CROWN: Curated Repository Of Well-resolved Noncovalent interactions

CROWN introduces a machine learning-ready dataset of 153,005 high-quality protein-ligand complexes that reconciles the trade-off between structural reliability and data diversity by applying a comprehensive automated pipeline with a novel energy minimization step to the PLInder database, offering a geometry-centric resource for training and benchmarking interaction prediction models.

The Gist

Imagine you are trying to teach a robot chef how to cook the perfect meal. To do this, you need a massive library of recipes and photos of finished dishes. But here's the problem: your library is a mess.

Some books are written in perfect, clear handwriting (high-quality data), but there are only a few of them. Other books are written in messy scrawl, with missing pages, torn photos, and ingredients listed that don't even exist (low-quality data), but there are thousands of them.

If you teach the robot with only the few perfect books, it won't know how to cook diverse meals. If you teach it with the messy thousands, it will learn to burn the food and serve inedible dishes.

This is the exact problem scientists face when training AI to understand how drugs (ligands) stick to proteins (the "locks" in our bodies).

Enter CROWN: A new, massive, and perfectly organized library of protein-drug interactions.

The Problem: The "Too Small vs. Too Messy" Dilemma

Before CROWN, researchers had to choose between two bad options:

1. The "Curated" Library (like PDBBind): These are high-quality, hand-checked books. They are reliable, but there are only a few thousand of them. It's like having a library with only 100 recipes. The AI learns the basics but can't handle complex or rare ingredients.
2. The "Massive" Library (like PLInder): This is a warehouse with 650,000 books. It has every recipe imaginable, but many are torn, have missing pages, or list "unicorn horn" as an ingredient. If you feed this to an AI, it gets confused by the errors.

The Solution: The CROWN Factory

The authors of this paper built a fully automated factory (a computer pipeline) that takes the messy warehouse of 650,000 books and turns them into a pristine, high-quality library of 153,000 perfect books.

Here is how their factory works, step-by-step:

1. The Quality Control Gatekeepers

The factory starts by throwing out anything that doesn't meet basic standards.

• The Resolution Filter: If a photo of the dish is blurry (the crystal structure is low resolution), it gets tossed.
• The Ingredient Filter: If the recipe calls for "magic dust" (ions, crystallization artifacts, or weird metals the computer can't understand), it's removed. They only keep "drug-like" ingredients.
• The Pocket Check: Imagine trying to fit a key into a lock, but the lock is missing half its teeth. If the protein "pocket" (the lock) has missing atoms around the drug, the entry is rejected.

2. The Repair Crew

Once the good candidates are selected, a team of digital repair workers fixes the remaining issues:

• Missing Parts: If a recipe is missing a page, they use logic to guess what it should say and fill it in.
• Tangled Wires: Sometimes, the 3D models of the molecules have atoms crashing into each other (steric clashes). The crew gently pushes them apart so they fit naturally.
• The "Flat-Bottom" Magic Trick: This is the paper's secret sauce. Imagine you have a slightly wobbly table (the protein structure from the lab). You want to fix the wobble without moving the table legs too far from where they were originally placed.
  • They use a special "spring" that is loose if the atoms move a tiny bit (within the margin of error of the original photo).
  • But if an atom tries to move too far, the spring gets stiff and pulls it back.
  • Result: The structure becomes physically perfect (no weird bumps or gaps) but still looks exactly like the original experiment. It's like tuning a guitar: you tighten the strings just enough so they sound right, but you don't change the shape of the guitar.

3. The Final Polish

Finally, they check the work. If the "repair" changed the shape of the lock too much, they throw that entry out. They also make sure every "key" (drug) has the right number of hydrogen atoms (protons) to work at body temperature (pH 7.4).

Why is CROWN Special?

• It's Huge: It has 4 times more variety of proteins and species than the old "perfect" libraries. It covers a much wider range of biological life and chemical shapes.
• It's Clean: It has zero missing atoms, zero broken bonds, and zero weird overlaps. It's a "clean room" for AI training.
• It Doesn't Need "Taste Tests": Most old libraries rely on knowing exactly how well a drug worked (binding affinity). But that data is missing for most drugs. CROWN says, "We don't need to know if the drug worked; we just need to know exactly what the lock and key look like together." This allows them to include thousands of structures that were previously ignored.

The Bottom Line

CROWN is like taking a chaotic, dusty attic full of half-finished sketches and turning it into a museum of perfect, high-definition blueprints.

By giving AI models this clean, massive, and diverse dataset, scientists hope to build better "robot chefs" that can design new medicines faster, predict how drugs will behave, and solve the mysteries of how our bodies work at a molecular level. It bridges the gap between having enough data and having good data.

Technical Summary

1. Problem Statement

The development of machine learning (ML) models for protein–ligand interactions is currently hindered by a trade-off in available structural databases:

• High-Quality, Low-Scale Datasets: Resources like PDBBind and HiQBind offer rigorously curated, high-reliability structures but are limited in size (tens of thousands of entries) and chemical diversity. They often rely on manually curated subsets of the Protein Data Bank (PDB).
• High-Scale, Low-Quality Datasets: Resources like PLInder provide broad coverage (nearly 650,000 entries) across the entire PDB but lack rigorous quality control. They contain unresolved atoms, steric clashes, incorrect bond assignments, and crystallization artifacts, which introduce systematic noise into ML training.
• Affinity Bias: Many existing benchmarks rely on experimentally measured binding affinities. This excludes the vast majority of structurally characterized complexes for which no affinity data exists and introduces biases due to heterogeneous assay conditions.
• Structural Inconsistency: Different crystallographers use varying refinement protocols, leading to heterogeneous structural quality and non-uniform geometries across datasets.

2. Methodology

The authors developed CROWN, a fully automated preprocessing pipeline applied to the PLInder database (version 2024-06/v2) to create a machine learning-ready dataset. The pipeline consists of five quality filters interleaved with two structural processing stages:

A. Quality Filters

1. Structure Quality: Retains only high-confidence X-ray structures (resolution $\le$ 3.0 Å). Ligands must have a Real-Space R-value (RSR) $< 0.3$ and a Real-Space Correlation Coefficient (RSCC) $> 0.8$. The pipeline independently re-annotates these metrics to correct parsing errors in PLInder.
2. Ligand Quality: Removes crystallization artifacts, ions, covalent ligands, and ligands containing rare elements (e.g., B, Se, Si) unsupported by the OpenFF 2.2.0 force field. Exceptions are made for specific metal-coordinating cofactors (HEM, MGD, SF4) with available AMBER parameters.
3. Pocket Quality: Excludes complexes with missing protein atoms within 6 Å of the ligand. Non-standard amino acids within the binding pocket are removed to ensure compatibility with the AMBER ff19SB force field.
4. Interaction Quality: Filters for ligands with 10–100 heavy atoms and requiring $>10$ close contacts ($<4$ Å) with the protein. Redundant entries (identical binding poses within 0.1 Å RMSD) are removed.
5. Complex Stability: Post-minimization, entries are retained only if the Root-Mean-Square Deviation (RMSD) is $<0.6$ Å for ligand/pocket atoms and $<0.2$ Å for the protein scaffold, ensuring the minimization did not induce unrealistic structural rearrangements.

B. Structural Processing Stages

1. Structure Corrections:
   • Removal of crystallization artifacts (e.g., glycerol).
   • Selection of highest-occupancy conformers.
   • Reconstruction of missing residues/atoms using PDBFixer.
   • Substitution of non-standard residues (outside the pocket) with standard equivalents (e.g., hydroxyproline $\to$ proline).
   • Repair of inter-chain steric clashes and addition of capping groups (ACE/NME).
   • Inference of bond connectivity and orders based on 3D coordinates if SDF parsing fails.
2. Constrained Energy Minimization (The Core Innovation):
   • Protonation: Assigned at physiological pH (7.4) using PDBFixer (proteins) and Dimorphite-DL (ligands).
   • Force Field: OpenMM 8.4.0 with ff19SB (proteins), OL21 (nucleic acids), OPC3 (water), and OpenFF 2.2.0 (ligands).
   • Minimization Protocol: Uses a flat-bottomed tethering potential to balance crystallographic fidelity with strain relief.
     • Protein Scaffold: Subjected to stiff harmonic restraints ($k = 10^6$ kJ·mol$^{-1}$·nm$^{-2}$) to freeze the global structure.
     • Pocket & Ligand: Subjected to a soft, flat-bottomed potential. Atoms can move freely within a 0.25 Å tolerance (representing crystallographic uncertainty) without energy penalty. Beyond this, a smooth $C^2$-continuous quintic polynomial ramps up the restraint, transitioning to a linear regime to prevent harsh penalties.
     • Hydrogens: Left unrestrained to optimize hydrogen bonding networks.

3. Key Contributions

• Scale vs. Quality Reconciliation: CROWN bridges the gap between curated and large-scale datasets, offering 153,005 high-quality complexes (a 4-fold increase in protein/taxonomic diversity over PDBBind/HiQBind) while maintaining rigorous structural standards.
• Geometry-Centric Design: Unlike affinity-centric datasets, CROWN treats the 3D interaction geometry as a self-consistent source of information, allowing the inclusion of thousands of structures without affinity labels.
• Unique Minimization Step: It is the first protein–ligand dataset to systematically apply constrained energy minimization with custom flat-bottomed restraints. This step harmonizes heterogeneous refinement practices across the PDB, producing structurally uniform complexes without distorting experimentally observed binding geometries.
• Complete Annotation: Every entry has verified electron density metrics (RSR/RSCC), zero unresolved atoms, and zero steric clashes, eliminating "blind spots" common in other databases.
• Open Access: The dataset, web interface, and full preprocessing pipeline (open-source) are freely available.

4. Results

• Dataset Composition: From an initial 649,915 PLInder systems, the pipeline yielded 153,005 complexes.
  • Diversity: 55,208 unique PDB IDs, 12,352 UniProt IDs, 3,209 species, and 26,746 unique ligand CCD identifiers.
  • Chemical Space: Covers a broader range of molecular weights, rotatable bonds, and hydrogen-bond acceptors than PDBBind, including larger ligands (e.g., PROTACs, macrocycles) relevant to modern drug discovery.
• Quality Metrics:
  • Zero Defects: The final dataset contains 0 entries with unresolved ligand atoms, missing bonds, steric clashes, or non-standard pocket residues.
  • Minimization Impact: RMSD analysis shows the protein scaffold remains stable (negligible displacement), while pocket and ligand heavy atoms adjust by a median of 0.22–0.27 Å (within the 0.25 Å tolerance), and hydrogens show larger adjustments (0.52 Å) to optimize H-bond networks.
• Comparison: CROWN outperforms BioLiP2 in ligand diversity (removing ions/artifacts) and surpasses PDBBind/HiQBind in scale and taxonomic coverage.

5. Significance

CROWN addresses a critical bottleneck in structural biology and AI-driven drug discovery: the lack of large-scale, high-fidelity training data.

• For Generative Models: The dataset provides the volume and diversity necessary to train equivariant diffusion models and generative AI for predicting protein–ligand binding poses.
• For Scoring Functions: It enables the development of knowledge-based scoring functions and statistical potentials that are not biased by affinity labels or limited to small chemical spaces.
• Reproducibility: By releasing the full automated pipeline, the authors allow the community to regenerate the dataset as new PDB structures are deposited, ensuring long-term utility.
• Future-Proofing: While current limitations exist regarding metal-coordinating ligands and rare elements (due to force field constraints), the framework is designed to integrate future ML-based force fields to expand coverage.

In summary, CROWN represents a paradigm shift from "affinity-labeled" to "geometry-centric" datasets, providing a robust, scalable, and structurally uniform foundation for the next generation of molecular recognition models.