ProteinConformers: large-scale and energetically profiled descriptions of protein conformational landscapes

ProteinConformers is a comprehensive resource that provides 2.7 million geometry-optimized protein conformations with extensive energetic and similarity annotations, addressing critical gaps in conformational coverage, energy profiling, and benchmarking standards for protein dynamics and drug discovery.

The Gist

Imagine a protein not as a rigid, static statue, but as a living, breathing dancer. To understand how a protein works—how it fights disease, how it helps your body digest food, or how it can be targeted by a new drug—you have to watch its entire dance routine, not just one frozen pose.

This paper introduces ProteinConformers, a massive new digital library that maps out every possible move this "dancer" can make.

Here is the story of what they built, explained simply:

1. The Problem: We Only Had Snapshots

For a long time, scientists had a library of protein "photos" (static structures). It was like trying to understand a movie by looking at a single frame.

• The Limitation: Existing tools could only show the protein in its most comfortable, relaxed pose (like a dancer standing still). They struggled to show the protein stretching, twisting, or wobbling into different shapes, which is often where the real magic (and drug targets) happens.
• The Gap: We didn't have a map of the "energy landscape." Think of a protein's energy like a hilly terrain. The bottom of the valley is the most stable shape. But proteins sometimes need to climb a small hill to get to a different valley to do their job. Previous maps didn't show us the hills and valleys well enough.

2. The Solution: A Massive "Dance Floor" Simulator

The researchers built ProteinConformers, a super-computer simulation that acts like a giant, high-tech dance floor.

• The Scale: They didn't just simulate one dancer; they simulated 734 different proteins.
• The Moves: For each protein, they didn't just take one photo. They generated 2.7 million different poses (conformations).
• The Method: Instead of starting with the protein in its perfect pose, they started with hundreds of slightly "messy" or "twisted" starting positions (like asking a dancer to start in a weird stretch). Then, they let physics run a simulation (like a video game engine) to see how the protein naturally relaxed, bounced, and settled into different shapes.
• The Result: They created a continuous map from "totally messy" shapes to "perfectly folded" shapes, filling in the gaps we used to miss.

3. The "Scorecard": Is the Dance Real?

Just because a computer can make a protein twist doesn't mean the twist is physically possible. The authors had to prove their data was real.

• The Energy Check: They calculated the "energy cost" of every single pose using five different scientific formulas. It's like checking if a dancer's move requires too much energy to be realistic.
• The Benchmark: They created a smaller, high-quality "exam" set called ProteinConformers-lite. They used this to test other AI models (like AlphaFlow and BioEmu) to see if those AIs could generate good dance moves.
• The Verdict: Their data was as physically realistic as the best existing datasets, but it covered a much wider range of movements.

4. The New Tool: An Interactive Video Game for Scientists

The best part? They didn't just dump the data in a folder. They built a website (a portal) that anyone can use.

• The Dashboard: Imagine a website where you can search for a protein, and instead of a boring list, you see a 3D model you can spin around.
• The Filters: You can say, "Show me all the poses where the protein is twisted but still has low energy," or "Show me the poses that look like the native shape."
• The Download: If you are a researcher, you can download the entire "dance routine" for a specific protein to study it on your own computer.

Why Does This Matter?

• For Drug Discovery: Many drugs work by locking a protein in a specific shape. If we only know the "standing still" shape, we might miss the "twisted" shape that the drug actually needs to grab onto. This library gives us all the shapes to look at.
• For AI: It gives Artificial Intelligence a better textbook to learn from. By seeing millions of examples of how proteins move, AI can learn to predict new shapes more accurately.
• For Understanding Life: It helps us understand allostery—a fancy word for how a protein changes shape in one part of its body to trigger a reaction in another part (like a domino effect).

In short: The authors built the world's most detailed "motion capture" database for proteins. They turned static photos into a full-motion movie, complete with a scorecard to prove the moves are real, and put it all on a website so scientists can finally see the full dance of life.

Technical Summary

1. Problem Statement

Understanding protein function requires capturing dynamic structural interconversions across conformational landscapes, which are governed by underlying thermodynamic energy landscapes. However, existing resources and methods suffer from three major limitations:

• Limited Diversity: Molecular Dynamics (MD) simulations often start from native structures near the global energy minimum, failing to sample broad conformational spaces (from non-native to near-native states).
• Lack of Standardization: There are no standardized benchmarks for assessing the geometric plausibility and landscape-wide diversity of multi-conformer generators.
• Insufficient Annotation: Existing datasets lack comprehensive energetic annotations and structural similarity metrics, leaving the coupling between conformational variability and energetics poorly characterized.

2. Methodology

A. Dataset Construction (ProteinConformers)

The authors constructed a large-scale dataset comprising 734 proteins (sequence lengths 33–949 residues, avg. 247).

• Source Data: Decoys were manually curated from CASP (Critical Assessment of protein Structure Prediction) seasons 5–15.
• Preprocessing: A rigorous pipeline removed duplicates, fixed missing residues/atoms, corrected numbering, and filtered out non-protein entries or mismatched sequences.
• Multi-Seed Sampling: Instead of single-trajectory MD, the team initiated simulations from hundreds of diverse starting conformations (seed decoys) per protein.
• MD Protocol:
  • Software: GROMACS 2023 with the OPLS-AA force field and TIP3P water model.
  • Process: Steepest-descent minimization $\rightarrow$ NVT (100 ps, 300 K) $\rightarrow$ NPT (100 ps, 1 bar) $\rightarrow$ Production run (125–375 ps).
  • Output: Snapshots extracted every 25 ps. Failed simulations were discarded.
• Scale: Generated 2.7 million geometry-optimized conformations from ~352k seed decoys.

B. Energetic and Similarity Profiling

• Energetics: Each conformation was evaluated using five energy functions: RW, RWplus, EvoEF2, Rosetta (REF2015), and FoldX, resulting in 13.7 million energy evaluations.
• Similarity: Pairwise TM-scores and RMSDs were calculated against native structures, generating 5.5 million similarity annotations.

C. Benchmark Dataset (ProteinConformers-lite)

A curated subset of 87 proteins (from CASP14/15) containing 381,546 conformers was created to serve as a high-quality benchmark. Targets were selected for their challenging nature and high-quality seed decoys.

D. Evaluation Framework

The paper introduces a dual-axis benchmarking framework to evaluate generative models:

1. Diversity Metrics:
   • Projects conformations onto 2D free energy landscapes using PCA.
   • Calculates Interaction, Coverage, and Jaccard Index based on low-energy region overlap (thresholds: 5, 10, 20 kJ/mol) between the reference and generated ensembles.
2. Plausibility Metrics (Conformation Geometry Map - CGM):
   • Computes inter-residue geometric features: Euclidean Distance ($D$), Dihedral Angle ($\Omega$), Orientation ($\Theta$), and Planar Angle ($\Phi$).
   • Derives statistics (mean, std, skewness) for residue pairs.
   • CGMScos: Cosine similarity of descriptor vectors (directional agreement).
   • CGMSmah: Mahalanobis distance-based similarity (global geometric consistency).

3. Key Contributions

1. ProteinConformers Database: A publicly available resource with 2.7M conformations, 13.7M energy scores, and 5.5M similarity annotations, covering a spectrum from non-native to near-native states.
2. ProteinConformers-lite: A standardized benchmark set for evaluating multi-conformer generators, featuring a broader conformational spectrum than the ATLAS database.
3. Evaluation Framework: A novel methodology combining energy-threshold-based diversity analysis and geometry-map-based plausibility scoring (CGMS).
4. Interactive Web Portal: A Streamlit-based platform allowing users to query, visualize (3D alignment, distance/rotation maps), filter, and download data without local computation.

4. Key Results

Dataset Quality

• Stereochemical Quality: The local geometry (dihedral angles, bond lengths) of ProteinConformers-lite closely matches the high-quality Top2018 reference dataset.
• Ramachandran Outliers: Outlier rates decrease as TM-score increases, falling below the Top2018 average (13%) for near-native states (TM-score > 0.5), confirming physical plausibility.
• Diversity: The dataset spans a wider per-protein TM-score range and covers more diverse fold families (ECOD/CATH) compared to ATLAS.

Benchmarking Generative Models

Five models were tested: AlphaFlowMD/Dis, ESMFlowMD/Dis, and BioEmu.

• Diversity: BioEmu achieved the highest coverage of low-energy regions under strict thresholds (5 kJ/mol), indicating superior sampling of stable states. Distilled variants (e.g., AlphaFlowMD/Dis) showed reduced coverage, suggesting narrower exploration.
• Plausibility:
  • Models generally recover distance distributions better than orientation statistics.
  • BioEmu scored highest on distance components.
  • AlphaFlowMD/Dis achieved comparable global geometric consistency (CGMSmah) to BioEmu.
  • MD Refinement Impact: Models fine-tuned on MD data showed only modest improvements in local geometric realism compared to non-MD tuned counterparts.

5. Significance

• Foundation for Dynamics: Provides a rigorous, energetically annotated foundation for studying protein dynamics, allostery, and drug discovery.
• Standardization: Establishes a critical benchmark for the community to evaluate the next generation of protein conformation ensemble predictors.
• Accessibility: The interactive web portal democratizes access to complex conformational data, enabling rapid analysis without requiring local high-performance computing resources.
• Insight into Generative AI: The benchmarking results highlight current limitations in AI-generated protein ensembles, specifically regarding orientation statistics and the marginal benefit of MD refinement on local geometry.

Availability:

• Web Portal: https://zhanggroup.org/ProteinConformers
• Benchmark Data: https://huggingface.co/datasets/Jim990908/ProteinConformers
• Code: https://github.com/auroua/ProteinConformers