Planobilly
- 446
- 113
- TL;DR
- Computed dipole/quadrupole for all 550K AlphaFold proteins. Average scaling matches random prediction, but functional families show three distinct regimes. Helicase charge suppression survives composition controls. Zenodo DOI included. Seeking critique.
I want to be upfront: I don't have academic credentials in biology or bioinformatics. My background is in physics and engineering — I hold several patents and spent my career solving applied problems. I'm 78 years old, retired, and I run a small private AI lab out of my home in Florida purely for my own curiosity. The centerpiece is an NVIDIA DGX Spark (128 GB unified memory, Grace Blackwell architecture), which I use alongside local LLMs and custom code to explore whatever catches my attention. I also run a small independent news analysis site. None of this is funded or affiliated with any institution.
The Spark gave me the compute to do something I wouldn't have attempted otherwise. Over the past few months I wrote a custom C# electrostatics pipeline and ran it against all 550,122 proteins in the AlphaFold/SwissProt database (version 4) — computing dipole moments, quadrupole tensors, and solvation energies using AMBER ff14SB partial charges, Henderson-Hasselbalch protonation at pH 7.4, and OBC-II Generalized Born solvation. The full run completed with zero computational failures.
Here's what I found, and why I think it's worth someone with domain expertise taking a look.
The Global Result
The proteome-wide average dipole moment scales as N0.831. The random compact-sphere prediction is N5/6 = 0.833. A match to within 0.002.
That means the average protein's charge placement looks statistically random. But the average hides something.
Function-Specific Deviations
Different protein families organize their charges in fundamentally different ways that correspond to what they do. When I broke the dataset into 22 functional categories and ran ANCOVA, the likelihood ratio rejected a common scaling exponent decisively (χ² = 13,724 for quadrupole, p effectively zero). Three distinct regimes emerged:
The Strongest Evidence: Suppression
To rule out amino acid composition as the explanation, I ran a spatial null model: for 10,097 proteins, I shuffled partial charges across atomic positions 100 times each, preserving composition but randomizing spatial arrangement.
Helicases have a real exponent of 0.44 versus a shuffled expectation of 0.94. That gap of −0.50 means their charges are specifically arranged to cancel polarity. Composition can't explain it. Shape alone can't explain it (I controlled for radius of gyration).
Experimental Validation
I compared AlphaFold-predicted structures against 19,104 X-ray crystal structures (resolution ≤ 2.5 Å):
One Unexpected Result
Ribosomal proteins show a quadrupole exponent of essentially zero (0.09) — flat. Their charge architecture appears frozen, consistent with their status as some of the oldest conserved structures in biology, roughly 3.5 billion years old.
The full paper is published on Zenodo:
DOI: 10.5281/zenodo.20411754
I also wrote a plain-language version on Medium for anyone who wants the shorter explanation.
I'm not claiming this is groundbreaking or that I've discovered something the field has missed. I had the compute, I had a question, and these are the results. What I'd like to know:
— Billy Simmons
The Spark gave me the compute to do something I wouldn't have attempted otherwise. Over the past few months I wrote a custom C# electrostatics pipeline and ran it against all 550,122 proteins in the AlphaFold/SwissProt database (version 4) — computing dipole moments, quadrupole tensors, and solvation energies using AMBER ff14SB partial charges, Henderson-Hasselbalch protonation at pH 7.4, and OBC-II Generalized Born solvation. The full run completed with zero computational failures.
Here's what I found, and why I think it's worth someone with domain expertise taking a look.
The Global Result
The proteome-wide average dipole moment scales as N0.831. The random compact-sphere prediction is N5/6 = 0.833. A match to within 0.002.
That means the average protein's charge placement looks statistically random. But the average hides something.
Function-Specific Deviations
Different protein families organize their charges in fundamentally different ways that correspond to what they do. When I broke the dataset into 22 functional categories and ran ANCOVA, the likelihood ratio rejected a common scaling exponent decisively (χ² = 13,724 for quadrupole, p effectively zero). Three distinct regimes emerged:
| Regime | Families | Dipole Exp. | Interpretation |
|---|---|---|---|
| Superlinear | Kinases (1.10), Motors (1.09), Immunoglobulins (1.17), Receptors (1.00) | > 0.9 | Charges amplify with size — long-range electrostatic steering |
| Moderate | Proteases (0.90), Polymerases (0.85), Chaperones (0.79) | 0.65 – 0.9 | Near the random baseline — electrostatics not the primary constraint |
| Suppressed | Helicases (0.57), Cytochromes (0.64), Oxidoreductases (0.62) | < 0.65 | Active charge cancellation — the opposite of random |
The Strongest Evidence: Suppression
To rule out amino acid composition as the explanation, I ran a spatial null model: for 10,097 proteins, I shuffled partial charges across atomic positions 100 times each, preserving composition but randomizing spatial arrangement.
Helicases have a real exponent of 0.44 versus a shuffled expectation of 0.94. That gap of −0.50 means their charges are specifically arranged to cancel polarity. Composition can't explain it. Shape alone can't explain it (I controlled for radius of gyration).
Experimental Validation
I compared AlphaFold-predicted structures against 19,104 X-ray crystal structures (resolution ≤ 2.5 Å):
- Solvation energy correlation: r = 0.97
- Quadrupole: r = 0.79
- Dipole: r = 0.73
One Unexpected Result
Ribosomal proteins show a quadrupole exponent of essentially zero (0.09) — flat. Their charge architecture appears frozen, consistent with their status as some of the oldest conserved structures in biology, roughly 3.5 billion years old.
The full paper is published on Zenodo:
DOI: 10.5281/zenodo.20411754
I also wrote a plain-language version on Medium for anyone who wants the shorter explanation.
I'm not claiming this is groundbreaking or that I've discovered something the field has missed. I had the compute, I had a question, and these are the results. What I'd like to know:
- Does the methodology hold up? Are there obvious flaws in the electrostatics pipeline or the statistical approach?
- Is the composition-preserving shuffle a valid null model, or is there a better control?
- Has anyone seen similar functional clustering by electrostatic scaling in the literature? I searched but may have missed prior work.
- Are there alternative explanations for the helicase suppression result that I haven't considered?
— Billy Simmons