Scaffold-tuner: a property-directed compound editing tool for scaffold-preserving structural optimization

In the high-stakes world of drug discovery, scientists often face a puzzle that feels like trying to tune a radio to a specific frequency without changing the station. They have a promising molecule, a chemical structure that shows potential to fight a disease, but it needs fine-tuning. To make a drug work better in the human body, researchers must adjust specific physical traits, such as how many hydrogen bonds the molecule can form, how many ring-shaped structures it contains, or how many parts of it can twist and turn. These adjustments are critical because they determine whether a drug can be absorbed by the body or if it will simply pass right through. However, making these changes is notoriously difficult. If a scientist tries to tweak one part of the molecule to improve its absorption, they often accidentally break another part, ruining the drug's ability to bind to its target. The challenge lies in making precise, controlled edits to a complex chemical structure without dismantling its core identity.

To solve this problem, a researcher at Shiga University has developed a new digital tool called Scaffold-tuner. Think of a molecule as a house with a sturdy foundation and a unique roof; the foundation is the core structure that defines what kind of building it is, while the roof and walls can be modified. In chemistry, this unchangeable core is called the scaffold. The goal is to keep that scaffold intact while swapping out the "walls" to change the house's properties. Previous methods for doing this were often like throwing darts in the dark: they would generate thousands of random variations, hoping to stumble upon one that worked, or they would rely on complex artificial intelligence that sometimes changed the foundation along with the walls, creating a completely different building. These approaches were inefficient and often produced results that were too different from the original drug to be useful.

The new tool, Scaffold-tuner, takes a different, more deliberate approach. Instead of guessing or using a black-box algorithm, it uses a curated library of chemical fragments—small, pre-tested pieces of molecules known to perform specific jobs. If a researcher needs to add a hydrogen-bond donor, the tool selects a fragment designed specifically for that task. If they need to remove a rotatable bond, it chooses a fragment that locks the structure in place. By combining these targeted fragments with the original molecule's core scaffold, the tool generates new candidate compounds that are guaranteed to keep the original structure's heart while achieving the desired physical changes. It is a method of editing that prioritizes control and predictability over sheer volume.

To test if this method actually worked, the researcher put it through a rigorous trial using one thousand drug-like compounds taken from a public database called ChEMBL. These compounds served as the starting point, or "parents," for the experiment. The tool was asked to perform eight different types of edits, such as adding or removing a hydrogen-bond donor, adding or removing an aromatic ring, or changing the number of rotatable bonds. For comparison, the same starting compounds were fed into three other methods: a simple random editing tool, a method called CReM that makes chemically reasonable but unguided changes, and a sophisticated artificial intelligence system known as REINVENT4. The goal was to see which method could successfully create a new molecule that met the specific target change while staying true to the original structure.

The results showed that Scaffold-tuner excelled at precision, particularly when researchers could only look at a small number of options. In a scenario where only the very first candidate generated by each method was considered, Scaffold-tuner succeeded in creating the desired change in 42.3 percent of the cases. This was significantly higher than the other methods, which succeeded in only 13.9 percent, 15.6 percent, and 17.1 percent of cases, respectively. This finding suggests that when a scientist needs a quick, reliable suggestion to move a project forward, Scaffold-tuner is far more likely to provide a useful answer immediately. While the artificial intelligence tool eventually caught up and produced a wider variety of successful changes when allowed to generate up to one hundred candidates, it did so by making much larger, more drastic changes to the molecules. These AI-generated molecules often looked very different from the original, making it harder for scientists to understand how the changes affected the drug's behavior.

In contrast, the molecules produced by Scaffold-tuner remained very similar to their parent compounds. On average, the successful candidates retained about 75 percent of the original core scaffold structure, and their overall similarity to the starting molecule was high. This is a crucial advantage because in drug development, scientists need to know exactly which change caused a change in activity. If the structure is too different, it becomes impossible to tell if the new property came from the intended tweak or from a completely new arrangement of atoms. The study found that Scaffold-tuner was particularly effective at adding aromatic rings, a task where the other methods struggled or failed almost entirely. It also managed to increase or decrease rotatable bonds with a level of success that, while not perfect, was consistent and reliable.

However, the tool is not without its limitations. The study noted that the success of the edits depended heavily on the specific type of change requested. For instance, while the tool was excellent at adding rings, it struggled to remove them, achieving a success rate of zero percent for that specific task. Similarly, increasing the number of rotatable bonds was difficult, with a success rate of only 15.2 percent. This indicates that the tool is not a universal fix-all but rather a specialized instrument that works best for certain types of structural adjustments. The researchers also observed that while they could successfully change one property, it was often difficult to do so without accidentally altering other properties at the same time. For example, changing the number of rotatable bonds often came with unintended changes to other parts of the molecule, a challenge that remains difficult for any editing tool to fully overcome.

Despite these constraints, the study concludes that Scaffold-tuner offers a valuable new framework for translating abstract chemical goals into concrete molecular structures. It bridges the gap between theoretical models, which predict how a drug should behave, and the practical reality of building that drug. By providing a way to make specific, interpretable changes while preserving the core identity of the molecule, it allows researchers to explore the relationship between structure and function with greater clarity. The tool does not replace the need for broader, more exploratory methods, but it fills a specific niche where precision and speed are paramount. For scientists trying to refine a lead compound, it offers a way to get a handful of highly relevant options quickly, rather than sifting through thousands of random variations. The source code for the tool is now available for others to use and improve, suggesting that this property-directed approach could become a standard part of the drug discovery toolkit, helping to turn the complex art of molecular editing into a more predictable and efficient science.