Per-residue optimisation of protein structures: Rapid alternative to optimisation with constrained alpha carbons

This paper introduces PROPTIMUS RAPHAN, a computationally efficient method that optimizes protein structures by dividing them into overlapping substructures to achieve linear scaling and results comparable to traditional constrained alpha-carbon optimization in significantly less time.

The Gist

The Problem: The "Rough Draft" Protein

Imagine you have a master architect (like AlphaFold, a famous AI) who draws a blueprint for a massive, intricate castle (a protein). The architect is amazing at getting the big picture right: where the towers are, how the walls connect, and the general shape of the building.

However, the blueprint isn't perfect. If you look closely at the bricks, the mortar, and the tiny details of the windows, they might be slightly crooked, too loose, or in the wrong place. In the world of science, these tiny details are called bond lengths and angles.

If you try to use this "rough draft" castle for a delicate experiment (like testing how a key fits into a lock), the slight crookedness of the bricks could ruin the result. Scientists need to "fix" or optimize these structures to make them perfect.

The Old Way: The "Heavy Lifting" Problem

Traditionally, to fix these structures, scientists used a method called Force Field Optimization. Think of this like trying to straighten every single brick in a 100-story skyscraper at the exact same time.

• The Issue: As the building gets bigger, the work doesn't just get a little harder; it gets exponentially harder. If you double the size of the protein, the time it takes to fix it quadruples.
• The Workaround: To make it faster, scientists would "freeze" the main pillars (the alpha carbons) in place and only try to fix the bricks in between. This helps, but it's still slow and requires a massive amount of computer memory (RAM), often causing the computer to crash on large proteins.

The New Solution: PROPTIMUS RAPHAN (The "Neighborhood" Approach)

The authors of this paper invented a new method called PROPTIMUS RAPHAN. Instead of trying to fix the whole castle at once, they use a "Divide and Conquer" strategy.

The Analogy: The Neighborhood Renovation
Imagine you want to fix a whole city. Instead of hiring one giant crew to fix every house simultaneously (which is chaotic and slow), you break the city into small, overlapping neighborhoods.

1. Zoom In: The method looks at just one house (a residue) and its immediate neighbors.
2. Fix Locally: It fixes the bricks and windows for just that small neighborhood. Because the neighborhood is small, this is very fast.
3. Move On: It moves to the next neighborhood, fixes it, and so on.
4. Overlap: Crucially, these neighborhoods overlap. When they move to the next block, they make sure the shared walls between the two neighborhoods are still aligned.

By doing this, the time it takes to fix the whole city grows linearly. If the city is twice as big, it takes exactly twice as long, not four times as long. It's like walking down a street fixing houses one by one, rather than trying to fix the whole street in one giant leap.

The Results: Fast, Light, and Accurate

The researchers tested this new method on 461 different protein structures (some very large) and compared it to the old "frozen pillar" method.

• Speed: The new method is incredibly fast. It can fix about 5,000 atoms per hour on a standard computer. The old method would take much longer or crash.
• Memory: It uses very little computer memory. While the old method needed a supercomputer's worth of RAM for large proteins, the new method can run on a standard laptop.
• Accuracy: The results are almost identical to the high-precision method. The "bricks" are just as straight.
  • One small catch: Sometimes, the new method finds a slightly different "perfect" shape than the old method. Imagine a flexible door hinge; there are two slightly different ways it can hang perfectly. The new method might pick one, and the old method picks the other. Both are valid, but they are slightly different. This happens mostly in the "floppy" parts of the protein that don't have many connections holding them tight.

The Bottom Line

PROPTIMUS RAPHAN is like a smart, efficient renovation crew. Instead of trying to lift the whole building to fix it, they fix it room-by-room, neighborhood-by-neighborhood.

This means scientists can now take the "rough draft" proteins generated by AI and turn them into high-quality, detailed models much faster and on cheaper computers. This opens the door for more accurate drug discovery and biological research without needing a supercomputer for every single experiment.

Technical Summary

1. Problem Statement

While predictive algorithms (e.g., AlphaFold) and experimental methods can determine the relative positions of protein residues (specifically $\alpha$-carbons) with high accuracy, the local quality of these structures often lacks precision. Issues include inaccuracies in bond lengths, bond angles, and the precise positioning of individual atoms.

• Current Limitation: To correct these local defects, protein structures are typically optimized using force fields. However, optimizing entire protein structures (often containing >10,000 atoms) is computationally expensive, exhibiting quadratic time complexity relative to the number of atoms.
• Existing Workarounds: A common technique is to constrain $\alpha$-carbons to their predicted positions during optimization. While this reduces degrees of freedom, the computational cost remains quadratic, making it impractical for very large structures or high-throughput applications.

2. Methodology: PROPTIMUS RAPHAN

The authors introduce PROPTIMUS RAPHAN (Per-residue optimisation of protein structures: Rapid alternative to optimisation with constrained alpha carbons), a general iterative method based on a divide-and-conquer strategy (specifically the "Cover" approach).

Core Algorithm:

1. Decomposition: Instead of optimizing the whole protein at once, the method divides the structure into overlapping residual substructures. Each substructure corresponds to a specific residue and includes its neighbors.
2. Atom Classification: Within each substructure, atoms are categorized:
   • Optimised atoms: All atoms of the residue except the $\alpha$-carbon and N-H peptide bond atoms (which are replaced by atoms from the subsequent residue to maintain connectivity).
   • Flexible atoms: Atoms within 4 Å of an optimised atom (excluding $\alpha$-carbons).
   • Constrained atoms: The remaining atoms in the substructure, whose coordinates are fixed during the local optimization.
3. Local Optimization: Each substructure is optimized independently using the GFN-FF (Geometries from Force Fields) force field, which offers near-quantum mechanical (QM) accuracy. An implicit solvent model (ALPB) is used.
4. Reconstruction: The optimized positions of the "optimised atoms" are used to reconstruct the global protein structure.
5. Iteration & Convergence: The process repeats iteratively. Residues that have converged (positions no longer change) are excluded from subsequent iterations to speed up the process.
   • Refinement: The entire process is run twice. The second run uses a larger cutoff (8 Å vs. 6 Å) and treats all non-$\alpha$-carbon atoms as optimised to increase accuracy.

Implementation:

• Reference Implementation: PROPTIMUS RAPHANGFN-FF, written in Python, utilizing BioPython, RDKit, and the xtb software package for GFN-FF calculations.
• Scalability: The method is highly parallelizable.

3. Key Contributions

• Linear Time Complexity: By optimizing small, overlapping fragments rather than the whole system, the computational time scales linearly with the size of the protein structure, a significant improvement over the quadratic scaling of traditional methods.
• High Accuracy: The method utilizes GFN-FF, a force field known for its near-QM accuracy, ensuring that the resulting structures are chemically realistic.
• Memory Efficiency: The divide-and-conquer approach drastically reduces memory requirements, allowing the optimization of massive structures on standard hardware.
• Open Source: The authors provide a freely available, parallelizable, and easy-to-install Python implementation.

4. Results

The method was tested on 461 protein structures from the AlphaFold DB (ranging from 200 to 5,000 heavy atoms) and compared against the standard GFN-FFCα (GFN-FF with constrained $\alpha$-carbons) method.

Accuracy Comparison:

• Similarity to Standard Method: Structures optimized by PROPTIMUS RAPHAN (SETRAPHAN) showed high similarity to those optimized by GFN-FFCα (SETCα).
  • Bond Lengths: Mean Absolute Deviation (MAD) of 0.075 pm (comparable to PDB format accuracy).
  • Bond Angles: MAD of 0.136°.
  • Atomic Positions: MAD of 0.074 Å.
• Local Minima Analysis: The study revealed that PROPTIMUS RAPHAN and GFN-FFCα often converge to different local minima on the potential energy surface, particularly in flexible regions (non-polar residues, surface residues) with few hydrogen bonds.
  • However, when SETRAPHAN structures were further optimized by GFN-FFCα (SETRAPHAN+Cα), the deviation was minimal (MAD of 0.033 Å for atomic positions), indicating that PROPTIMUS RAPHAN finds alternative, energetically valid conformations very close to the global minimum.

Performance Comparison:

• Speed: PROPTIMUS RAPHAN achieves an average throughput of 5,000 atoms per hour on a single CPU. In contrast, GFN-FFCα scales quadratically and becomes prohibitively slow for large proteins.
• Memory Usage:
  • GFN-FFCα: Failed to optimize 15 structures due to exceeding 196 GB RAM.
  • PROPTIMUS RAPHANGFN-FF: Optimized the largest test structure (9,940 atoms) using only 0.5 GB RAM (single CPU) or 3 GB RAM (16 CPUs).

5. Significance

• Enabling Large-Scale Optimization: The linear scaling and low memory footprint make it feasible to optimize the hundreds of millions of structures available in databases like AlphaFold DB and ESM Metagenomic Atlas on standard desktop computers.
• Quality Control: It provides a rapid, high-accuracy preprocessing step for applications sensitive to structural quality, such as molecular docking, QSPR modeling, and QM/MM calculations.
• Alternative Conformations: The method's ability to find alternative local minima in flexible regions suggests it may capture biologically relevant conformational states that rigid-constraint methods might miss.

In conclusion, PROPTIMUS RAPHAN offers a computationally efficient, memory-light, and highly accurate alternative to traditional protein structure optimization, bridging the gap between the speed of machine learning predictions and the precision required for detailed computational chemistry.