De novo SBDD with Automation: Current State and Perspectives

Structure-based drug design (SBDD) is a rational and reasonable way for the creation of more feasible and acceptable molecules to the target protein of interest. Generative AI is changing our way of life, and it has a huge impact on scientists and researchers. Recent development of generative AI-mediated structure-based computational design is opening the utilization of automated process. Now is the age of making full use of computational methodologies.

Recent advances in automation in the field of SBDD with de novo creation of compound structures is summarized in a review.¹⁾ It comprehensively summarized the very beginning²⁾ of structure-based computer-aided drug design (SB-CADD), especially focusing on the recent ones. The authors categorized CADD approaches into three (below) and picked up the examples.

-　Fragment-based ligand designs
-　Evolutionary algorithms
-　Deep learning de novo drug design methods

In SB-CADD, the designed molecules de novo needs to be evaluated by scoring. Applying lots of conditions and calculations demand the computational capability. Thus the authors focused on the balance of time and accuracy to figure out the character of each methodology.

Their research was literature-based and no additional experiments were performed for further validation. But it is a fair way to understand the nature of SB-CADD methods. They put comments on the way of ligand generation, synthetic accessibility evaluation and validation of the methodology for every method. We can easily compare the recent methodologies as well as available data and software with the characteristics to apply for the drug discovery and development.

There is a caution for those who are interested in deep learning (DL)-based generative AI for de novo SBDD.³⁾ DL-based generative AI shows interesting chemical structures, but there is no explanation for the structural generation or scoring. Scientists need to analyze the dataset and agathism to understand the way of generation. In terms of validation of methodology with DL, docking based Monte Carlo tree search (MCTS) needs to be considered because it can differentiate the molecular optimization.⁴⁾,⁵⁾

There are still challenges to practically use SB-CADD de novo design for drug discovery. Even though the generative models are now flexibly constructable and generally useful methods are available through the website, scoring is of an issue. The problem is understandable through the investigations so far that observed false positives.⁶⁾,⁷⁾

Automation of SBDD is becoming possible but the validation is a must. The recent trend focuses on AI and DL, but the dataset would change the result of de novo design of molecules. In CB-SBDD, it is necessary to take the common benchmarks into consideration. There is no standardized benchmark for CB-SBDD. It depends on the target biomolecule of interest and you need to define it.

This is our current understanding on de novo SBDD with recent technologies. Full automation is still difficult in drug discovery field. We are eager to utilize novel techniques for SBDD, but in combination with talented scientists. We think it is the way to keep running the cutting edge in drug discovery with the aid of SBDD.

1) https://doi.org/10.1021/acs.jcim.4c00247
2) https://doi.org/10.1016%2FS0040-4020(01)86503-0
3) https://doi.org/10.1039%2FD1SC05976A
4) https://doi.org/10.1039%2FD1SC04444C
5) https://doi.org/10.1021%2Facs.jcim.1c00679
6) https://doi.org/10.1371%2Fjournal.pone.0240450
7) https://doi.org/10.1021%2Facs.jcim.8b00545