and manual intervention. We present Ensemble Analyzer (EnAn), an open-source Python framework that automates this workflow. EnAn’s modular and extensible architecture integrates seamlessly with widely known calculators and allows automated generation and comparison of electronic and vibronic spectra. EnAn effectively manages the reproducible exploration of conformational spaces.

Recent years have seen a rise in applications of deep learning to problems in the molecular sciences. Among them, the diffusion model DiffDock stands out as a method for docking small molecules into p...

The serotonergic system remains a critical focus of neuropsychopharmacology due to its widespread influence on mood, cognition, and behavior. Despite the clinical success of selective serotonin reupta...

Introducing ggsql, a grammar of graphics for SQL that lets you describe visualizations directly inside SQL queries.

ピアニスト堀江沙知による和製ジャズ・ピアノインストプロジェクト。