Instalar com Codex ou Claude Copie este prompt, cole no Codex, Claude ou outro assistente e deixe que ele revise a página da skill e instale para você.
Um comando direto ignora o prompt de revisão. Verifique a origem antes de executá-lo.
Biological sequence analysis — gene/protein sequence retrieval (NCBI, Ensembl, UniProt), nucleotide/protein search, ortholog discovery, and FASTQ QC + alignment workflows (Trimmomatic, BWA, samtools, coverage depth). Use for sequence retrieval, sequence comparison, FASTQ QC analysis, and read alignment pre-processing.
disable-model-invocation
true
Biological Sequence Analysis
⚠️ TOP-OF-MIND RULE: Trimmomatic "reads completely discarded" = F + R + 2*D, summed across samples
When a question asks about Trimmomatic "reads completely discarded" / "reads thrown out" /
"reads not in any output", do NOT report the Dropped field alone. Dropped counts
PAIRS where both mates failed; each pair = 2 individual reads. Plus the Forward-only and
Reverse-only buckets also discard one read per pair.
reads_discarded = sum over samples of (Forward_Only + Reverse_Only + 2 * Dropped)
❌ WRONG: sum(Dropped per sample) — typically reports ~thousands, GT is 100×+ higher
✅ RIGHT: sum(F + R + 2*D per sample)
Full counter table is in the FASTQ section below.
RULE ZERO — Check for pre-computed results FIRST
Before following any instruction below, scan the data folder for:
*_executed.ipynb → read with tu run read_executed_notebook '{"data_folder":"<path>","search":"<keyword>"}' and cite its cell outputs as the authoritative answer
Pre-computed result files (CSV/TSV with names like *results*, *deseq*, , , ) → read directly and report the requested value
*enrich*
*stats*
*_simplified.csv
Canonical analysis scripts (analysis.R, run_*.py, find_*.R, *.Rmd) → execute as-is and read the output
Only follow this skill's re-analysis recipe below if none of the above exist. Re-running from raw data produces different numbers than the published answer and is much slower (often 5-10× turn count).
Retrieve, annotate, and compare biological sequences from NCBI, Ensembl, and UniProt. Covers nucleotide search, sequence fetching, gene summaries, ortholog discovery, and protein sequence extraction.
FASTQ QC, Trimmomatic, and read alignment (when raw reads are present)
When the data folder has *.fastq files and the question involves Trimmomatic, BWA, samtools, FastQC, or coverage depth, this skill is the entry point — but the actual work is shell-level (no specific ToolUniverse data tool).
Trimmomatic PE — counting "completely discarded" reads
Trimmomatic PE classifies each input pair as:
Both Surviving (B): R1 and R2 both pass → kept as paired
Forward Only Surviving (F): R1 passes, R2 dropped → R1 kept as singleton, R2 fully discarded
Reverse Only Surviving (R): R2 passes, R1 dropped → R2 kept as singleton, R1 fully discarded
Dropped (D): both fail → BOTH R1 and R2 fully discarded
Counter selection — read the question carefully. CRITICAL: "READS completely discarded" ≠ "PAIRS dropped". The Trimmomatic Dropped count counts PAIRS (each = 2 individual reads). When the question asks about reads (not pairs), translate every counter to per-read terms:
Question phrasing
Formula (per-sample, then SUM across all samples)
"reads completely discarded", "reads thrown out", "reads not in any output"
F + R + 2*D (every individual R1 or R2 not in any output FASTQ)
"read pairs dropped", "pairs where both mates failed"
D
"individual R2 reads dropped" (R1 kept as singleton)
F
"individual R1 reads dropped" (R2 kept as singleton)
R
"reads passing QC" / "surviving reads"
2*B + F + R
Trimmomatic's stderr summary gives Input Read Pairs: N Both Surviving: B (b%) Forward Only Surviving: F (f%) Reverse Only Surviving: R (r%) Dropped: D (d%). Always sum across ALL input sample pairs (e.g., SRR1 + SRR2 + ...).
DO NOT report just D as "reads completely discarded" — that's pair count, not read count, and is off by ~100×. The "Forward Only" R2 mate IS discarded; the "Reverse Only" R1 mate IS discarded; "Dropped" pairs lose BOTH reads.
Coverage depth (samtools depth / mosdepth)
For "average coverage depth", run samtools depth -a alignment.bam | awk '{sum+=$3; n++} END {print sum/n}' — the -a flag includes positions with zero coverage (otherwise the average is biased upward). For per-chromosome coverage, group by $1.
When to Use
"Get the mRNA sequence for BRCA1"
"Search NCBI for E. coli K-12 complete genome"
"Find orthologs of TP53 across species"
"Fetch the protein sequence for UniProt P04637"
"Get the CDS sequence for Ensembl transcript ENST00000269305"
NCBIGene_search: term (string REQUIRED, format "TP53[Symbol] AND Homo sapiens[Organism]"), retmax (int, default 10). Returns {status, data: {esearchresult: {idlist: ["7157"]}}}.
NCBIGene_get_summary: id (string REQUIRED, e.g., "7157"). Returns {status, data: {result: {"7157": {name, description, summary, chromosome, maplocation, genomicinfo, mim}}}}. Result is keyed by gene ID string.
EnsemblCompara_get_orthologues: gene (gene symbol, e.g., "CFTR"), species (e.g., "human"). User-friendly alternative to NCBIDatasets_get_orthologs — accepts gene symbols directly.
Phase 6: Variant and Clinical Context
EnsemblVEP_annotate_hgvs: hgvs_notation (e.g., "NM_000492.4:c.1521_1523del"). Returns consequence, protein impact, genomic coordinates.
ClinVar_search_variants: gene (gene symbol). Returns variant count and IDs for clinical significance lookup.
PubMed_search_articles: query, limit. Literature context for gene/variant findings.
Tool Parameter Quick Reference
Tool
Correct Param
Common Mistake
NCBIGene_search
term (with [Symbol] syntax)
query or gene
NCBIGene_get_summary
id (string)
Integer type
NCBI_fetch_accessions
uids (array)
accessions
NCBI_get_sequence
accession (string)
Passing UID
NCBIDatasets_get_orthologs
gene_id (string)
Gene symbol
EnsemblSeq_get_id_sequence
ensembl_id
id
ensembl_get_sequence
id + multiple_sequences
Omitting multiple_sequences for gene+CDS
UniProt_get_sequence_by_accession
accession
Response is result not data
Fallbacks
Gene not found -> try NCBIDatasets_get_gene_by_symbol with explicit taxon
No accessions from search -> broaden query (remove strain/seq_type filters)
Ensembl error for gene+CDS -> use transcript ID (ENST) or set multiple_sequences=true
UniProt accession unknown -> NCBIDatasets_get_gene or UniProt_search for cross-refs
Ortholog search empty -> verify gene_id is numeric NCBI Gene ID
Sequence Analysis Reasoning (CRITICAL)
LOOK UP DON'T GUESS -- always fetch sequences, coordinates, and domain boundaries from databases. Do not reconstruct them from memory.
When to Use Which Tool
Question Type
Tool Choice
Why
"Find similar sequences"
BLAST_protein_search
Homology search against databases; returns E-values and identity
"What domains does this protein have?"
InterPro_get_entries_for_protein or Pfam_get_protein_annotations
Domain architecture with exact residue coordinates
"Get the sequence of gene X"
NCBI_search_nucleotide -> NCBI_get_sequence
Nucleotide retrieval by gene name
"Compare orthologs"
NCBIDatasets_get_orthologs or EnsemblCompara_get_orthologues
Cross-species gene comparison
"What is the protein impact of variant X?"
EnsemblVEP_annotate_hgvs
Consequence prediction with protein coordinates
"Align two sequences"
BLAST (pairwise)
Quick pairwise comparison with scoring
Reading Frame Selection Strategy
When translating a DNA sequence to protein:
Do NOT guess the reading frame -- preferred: use DNA_translate_reading_frames tool; fallback: translate_dna.py which tries all 3 frames automatically
The correct frame is the one with the LONGEST open reading frame (no premature stops)
If the sequence starts with ATG, frame 1 is likely correct -- but verify
If all 3 frames have early stop codons, the sequence may be: (a) non-coding, (b) reversed, or (c) contains sequencing errors. Try reverse complement first.
Protein Domain Interpretation
When asked about protein function or structure:
Get domain architecture first: InterPro_get_entries_for_protein returns all annotated domains with positions
Variants in conserved domains are more likely pathogenic than those in linker regions
LOOK UP domain boundaries from the database -- do not estimate positions from memory
Reasoning for Protein Feature Questions
When asked "how many X residues in region Y of protein Z":
Identify the correct protein — Gene names are ambiguous. GABAA has many subunits (GABRA1, GABRB2, GABRR1...). Read the question carefully for the specific subunit. Use proteins_api_search with gene name + "human" to find the right accession.
Find the region boundaries — Use proteins_api_get_features with the accession to get annotated domains (TRANSMEM, DOMAIN, REGION). Don't guess positions — get them from the database.
Count residues in the region — Fetch the sequence, extract the region, count. WRITE Python code for this — don't try to count manually.
For residue counting questions, ALWAYS use the script or sequence[start:end].count('C'). Do NOT estimate or count from memory.
Account for multimers — READ THE QUESTION for "homomeric", "pentamer", "tetramer", "dimer". If the question asks about a homomeric receptor (e.g., "homomeric GABAAρ1"), every subunit is identical. Count the residues in ONE subunit, then multiply:
Homomeric pentamer (most ligand-gated ion channels like GABAA ρ1): × 5
Homotetramer (many ion channels): × 4
Homodimer: × 2
If the question says "in the TM3-TM4 linker domains" (plural), it means across all subunits in the complex.
Bundled Computation Scripts
Never manually count residues, compute GC%, or write reverse-complement logic inline. Run these scripts instead — they are tested and handle edge cases.
Use this script to look up commonly-confused biology facts instead of relying on memory. It covers receptor types, ion channel stoichiometry, neurotransmitters, immune cell markers, and gene naming confusions.
Use this script for any question about the genetic code, codon degeneracy, amino acid chemistry, codon usage bias, or tRNA wobble pairing. All outputs are JSON.
Any question about how many codons encode a given amino acid (degeneracy)
Any question about rare vs. common codons for protein expression optimisation
Any question about tRNA anticodon recognition / wobble base pairing
Any question about amino acid physical-chemical properties (MW, pKa, hydrophobicity, polarity, charge)
Any question about the names of stop codons (Amber/Ochre/Opal)
Before manually stating codon degeneracy — verify with codon_table
Wobble rules: I pairs U/C/A (3 codons); G pairs U/C; U pairs A/G; C pairs G only; A pairs U only (rare). Use --type wobble --anticodon "GAU" to verify.
Amino acid lookup: accepts full name (--name "Cysteine"), 1-letter (--code C), or 3-letter (--code CYS).
Codon-Anticodon Matching Reasoning (CRITICAL for tRNA problems)
When solving "which codons does this tRNA recognize" or "which tRNA reads this codon":
Anticodon is written 3'->5' but conventionally listed 5'->3'. The FIRST position of the anticodon (5' end) is the WOBBLE position and pairs with the THIRD position of the codon (3' end).
Anticodon-codon pairing is ANTIPARALLEL: anticodon 5'-X-Y-Z-3' pairs with codon 3'-X'-Y'-Z'-5' (i.e., codon 5'-Z'-Y'-X'-3').
Wobble position rules (anticodon 5' base -> codon 3' base it can pair with):
C -> G only (1 codon)
A -> U only (1 codon; rare in bacteria, common in mitochondria)
U -> A or G (2 codons)
G -> C or U (2 codons)
I (inosine, deaminated A) -> U, C, or A (3 codons)
Minimum tRNA set: Because I reads 3 bases and G/U each read 2, a 4-codon family (e.g., GCN = Ala) needs only 2 tRNAs: one with I at wobble position (reads 3 of 4 codons) and one with C or U at wobble (reads the remaining 1-2).
ALWAYS use the script: python3 skills/tooluniverse-sequence-analysis/scripts/amino_acids.py --type wobble --anticodon "IAU" to verify rather than reasoning from memory.
translate_dna.py — DNA to protein translation
Preferred: use DNA_translate_reading_frames tool (via MCP/SDK) with sequence parameter. Fallback: run translate_dna.py directly.
count_residues: Count residue in full sequence. --sequence "ACDE..." --residue C
count_region: Count in region (1-based inclusive). --sequence "MAC..." --start 5 --end 20 --residue C OR --accession P24046 --start 318 --end 440 --residue C (fetches from UniProt live)
gc_content: GC% of DNA. --sequence "ATGCGATCG"
reverse_complement: DNA reverse complement. --sequence "ATGCGATCG"
stats: Auto-detect DNA/RNA/Protein, compute length, MW for protein. --sequence "ATGCG..."
ALWAYS use count_region --accession when the user gives a UniProt accession + region -- do not count manually.
Interpretation Framework
Sequence Quality Assessment
Indicator
High Quality
Acceptable
Caution
RefSeq status
NM_/NP_ (curated)
XM_/XP_ (predicted)
No RefSeq (GenBank only)
Sequence version
Latest version (.N)
Previous version
Removed/replaced
Annotation
Reviewed (UniProt Swiss-Prot)
Unreviewed (TrEMBL)
No annotation
Gene symbol
HGNC approved
Alias/synonym
Locus tag only
Synthesis Questions
Is this the correct sequence? (verify organism, gene symbol, isoform)
Is it the canonical isoform? (RefSeq MANE Select or UniProt canonical)
How well-annotated is it? (SwissProt > TrEMBL > GenBank predicted)
Are there known variants? (ClinVar pathogenic variants in this sequence)
Answer Formatting (CRITICAL)
TRIM YOUR ANSWER: If the question asks "what protein", answer with JUST the protein name. Do not add parenthetical abbreviations, descriptions, or qualifications. Example: answer "Glucose-6-phosphate 1-dehydrogenase", NOT "Glucose-6-phosphate 1-dehydrogenase (G6PD, EC 1.1.1.49)". When identifying a protein from a sequence, use BLAST/UniProt and report the top hit name exactly as it appears in the database — no embellishment.
