How is PicoFold different from PSIPRED and JPred?

PSIPRED and JPred predict secondary structure from protein amino acid sequence. PicoFold predicts from DNA codons — the only tool with this capability. PicoFold can distinguish synonymous codons (same amino acid, different DNA). However, PSIPRED (~84% Q3) and JPred (~82% Q3) have significantly higher accuracy than PicoFold (37.7% Q3).

Why is PicoFold's accuracy lower than PSIPRED?

PicoFold uses a deterministic physics-based model mapping each DNA codon to a secondary structure type, without machine learning or evolutionary profiles. Its Q3 of 37.7% on 113 proteins is above random baseline (34.7%) with confirmed helix signal, but below ML tools. PicoFold answers a different question: how does codon choice relate to secondary structure.

Can PicoFold replace AlphaFold or PSIPRED?

No. PicoFold is complementary. Use PSIPRED or JPred for accurate SS prediction from protein sequence. Use AlphaFold for 3D structure. Use PicoFold when you want to explore how DNA codon choice relates to secondary structure — for research, education, or studying synonymous codon effects.

What can PicoFold do that no other tool can?

PicoFold is the only tool that predicts secondary structure from DNA codons. It can show how synonymous codon substitutions affect predicted SS, evaluate codon optimization choices, detect potential effects of synonymous SNPs, and study the codon-SS relationship. No protein-based tool can do this.

Is PicoFold useful for codon research despite lower accuracy?

PicoFold provides a unique data point for codon research: it maps each codon to a secondary structure type, so different synonymous codons produce different SS assignments. Chi-square analysis shows 6/18 amino acid families have significant codon-SS associations (Ser p=0.013, Gly p=0.016). This is a research signal, not a production-grade predictor — use it alongside established tools.

Comparison

Do synonymous codons affect protein structure?

Q: Do synonymous codons affect protein structure?

Evidence increasingly says yes. PicoFold is the only tool that lets you explore this: it maps each DNA codon to a secondary structure type, so synonymous codons produce different SS assignments. Validation on 113 proteins shows 6/18 amino acid families with significant codon-SS association. Independent literature (Nature Communications 2022, PNAS 2020) confirms synonymous codons affect backbone geometry and protein properties.

PicoFold is a research tool for studying the codon–SS relationship. It maps DNA codons to secondary structure types — no other tool does this. Not a replacement for PSIPRED or AlphaFold. A different instrument for a different question.

Silent mutations change the codon but not the amino acid — do they affect structure?

Synonymous codons encode the same amino acid but use different DNA triplets. Protein-based tools see them as identical. PicoFold sees the difference.

Example: two alanine codons

GCT Ala (T-ending)

GCC Ala (C-ending)

→

PSIPRED / JPred Same result

GCT Ala (T-ending)

GCC Ala (C-ending)

→

PicoFold Different SS types

GCT → pi-helix (T-ending) · GCC → alpha-helix (C-ending) — same amino acid, different predicted structure.

Who benefits

Five audiences, one unique capability

PicoFold is the only tool that gives structural feedback at the DNA codon level.

Codon–Structure Researchers

Study how synonymous codons — same amino acid, different DNA — correlate with different secondary structure types. PicoFold is the only tool that maps codons to SS, letting you explore this relationship across any gene. Chi-square validation confirms the signal in 6/18 amino acid families.

Codon Usage Bias Researchers

Explore how organism-specific codon preferences correlate with protein secondary structure composition. Compare codon bias patterns across species. PicoFold turns codon tables into structural mappings — a unique data layer for codon usage analysis.

Molecular Biology Educators

Demonstrate to students that DNA codons carry more information than amino acid identity alone. Show how a silent mutation changes the SS mapping. Instant, visual, runs in the browser — no installation required. The best honest use case for PicoFold.

Synonymous Mutation Researchers

Investigate whether synonymous SNPs — invisible to protein-based tools — show different codon–SS mappings. PicoFold provides a hypothesis-generating signal, not a clinical prediction. Explore the data, then validate experimentally.

Structural Bioinformatics Labs

Add a unique data point to your analysis pipeline. PicoFold's codon–SS mapping provides a complementary signal that no protein-sequence tool offers. Useful as an additional feature in multi-method research, not as a standalone predictor.

Side by side

Detailed comparison

PicoFold is not a replacement for PSIPRED or AlphaFold — it answers a different question.

	PicoFold	PSIPRED 4.0	JPred 4	NetSurfP-2.0	AlphaFold 3
Input	DNA (FASTA / EMBL / GenBank)	Protein sequence	Protein sequence	Protein sequence	Protein sequence
Method	Ring-Granular Model (deterministic lookup)	Neural network + PSI-BLAST profiles	Jnet neural network + HMM profiles	ESM-2 language model embeddings	Evoformer deep learning
Primary output	SS from codons (H/E/C) + PDB	SS (H/E/C)	SS (H/E/C) + burial	SS + RSA + disorder	3D structure (SS derived)
Q3 accuracy	37.7%*	~84%	~82%	~85%	N/A (3D tool)
Speed	<10 ms	1–5 min	30–60 s	~5 s	Minutes–hours
GPU required	No	No	No	Recommended	Yes
Deterministic	Always	Yes†	Yes†	Yes†	No
Synonymous codon sensitivity	Yes (unique)	No	No	No	No
Commercial API	Yes	Web only	Web only	Yes	Restricted
Self-hosted	Docker	Open source	No	Open source	Open source

* PicoFold Q3 measured on 113 proteins (21,797 residues) against DSSP ground truth. Random baseline: 34.7%. See validation details.

† Deterministic for the same protein input, but cannot distinguish synonymous codon variants.

PSIPRED: Buchan & Jones, 2019. JPred: Drozdetskiy et al., 2015. NetSurfP-2.0: Klausen et al., 2019.

Recommendations

When to use what

Different tools for different questions. Here's our honest recommendation.

Best SS accuracy

PSIPRED / JPred / NetSurfP

You have a protein sequence and need the most accurate SS prediction
You don't need to compare synonymous codon variants
You want a well-established, peer-reviewed benchmark
Q3 accuracy matters more than speed

3D structure

AlphaFold / ESMFold

You need a full 3D folded structure, not just SS assignments
You're working with known proteins that have homologs in PDB
You have GPU access and can wait minutes–hours
Highest possible structural accuracy is critical

Codon–SS research

PicoFold

You want to explore how synonymous codons map to different SS types
You're studying the codon–structure relationship in specific genes
You need a teaching tool to show students that codons carry structural information
You want instant, deterministic codon–SS mappings for research
You're investigating synonymous SNP effects at the DNA level

Transparency

Known limitations

We believe honest disclosure builds more trust than marketing.

Lower Q3 accuracy

PicoFold Q3 = 37.7% vs ~82–85% for PSIPRED/JPred/NetSurfP. The model uses a deterministic codon–SS lookup, not ML trained on thousands of structures. It is above random baseline (34.7%) and helix signal is confirmed, but absolute accuracy is significantly lower.

Beta prediction anti-correlates

The model maps A-ending codons to beta-strand, but validation shows A-ending codons have the lowest strand rate (18.8%). Beta prediction does not work in the current model. Helix prediction is the confirmed signal.

3D fold not reliable

PicoFold's experimental 3D coordinates do not match experimental structures (TM-score 0.047 on insulin, below random 0.17). The 3D viewer is a bonus visualization, not a structural prediction. Focus is on secondary structure assignments.

Best for helix-rich targets

Top results on highly helical proteins: Tropomyosin α Q3=0.82, Vimentin Q3=0.70. For beta-sheet-rich or mixed-fold proteins, accuracy drops. The model captures local helix bias from codons, not long-range strand interactions.

Deterministic, not adaptive

The Ring-Granular Model is a fixed lookup table — it cannot learn from new data. ML-based tools improve as training sets grow. PicoFold's accuracy is bounded by the model's current parameterization.

DNA input only

PicoFold requires DNA sequence (FASTA/EMBL/GenBank). If you only have a protein sequence, PicoFold cannot help — use PSIPRED or JPred instead. You need the actual coding DNA to make predictions.

Evidence: synonymous codons are not neutral for structure

Despite lower absolute accuracy, PicoFold's core hypothesis — that synonymous codons are not neutral for structure — is supported by independent research.

Chi-square validation: 6 of 18 amino acid families show statistically significant codon–SS association (p<0.05). Serine (p=0.013) and Glycine (p=0.016) survive Bonferroni correction. Data: 113 proteins, 21,797 residues, DSSP ground truth.

Nature Communications 2022: 57 of 87 synonymous codon pairs show different backbone dihedral angles in crystal structures.

PNAS 2020: Synonymous substitutions in antibodies alter solubility and binding properties.

Annual Reviews Biophysics 2024: Codon effects on cotranslational folding confirmed across multiple experimental systems.

Explore the codon–SS relationship

5 free analyses, no card required. Upload a DNA sequence and see how each codon maps to a secondary structure type.

Create free account Learn more