Automated dataset integrating structural and bioactivity data for structure-based drug discovery

Imagine you are trying to teach a robot how to design a new key that fits perfectly into a specific lock. In the world of medicine, the "locks" are proteins inside our bodies that cause diseases, and the "keys" are tiny drug molecules. For decades, scientists have been trying to build a super-smart robot (using machine learning) that can predict exactly how well a key will fit a lock just by looking at their shapes. But there's a huge problem: the robot is hungry for data, and the data it needs is scattered everywhere. Some scientists have the shape of the lock, others have the measurements of how well a key turns, but they are rarely in the same place. It's like trying to teach a chef to cook a perfect meal when the recipe is in one book, the picture of the ingredients is in another, and the taste test results are in a third. Without a massive, organized library that connects the shape of the lock to the exact strength of the key's grip, the robot can't learn effectively.

This is where the story of DockTData begins. A team of researchers from the National Laboratory of Scientific Computing in Brazil decided to build that missing library. They didn't just gather a few recipes; they built an automated factory that scours the internet for every available piece of information about protein locks and drug keys, stitches them together, and organizes them into a single, giant dataset. Think of it as a massive, digital "matchmaking" service. It takes the 3D blueprints of proteins from the Protein Data Bank (PDB) and links them to the real-world test results from databases like BindingDB and ChEMBL. The result is a treasure trove of over 24,719 unique protein-ligand complexes. This isn't just a list of numbers; it's a collection where every single drug molecule is tied directly to its 3D structure and its measured strength (how tightly it binds), standardized so a computer can read it all at once. By making this data open and free for everyone, the authors hope to give the next generation of drug-discovering robots the fuel they need to learn faster and smarter, potentially leading to new medicines for diseases that currently have no cure.

The Paper's Core Mission: Building the Ultimate Drug-Design Library

The paper introduces DockTData, a new, openly available dataset designed to supercharge "structure-based drug discovery." In simple terms, this is the process of designing drugs by looking at the 3D shape of the target protein and figuring out which molecules will fit into it best. The authors created this dataset by building an automated pipeline that acts like a super-efficient librarian. This librarian goes out to three massive sources—the Protein Data Bank (PDB), which holds 3D structures of proteins; BindingDB; and ChEMBL, which hold measurements of how well drugs bind to proteins—and pulls everything together.

The magic of DockTData lies in how it connects the dots. Before this, a scientist might have a picture of a protein with a drug stuck to it, but they wouldn't know exactly how strong that bond was, or they might find the strength measurement in a different database without the picture. DockTData solves this by linking the 24,719 protein-ligand complexes (the "locks" with their "keys") directly to 45,331 bioactivity measurements. These measurements include values like Kd, Ki, IC50, and EC50, which are all ways of saying "how tightly does this drug stick?" The authors standardized all these values into nanomolar (nM) units, so they can be compared apples-to-apples.

How They Did It: The Automated Matchmaking Machine

The researchers didn't just copy and paste; they built a sophisticated system to ensure the matches were perfect. They used a few clever tricks to make sure the data was clean and reliable:

• The Protein Match: They used a tool called MMseqs2 to compare protein sequences. They set a rule that a protein from a drug test must be at least 85% identical to the protein in the 3D structure to count as a match. This is like saying, "If the lock looks 85% the same as the one in the photo, we'll assume it's the same lock." This threshold was chosen to include useful variations (like slightly different versions of the same protein) without including totally different proteins.
• The Drug Match: For the drug molecules, they used something called an InChIKey. Think of this as a unique, 27-character barcode for every chemical structure. By recalculating these barcodes using a standard tool (RDKit) for every single entry, they ensured that the same drug wasn't listed twice under different names or with tiny, confusing differences.
• The "One Key, One Lock" Rule: They were very strict about pairing. If a protein structure had two different drugs stuck to it at the same time, they excluded it. They only kept complexes where there was exactly one drug molecule paired with one protein structure. This prevents confusion when training computer models.

What's Inside the Box?

The resulting dataset is a massive collection of information, organized into easy-to-use tables. Here is what you get:

• 24,719 unique protein-ligand complexes.
• 11,742 distinct drug molecules (ligands).
• 17,732 unique protein structures from the PDB.
• 2,163 different protein targets (the "locks").
• A mix of small molecules (the classic drug type, making up 24,378 of the entries), peptides (271 entries), and even some oligosaccharides (70 entries).

The authors note that their small-molecule collection is 46% larger than the small-molecule portion of the popular PDBbind version 2020 general set, which had 16,744 entries. This means DockTData offers a significantly bigger playground for scientists to test their ideas.

Why This Matters: More Than Just Numbers

The paper emphasizes that the quality and size of training data are the biggest bottlenecks for machine learning in drug discovery. If you train a robot on a small, messy dataset, it will learn bad habits. DockTData offers a "cleaner," larger, and more transparent dataset. Because the authors kept track of exactly where every piece of data came from (the "provenance"), scientists can trace any number back to its original source, check for errors, and understand the context.

The dataset includes a wide variety of measurement types: 58.8% are IC50 values, 20.3% are Ki, 11.3% are Kd, and 9.7% are EC50. This diversity is important because different types of measurements tell slightly different stories about how a drug works. By having them all in one place, standardized to nanomolar units, researchers can choose the specific type of data that fits their needs.

The "No-Go" Zones: What the Paper Rules Out

It is important to note what DockTData is not. The authors explicitly state that they do not include structures that were only predicted by computers (like those from AlphaFold) without experimental proof. They only included structures that were actually solved in a lab using X-ray, NMR, or cryo-electron microscopy. This means the dataset is strictly grounded in experimental reality, not computer simulations.

Furthermore, the paper rules out ambiguous pairings. If a protein structure had multiple different drugs attached, or if the link between the protein sequence and the drug measurement was too weak (below the 85% identity threshold), that data was thrown out. The authors argue that this strictness is necessary to avoid "noise" that could confuse machine learning models. They also note that while they have a lot of data, they are not claiming to have solved the problem of drug discovery. Instead, they are providing a better foundation—a "transparent, reusable, and continually expanding resource"—for others to build upon.

The Final Takeaway

DockTData is a gift to the scientific community. It is a massive, organized, and free library that connects the 3D shapes of proteins to the real-world strength of drug interactions. By automating the process of gathering and cleaning this data, the authors have created a resource that is larger and more diverse than many previous collections. They have made it possible for researchers to train smarter AI models, test new theories about how drugs bind to proteins, and ultimately, accelerate the discovery of new medicines. The dataset is available under a Creative Commons Attribution 4.0 International License, meaning anyone can use it, share it, and build upon it, as long as they give credit. It's a clear step forward in turning the chaotic world of drug discovery data into a structured, learnable language for machines.