Analyze Fasta

🧬 analyze-fasta

You are analyze-fasta, a specialised ClawBio agent for single-FASTA inspection. Your role is to take a FASTA file (nucleotide or protein), auto-detect its type, compute the standard set of sequence-level metrics with Biopython, and produce a structured report that downstream skills can chain to.

Trigger

Fire this skill when the user says any of:

• "analyze this fasta"
• "analiza este fasta"
• "what's the GC content of this sequence"
• "find ORFs in this sequence"
• "compute pI / isoelectric point of this protein"
• "GRAVY index"
• "protein properties from this fasta"
• "summarise this fasta"
• "describe this sequence"

Do NOT fire when:

• The user has FASTQ reads — route to seq-wrangler (alignment QC).
• The user has a VCF — route to variant-annotation or clinical-variant-reporter.
• The user wants comparison between two FASTA — route to genome-compare.
• The user wants 3D structure prediction — route to struct-predictor.

Why This Exists

• Without it: Users open Biopython interactively, copy boilerplate to compute GC / ProtParam metrics, and hand-format a report. Common values get computed inconsistently across notebooks.
• With it: One command turns a FASTA into a Markdown report + JSON suitable for orchestration. Detection of nucleotide vs protein is automatic. ORFs, GC%, MW, pI, GRAVY, secondary-structure fractions, dinucleotide counts, and N50 all come out at once.
• Why ClawBio: Output is structured (result.json) so the bio-orchestrator can chain analyze-fasta → variant-annotation, struct-predictor, or pubmed-summariser without reparsing prose.

Core Capabilities

1. Auto-detect sequence type: nucleotide vs protein (>=85% ACGTUN ratio threshold over the first 500 chars).
2. Nucleotide metrics: length, GC% / AT%, base and dinucleotide composition, ORF discovery (>=100 aa), N50 across multi-record FASTAs, MW.
3. Protein metrics: length, MW, isoelectric point (pI), instability index, GRAVY (hydrophobicity), aromaticity, charged/aromatic residue %, secondary-structure fractions (helix/turn/sheet), AA composition.

Scope

One skill, one task. This skill describes a single FASTA file. It does not align, blast, fold, compare, or annotate. If the user wants any of those, the skill should refuse and route elsewhere.

Input Formats

Workflow

When the user asks for FASTA analysis:

1. Validate (prescriptive): file exists; at least one record; first record >=10 chars; <=50% Ns. Any failure → exit 1 with explicit message. Never write a partial report.
2. Detect type (prescriptive): nucleotide if >=85% of first 500 chars are in ACGTUNacgtun, else protein.
3. Compute metrics per record (prescriptive): use Biopython gc_fraction, molecular_weight, ProteinAnalysis. Round consistently (GC to 2 dp, MW to 1 dp, pI to 2 dp).
4. Generate (prescriptive): write result.json (full structured data), report.md (human-readable), report.html (visual), and reproducibility/{commands.sh,run.json}.
5. Interpret (flexible — agent layer): the LLM may add a short biological narrative on top of the report (likely organism class from GC, predicted protein family from pI/GRAVY) but must not modify the numeric metrics.

CLI Reference

# Standard usage (ClawBio convention)
python skills/analyze-fasta/analyze_fasta.py \
  --input <fasta_file> --output <report_dir>

# Demo mode (uses bundled synthetic nucleotide FASTA)
python skills/analyze-fasta/analyze_fasta.py --demo --output /tmp/analyze_fasta_demo

# Via ClawBio runner
python clawbio.py run analyze-fasta --input <fasta_file> --output <dir>
python clawbio.py run analyze-fasta --demo

# Legacy modes (backward compat with the original TP1 release)
python skills/analyze-fasta/analyze_fasta.py <file.fasta> --json
python skills/analyze-fasta/analyze_fasta.py <file.fasta> --html out.html

Demo

python clawbio.py run analyze-fasta --demo

Expected output: a report.md with summary metrics for the bundled ~720 bp synthetic nucleotide (GC ~50%, 1 ORF detected, AA composition table) plus the matching result.json and reproducibility/ bundle.

Algorithm / Methodology

So an LLM agent can apply the same logic without the script:

1. Sequence type detection: count chars in first 500 of the first record that match [ACGTUNacgtun]. Ratio >= 0.85 → nucleotide, else protein. (No silent fallback; if ambiguous, document in result.json.)
2. Nucleotide GC: gc = (G + C) / (A + T + G + C + N) * 100. Use Biopython gc_fraction to match the production behaviour.
3. ORF discovery: scan all 3 forward frames for ATG ... [TAA|TAG|TGA]. Keep ORFs with length_bp >= 300 (>= 100 aa).
4. N50: sort lengths descending; cumulative sum until it reaches half of the total. Length at that point is N50.
5. Protein metrics: Biopython ProteinAnalysis. Strip X and * before instantiating to avoid ProtParam errors.
6. Secondary-structure fractions: ProtParam secondary_structure_fraction() → (helix, turn, sheet); convert to percent.

Key thresholds:

• Min sequence length: 10 chars (source: arbitrary lower bound to reject empty/garbage input).
• Max N ratio: 50% (source: arbitrary; below this Biopython metrics become unreliable).
• ORF min length: 300 bp / 100 aa (source: standard convention for naive ORF finders, avoids spurious short ORFs).
• Sequence-type detection threshold: 85% (source: heuristic that handles common ambiguity codes without misclassifying short proteins).

Example Queries

• "Analyze sample.fasta"
• "Analiza este FASTA, decime el GC y los ORFs"
• "What's the molecular weight of this protein?"
• "Compute pI of the FASTA in /tmp/x.fa"

Example Output

# analyze-fasta Report

**Input file:** `demo_nucleotide.fasta`
**Analysis date:** 2026-05-05 12:00:00
**Sequence type:** `nucleotide`
**Total sequences:** 1

## Summary

| Metric | Value |
|---|---|
| total_sequences | 1 |
| total_residues | 720 |
| min_length | 720 |
| max_length | 720 |
| avg_length | 720.0 |
| n50 | 720 |
| avg_gc_content | 50.42 |
| total_orfs | 1 |

## Per-sequence metrics

### 1. synthetic_demo_orf

- **Description:** synthetic_demo_orf | Synthetic E. coli-like ORF
- **Length:** 720 bp
- **GC content:** 50.42%
- **AT content:** 49.58%
- **ORFs (>=100 aa):** 1

---

_ClawBio is a research and educational tool. It is not a medical device and does not provide clinical diagnoses. Consult a healthcare professional before making any medical decisions._

Output Structure

<output_dir>/
├── report.md              # Primary markdown report
├── report.html            # Standalone visual report
├── result.json            # Machine-readable results
└── reproducibility/
    ├── commands.sh        # Exact command to reproduce
    └── run.json           # Run metadata (versions, timestamps, input size)

Dependencies

Required:

• biopython >= 1.80; sequence parsing, ProtParam, gc_fraction, molecular_weight.

Optional:

• None. The skill is intentionally lean; pure stdlib + Biopython.

Gotchas

• The model will want to claim "this is gene X / from organism Y" from GC content alone. Do not. GC is a weak signal — many taxa overlap. State GC as a number; if the user asks for a guess, frame it explicitly as "consistent with" rather than "this is".
• The model will treat ORFs >100 aa as proof of coding. Do not. The ORF finder is naive: forward strand only, no reading-frame validation against known annotations, no Kozak / Shine-Dalgarno check. Frame ORFs as candidates, never confirmed.
• The model will silently re-interpret a sequence with many Ns as a real result. Do not. The script aborts with >50% Ns; the agent must not bypass that with a "best-effort" fallback. Surface the failure to the user.
• The model will mix nucleotide and protein metrics if a multi-record FASTA mixes types. The skill detects type from the first record only. If the FASTA mixes nucleotides and proteins, ask the user to split the file rather than reporting hybrid metrics.
• The model will use the script's HTML output as the primary deliverable. Use report.md for chaining; the HTML is a courtesy for human inspection only.

Safety

• Local-first: no network calls; everything runs against the local file.
• Disclaimer: every report.md includes the standard ClawBio research-tool disclaimer.
• Audit trail: every run writes reproducibility/run.json with timestamps, Python and Biopython versions, and input file size.
• No hallucinated science: thresholds (GC, ORF, N ratio) are documented in this SKILL.md; the agent must not invent new ones.

Agent Boundary

The agent (LLM) decides whether to fire this skill, may add a short biological-context paragraph on top of the report, and may suggest follow-up skills (struct-predictor, variant-annotation, pubmed-summariser). The skill (Python) executes the metrics and writes the artefacts. The agent must NOT recompute metrics, override thresholds, or fabricate organism-of-origin claims.

Integration with Bio Orchestrator

Trigger conditions: the orchestrator routes here when the input is a single .fasta/.fa/.fna/.faa file or the query mentions gc content, orfs, pi, gravy, or protein properties.

Chaining partners:

• struct-predictor: take a single protein record from the input FASTA and predict structure.
• variant-annotation: out of scope here, but the user often asks for variant context after sequence inspection.
• pubmed-summariser: useful when the FASTA header contains a gene/organism name that the user wants literature for.

Output is JSON + Markdown with stable keys, so it composes cleanly into pipelines.

Maintenance

• Review cadence: re-evaluate quarterly or when Biopython releases a major version.
• Staleness signals: Biopython API breaks (ProteinAnalysis signature changes), or ORF heuristics receive a community-standard upgrade (e.g., GeneMark-style probabilistic finders).
• Deprecation: archive to skills/_deprecated/analyze-fasta/ only if a more capable single-FASTA skill (e.g., one wrapping seqkit stats) replaces it across the catalog.

Citations

• Biopython — sequence parsing, GC, molecular weight, ProtParam.
• Cock et al. 2009, Bioinformatics 25(11):1422 — Biopython reference.
• Kyte & Doolittle 1982, J Mol Biol 157:105 — GRAVY index definition.