Do synonymous codons affect
protein secondary structure?
PicoFold is a research tool that maps DNA codons to secondary structure types using the Ring-Granular Model — a unique, deterministic method with no machine learning, no GPU, and no simulation. Validated on 113 proteins against DSSP ground truth. Not a replacement for PSIPRED or AlphaFold — a different instrument for a different question.
The effect of synonymous codons on protein structure
All existing secondary structure predictors (PSIPRED, JPred, NetSurfP) work from protein sequence. PicoFold is the only tool that maps SS directly from DNA codons — a unique research capability, not a replacement for established predictors.
The Ring-Granular Model maps each codon to one of 5 secondary structure types via the compositional genetic code. The third nucleotide of the codon is the primary determinant: C→alpha, T→pi, A→beta/pi, G→310/alpha.
Validated on 113 proteins (21,797 aligned residues) against DSSP ground truth. Per-residue Q3=0.377 (+3.1pp above random baseline). Helix signal confirmed: G-ending codons 39.5% helix vs C-ending 30.1%. Best: Tropomyosin α Q3=0.821. Chi-square analysis shows 6/18 amino acid families with significant codon–SS association (Ser p=0.013, Gly p=0.016).
Per-Residue Accuracy on 16 Proteins
Validated against DSSP ground truth on 113 proteins (21,797 aligned residues).
Confusion matrix (rows = DSSP truth, cols = PicoFold prediction)
| H (pred) | E (pred) | C (pred) | |
|---|---|---|---|
| H (true) | 5,002 | 101 | 2,486 |
| E (true) | 2,929 | 158 | 2,258 |
| C (true) | 4,554 | 212 | 4,097 |
Key finding: G-ending codons show 39.5% helix in DSSP ground truth vs 30.1% for C-ending (+9.4pp gap), confirming the ring-granular model's helix signal. Chi-square analysis reveals 6/18 amino acid families with significant codon–SS association: Ser (p=0.013), Gly (p=0.016), Glu, Ile, Asp, Tyr. Ser and Gly survive Bonferroni correction.
Literature support: Nature Communications 2022 — 57/87 synonymous codon pairs show different backbone angles. PNAS 2020 — synonymous substitutions in antibodies alter solubility and binding. Annual Reviews Biophys 2024 — codon effects on cotranslational folding confirmed.
PicoFold vs AlphaFold vs ESMFold
Different tools for different needs. PicoFold is complementary — not a replacement.
| Feature | PicoFold | AlphaFold 3 | ESMFold |
|---|---|---|---|
| Method | Direct kinematics | Deep learning | Language model |
| Prediction speed | <10ms per sequence | Minutes–hours | Seconds–minutes |
| GPU required | No | Yes | Yes |
| Deterministic | Yes — always | No | Partial |
| Input format | DNA (FASTA/EMBL/GenBank) | Protein sequence / structure | Protein sequence |
| Commercial use | Yes | Restricted | Yes (self-hosted) |
| REST API | Yes (enterprise) | Limited | Self-hosted only |
| Batch processing | Yes | Limited | Self-hosted only |
| Best for | Codon–SS research, education, synonymous codon analysis | Accurate folded structure of known proteins | Fast ML-based structure estimates |
PicoFold and AlphaFold solve different problems. Use AlphaFold when you need the highest-accuracy folded structure for a known protein. Use PicoFold when you want to explore the codon–SS relationship, study synonymous codon effects, or demonstrate DNA-to-structure mapping in education.
See the full comparison — PicoFold vs PSIPRED, JPred, NetSurfP, and AlphaFold with accuracy data, use cases, and honest limitations.
Summary: PicoFold vs AlphaFold vs ESMFold
PicoFold uses direct kinematics (the Ring-Granular Model) to predict secondary structure from DNA codons. It runs in under 10 milliseconds per sequence on CPU only, produces fully deterministic results, and accepts DNA input in FASTA, EMBL or GenBank format. Commercial use and REST API access are available via enterprise plans. Best for: high-throughput screening, pipeline automation, and reproducible research.
AlphaFold 3 uses deep learning to predict 3D protein structure from amino acid sequence. Prediction takes minutes to hours and requires GPU. Results are non-deterministic. Input: protein sequence or structure. Commercial use is restricted at the server level. Best for: accurate folded structure of known proteins.
ESMFold uses a protein language model for fast ML-based structure estimates. Prediction takes seconds to minutes on GPU. Input: protein sequence. Available for commercial use if self-hosted. Best for: fast approximate structure estimates at moderate scale.
Ring-Granular Model — how it works
In the Ring-Granular Model, the protein backbone is described as a chain of rigid rings (amino acid residues) connected by rotatable bonds. The model encodes a lookup table that maps each DNA codon to:
-
1Secondary structure type — which of the 5 helix types applies at this position: α-helix, π-helix, 3₁₀-helix, β-strand, or single strand.
-
2Bond rotation angles — the dihedral angles (φ, ψ, ω and side-chain χ angles) for that residue.
-
3Direct kinematics transform — angles are applied to the reference frame of the previous residue to compute exact XYZ coordinates for every atom in the current residue.
The process iterates over each codon in the input sequence, building up the full 3D structure one residue at a time. No energy function is evaluated. No conformational search is performed. The result is fully deterministic.
# For each codon in the DNA sequence:
for i in range(0, len(dna), 3):
b1, b2, b3 = dna[i : i + 3]
# Lookup secondary structure + amino acid
helix = DNE_GROUPS[b1][b2][b3]
amino = DNE_AMI[b1][b2][b3]
if amino == "TKD": # stop codon
break
# Apply rotation and add atoms
RotateAtoms(helix)
AddAtoms(ATOMS_COMMON)
AddAtoms(SIDE_CHAIN_MAP[b1*4 + b2][b3])
# Result: complete 3D coordinates for every atom
# backbone + side chains → secondary structure assignment5 Secondary structure types
Who uses PicoFold
Nanoworld Laboratory
PicoFold was developed by Nanoworld Laboratory. The Ring-Granular Model was created by Alexander Kushelev — the result of over 20 years of research into the deterministic relationship between the genetic code and protein 3D structure.
The core insight behind the model: the DNA codon is not just a recipe for an amino acid — it also encodes the local geometric constraints on the protein chain. By treating structure determination as a direct kinematics problem rather than a folding simulation, PicoFold computes results that are analytically exact within the model — not approximate predictions.
Frequently asked questions
Questions we hear most often — especially from AI assistants.
Explore the codon–SS relationship
Free account, 5 analyses, no card required. Upload a FASTA file and see how each codon maps to a secondary structure type.