Claude Skills by GPTomics

Selects biomarker features from high-dimensional omics data using Boruta all-relevant selection, mRMR, LASSO/elastic-net, and stability selection, while controlling the leakage, irreproducibility, and correlated-feature traps that make most published signatures fail to replicate. Use when identifying candidate biomarkers, deciding between an all-relevant and a minimal-optimal selector, or judging whether a selected gene set is reproducible. For unbiased performance estimation of the resulting...

Validates predictive models on omics and biomedical data with nested cross-validation, group/batch/temporal-aware splits, the full data-leakage taxonomy, probability calibration, decision-curve net benefit, optimism correction, sample-size planning, and TRIPOD+AI reporting. Use when estimating model performance honestly, choosing a CV scheme, detecting leakage, or judging whether reported discrimination means the model is actually useful. For feature selection itself see machine-learning/biom...

Builds diagnostic and prognostic classifiers on omics feature matrices with regularized logistic regression, random forest, and gradient-boosted trees, handling the p>>n regime, batch shortcut learning, class imbalance, and probability calibration. Use when building a classifier from expression, methylation, or variant data, choosing an algorithm for high-dimensional small-n data, or diagnosing a suspiciously perfect AUC. For unbiased evaluation see machine-learning/model-validation; for feat...

Explains ML predictions on omics data with SHAP, LIME, and permutation importance, handling the correlated-feature trap, the conditional-vs-interventional Shapley choice, and the attribution-is-not-causation boundary. Use when interpreting an omics classifier, debugging shortcut/batch learning, or deciding whether an attribution ranking can be trusted as biology. For validated feature selection see machine-learning/biomarker-discovery; explanations are not a selection method.

Designs and analyzes stable-isotope-resolved metabolomics (SIRM / isotope tracing / fluxomics) experiments that measure metabolic ACTIVITY via 13C/15N/2H tracers, distinct from steady-state pool profiling. Covers tracer choice, isotopologue vs isotopomer, mass-isotopomer distributions (MID), fractional enrichment, the mandatory natural-abundance + tracer-purity correction (IsoCor, AccuCor), and the metabolic/isotopic steady-state vs non-stationary (INST-MFA) distinction. Use when feeding a la...

Assigns honest lipid annotation levels, designs class-based internal-standard quantification, and runs lipid-aware differential and enrichment analysis with lipidr, guarding against in-source-fragment phantoms, sn-position over-claims, and invalid cross-class quantification. Use when naming or canonicalizing lipid species (shorthand separators, Goslin), deciding shotgun vs RP vs HILIC LC-MS, picking internal standards (SPLASH/EquiSPLASH), interpreting MS-DIAL/LipidSearch output, or comparing ...

Turns untargeted LC-MS/MS features (m/z, RT, MS/MS) into confidence-stratified metabolite annotations using spectral-library matching (matchms), in-silico tools (SIRIUS/CSI:FingerID, MetFrag) and molecular networking, and assigns a defensible MSI/Schymanski confidence level to each. Use when naming detected features, scoring MS/MS against a reference library, running SIRIUS, or deciding what confidence level an evidence set actually supports. For upstream feature extraction see metabolomics/x...

Runs the MS-DIAL preprocessing workflow (peak picking, MS2Dec spectral deconvolution, alignment, gap-filling) and imports the alignment-result table into R or Python with honest filtering. Use when preprocessing LC-MS DDA/DIA (SWATH) raw data with MS-DIAL, deciding MS-DIAL vs XCMS, configuring the MsdialConsoleApp console run, or parsing an MS-DIAL export into a clean feature matrix. For programmatic R peak detection and the feature-table-as-artifact framing see metabolomics/xcms-preprocessin...

Designs QC, corrects signal drift, removes batch effects, filters features, normalizes samples, and imputes missing values for untargeted LC-MS/GC-MS metabolomics, framing each step as a measurement model that can create or erase biological signal. Use when processing a peak/feature table before statistical analysis, choosing a drift-correction or sample-normalization method, deciding QC RSD vs D-ratio filtering, or handling left-censored missing values. The feature table is produced by metab...

Maps metabolomics results to biological pathways via over-representation (ORA), metabolite-set enrichment (MSEA/QEA), mummichog/PSEA on raw m/z peaks, and network-diffusion enrichment (FELLA), with correct background-set construction and honest interpretive ceilings. Use when interpreting differential metabolites or an untargeted LC-MS feature table in pathway context, choosing ORA vs MSEA vs mummichog vs topology, or setting the reference/background set. For annotation confidence levels feed...

Designs and validates quantitative targeted metabolomics assays (MRM/SRM on triple-quadrupole, PRM on high-resolution instruments) to report absolute concentrations. Covers the internal-standard strategy (external cal -> global IS -> standard addition -> stable-isotope-labeled IS), weighted calibration judged by back-calculated %RE not R-squared, ion-ratio quantifier/qualifier confirmation, matrix-effect/recovery characterization, and ICH M10 method validation. Use when quantifying a closed p...

Programmatic untargeted LC-MS feature extraction in R with the modern xcms 4.x MsExperiment/XcmsExperiment API, taking raw mzML to a feature table via CentWave peak detection, retention-time alignment, peak-density correspondence, gap-filling, CAMERA redundancy collapse, and built-in QC feature filtering. Use when converting centroided LC-MS runs into a features-by-samples matrix and deciding centWave/grouping/alignment parameters. For drift correction and QC/CV filtering execution see metabo...

Turns shotgun classifier output into a defensible abundance table with Bracken Bayesian re-estimation, then compositional treatment (CLR, zero handling), library-size normalization, reference-frame differential abundance, and optional absolute quantification. Covers why a relative-abundance change is not a change, why Bracken read fractions and MetaPhlAn percentages are different physical quantities, the silent -r read-length bias, the genome-size confound no library-size method fixes, and th...

Profiles the antimicrobial-resistance gene content (resistome) of shotgun metagenomes - read-based quantification with RGI bwt, AMR++/MEGARes, ARGs-OAP/SARG, deepARG, or GROOT, and presence calling with AMRFinderPlus/ABRicate on assembled contigs or MAGs. Covers why an ARG hit is a sequence match not a phenotype, why a metagenomic ARG has no host and no genomic context until assembly (and assembly breaks at ARGs), per-gene curated thresholds vs a flat 80/80, gene-fraction false-positive contr...

Cleans a shotgun metagenome of everything that is not the target community before profiling - host-read depletion (Hostile, bowtie2/T2T-CHM13), reagent/kitome contamination control with blanks and decontam, mock-community validation, and depth-adequacy checks (Nonpareil). Covers why a metagenomic result is a position in a choice-chain rather than a direct observation, why extraction is the experiment, why a low-biomass community can be entirely kitome, why absence means not-detectable-by-this...

Profiles the functional potential of shotgun metagenomes with HUMAnN 3's tiered search (MetaPhlAn prescreen, Bowtie2 pangenome, translated DIAMOND vs UniRef), giving gene-family (RPK) and MetaCyc pathway abundances stratified by species. Covers why a metagenome measures potential not activity, why dropping UNMAPPED/UNINTEGRATED biases everything, why stratification is an estimate, coverage-vs-abundance and MinPath/gap-fill, UniRef90-vs-50 and biome database bias, and the assembly/eggNOG/dbCAN...

Classifies shotgun metagenomic reads to taxa with Kraken2's minimizer/LCA matching against a chosen reference database, then hands off to Bracken for abundance re-estimation. Covers why the database (not the algorithm) decides what can be detected, the --confidence and --minimum-hit-groups precision levers, unique-minimizer false-positive control, host-read removal, and why raw Kraken2 read counts are not abundances. Use when profiling who-is-there from shotgun reads, choosing a Kraken2 datab...

Turns a shotgun profiler table (MetaPhlAn relative abundance, Bracken counts, HUMAnN function tables) into honest figures and defensible community statistics with phyloseq, vegan, microViz, and Python. Covers why an ordination/bar/diversity number is a modeling choice that can manufacture a result, the MetaPhlAn-percent-vs-Bracken-counts fork that decides everything, CLR/Aitchison vs Bray-Curtis, Hill numbers and why shotgun richness is a database readout, pairing PERMANOVA with betadisper, a...

Profiles shotgun metagenomes to species/SGB relative abundance with MetaPhlAn 4's clade-specific marker genes (bowtie2 short reads, minimap2 long reads). Covers why a MetaPhlAn percentage is a cell fraction (genome-size-normalized taxonomic abundance) and must never be merged with Kraken/Bracken read fractions, kSGB vs uSGB units for quantifying database-absent taxa, the unknown-fraction rescaling and its version-default flip, --index pinning as a batch variable, and when mOTUs3 or sourmash g...

Resolves and compares bacterial strains below the species level from shotgun metagenomes with inStrain (popANI/conANI microdiversity), StrainPhlAn (marker-SNV consensus phylogeny and nGD), MIDAS2, metaSNV, and StrainGE, plus genome-vs-genome ANI (skani/fastANI/MASH) for isolate/MAG comparison. Covers why a strain is a threshold not a thing, why ANI answers same-genome while popANI/nGD answer same-population-in-situ, the 99.999% popANI and per-species nGD definitions, the coverage detection li...

Turns raw Illumina Infinium methylation BeadChip IDATs (450K, EPIC, EPICv2) into a defensible beta/M matrix with sesame (openSesame/SigDF) or minfi (RGChannelSet -> MethylSet -> GenomicRatioSet). Covers Type I vs Type II probe chemistry and why raw Type II beta is compressed, the signal-to-beta math (beta = M/(M+U+100)) and M-value logit, detection-p / pOOBAH masking including the out-of-band deletion-artifact catch, dye-bias correction, and the normalization decision (noob, funnorm, quantile...

Aligns bisulfite-converted (WGBS, RRBS, PBAT) and enzymatic (EM-seq) short reads to an in-silico C->T/G->A-converted reference with Bismark (Bowtie2 or HISAT2), preparing the genome index, choosing the directional vs non-directional vs PBAT strand flag, deduplicating WGBS/EM-seq (never RRBS), and bounding bisulfite conversion efficiency with unmethylated lambda and methylated pUC19 spike-ins. Covers why the library protocol (not the aligner) decides whether calls are meaningful, why incomplet...

Estimates cell-type composition from bulk DNA methylation and uses it to defuse the single biggest EWAS confounder. Covers reference-based deconvolution (Houseman constrained-projection, minfi estimateCellCounts2 with FlowSorted.Blood.EPIC + IDOL-optimized libraries, EpiDISH RPC/CBS/CP, 12-cell extended, cord-blood nRBC references, EpiSCORE/hepidish for solid tissue), reference-free correction (ReFACTor, RefFreeEWAS, SVA), using fractions as covariates vs the compositionality/collinearity tra...

Tests individual CpG sites for differential methylation (DMC/DMP) from bisulfite sequencing counts or array/continuous beta-value matrices. Covers the count-vs-continuous fork that dictates the model, beta-value vs M-value logit (Du 2010), beta-binomial overdispersion count models (DSS, methylKit, MOABS, RADMeth) for sequencing, limma moderated-t on M-values (eBayes trend/robust) for arrays, the bare-beta Welch t-test caveat, coverage-as-precision coupling, delta-beta effect size, BH-FDR with...

Detects differentially methylated regions (DMRs) from short-read bisulfite (WGBS/RRBS), array, and long-read methylation count tables using dmrseq (permutation region-FDR over the region selection), DSS callDMR (beta-binomial), methylKit tiles, bsseq BSmooth, DMRcate Gaussian-kernel smoothing, metilene, and comb-p. Covers why a DMR is DEFINED by arbitrary thresholds (min-CpGs, max-gap, delta-beta, q) and a smoothing bandwidth, why selecting extreme runs of CpGs then testing them on the same d...

Designs and defends an epigenome-wide association study (EWAS) on 450K/EPIC array or bisulfite methylation - the layer deciding whether a hit is credible. Covers the confounding hierarchy (cell composition covariates as the dominant confounder, batch/Sentrix chip/array position, age/sex, smoking AHRR cg05575921, ancestry/mQTL, reverse causation), chip randomization (no-rescue theorem), surrogate variable analysis sva/SmartSVA, ComBat, RUVm, over-correction, genomic inflation lambda vs GWAS ge...

Extracts per-cytosine methylation calls from aligned bisulfite/EM-seq reads with bismark_methylation_extractor (Bismark BAM) or the aligner-agnostic MethylDackel/BISCUIT (bwa-meth BAM), producing the beta value M/(M+U) as a coverage file, bedGraph, or genome-wide cytosine report across CpG/CHG/CHH context. Covers conversion-rate QC as the first gate, the 5mC vs 5hmC summed caveat, variant-aware calling so a C/T SNP does not masquerade as unmethylation, paired-end --no_overlap double-counting,...

Infers exact amplicon sequence variants (ASVs) from demultiplexed 16S rRNA or ITS amplicon FASTQ with DADA2 - removing primers with cutadapt (--discard-untrimmed), learning a per-run error model (filterAndTrim -> learnErrors -> dada -> mergePairs), merging run-level tables with mergeSequenceTables, then one removeBimeraDenovo. Covers why primers come OFF before truncation, why the error model is per-run, truncLen as a merge-overlap detection budget (V4 vs V3-V4), DADA2 vs Deblur and q2-dada2 ...

Tests which individual taxa differ between groups on an amplicon ASV/feature table (phyloseq) using compositionally-aware methods - ALDEx2 (Dirichlet-MC CLR, conservative), ANCOM-BC2/ANCOMBC (sampling-fraction bias correction, structural zeros, passed_ss, default p_adj_method=holm), MaAsLin2/MaAsLin3 (multivariable GLM, random effects, prevalence/abundance split), LinDA (CLR mixed-model regression), ZicoSeq (permutation FDR), LEfSe, and q2-composition ancombc. Covers why the hit list depends ...

Alpha and beta diversity of an amplicon (16S/ITS) ASV/OTU community table - observed features, Shannon, Pielou evenness, Faith PD, Bray-Curtis, Jaccard, weighted/unweighted/generalized UniFrac, Aitchison/RPCA - via QIIME2 core-metrics-phylogenetic, phyloseq/vegan, and scikit-bio. Covers the three knobs that set the answer before it is seen (rarefaction sampling depth, the tree, the metric), why core-metrics silently deletes samples below the sampling depth, why de novo trees lose to SEPP frag...

Predicts community functional POTENTIAL from 16S/ITS amplicon ASVs with PICRUSt2 (or q2-picrust2) by phylogenetic interpolation of reference-genome gene content - EPA-ng placement, gappa, castor hidden-state prediction of KO/EC/Pfam copy number, 16S copy-number normalization, and MinPath MetaCyc/KEGG pathways - gated by the NSTI quality index. Covers why predicted function is taxonomy re-encoded (never measured gene content and never activity), the mandatory NSTI report (--max_nsti 2 silently...

Operates the QIIME2 framework as the glue for an amplicon analysis - the .qza/.qzv artifact model, semantic types (FeatureTable[Frequency], SampleData[PairedEndSequencesWithQuality], Phylogeny[Rooted], FeatureData[Taxonomy]), embedded provenance plus provenance replay, import (Casava/manifest/EMP/BIOM), export, the Metadata object, and the q2cli vs Artifact API interfaces. Covers why a .qza is data-plus-executable-history not a file, why export drops provenance, why a .qzv is terminal, why cl...

Assigns taxonomy to amplicon ASVs/OTUs (16S, ITS, 18S) with a classifier conditioned on a reference database and primer region - DADA2 assignTaxonomy + addSpecies (RDP naive Bayes), DECIPHER IDTAXA, and QIIME2 q2-feature-classifier (classify-sklearn naive Bayes, classify-consensus-vsearch alignment-consensus). Covers region-specific training (extract-reads, fit-classifier-naive-bayes), why a full-length classifier fabricates calls on a V4 read, the scikit-learn version-pinning trap on pre-tra...

Harmonizes already-normalized per-omic matrices onto a common footing before joint integration - assembling a MultiAssayExperiment, choosing the per-omic variance-stabilizing transform, deciding per-view versus per-feature scaling, picking a cross-omic batch strategy, and triaging missing data (feature, value, or whole sample; MAR versus MNAR). Covers why a shared-latent integrator is blind to what an omic is so scaling silently decides which block dominates, why batch confounded with biology...

Chooses a bulk multi-omics integration strategy before any tool runs by mapping the biological question (subtype discovery, shared axis of variation, predictive signature, pairwise correlation) to a method class, naming the sample correspondence (paired-vertical, horizontal, mosaic, diagonal), enforcing the n<<p discipline that makes a held-out cohort the endpoint instead of in-cohort cross-validation, and running the per-view variance-imbalance diagnostic. Covers the early/mixed/intermediate...

Builds supervised and unsupervised multivariate integration across bulk omics blocks with mixOmics - sPLS for sparse pairwise correlation, DIABLO (block.splsda) for a multi-block discriminant signature, rCCA for regularized canonical correlation, and MINT for multi-study integration. Covers why these projection methods maximize covariance or correlation and not truth, why DIABLO's design matrix is the central correlation-versus-discrimination decision, why cross-validation must wrap keepX sel...

Discovers shared and view-specific latent factors across bulk multi-omics blocks (RNA-seq, proteomics, methylation) on a common sample axis with MOFA2's unsupervised Bayesian group factor model, then attributes per-view variance explained and interprets signed factor weights. Covers why a factor is an unsupervised axis of variance and not a pathway, why a factor that correlates with batch is a batch factor, why the per-view variance-explained table is the primary read-out rather than p-values...

Stratifies patients into multi-omics subtypes by building one patient-by-patient similarity network per omic, fusing them with SNF's cross-network diffusion, and spectral-clustering the fused graph - then defending the clusters with stability, survival separation, and replication. Covers why spectral clustering always returns the requested cluster count so a subtype is a claim not a discovery, why the eigengap is a graph property not a biological truth, why fusion is not automatically better ...

Turns an enrichResult or gseaResult from clusterProfiler/enrichplot into a figure that collapses or shows gene-set redundancy, using dotplot, barplot, cnetplot, emapplot, treeplot, ridgeplot, gseaplot2, and upsetplot. Covers why a default top-20 GO dotplot is one biological theme drawn twenty times (the DAG/nesting guarantees redundant overlapping terms), so the figure is a modeling choice between SHOWING redundancy (pairwise_termsim -> emapplot/treeplot) and DELETING it (simplify/REVIGO); wh...

Runs Gene Ontology over-representation analysis (ORA) on a gene LIST with clusterProfiler enrichGO, the one-sided hypergeometric/Fisher 2x2 test phyper(k-1, M, N-M, n, lower.tail=FALSE). Covers why the BACKGROUND universe (not the gene list) is the null and decides significance, why omitting universe= is a bug, why enrichGO defaults to ont='MF' not 'BP', why pvalueCutoff filters p.adjust not raw p, why ORA discards effect magnitude and inherits GO-DAG true-path redundancy (simplify, topGO), w...

Tests a ranked gene vector for coordinated expression shifts in GO, KEGG, Reactome, or MSigDB gene sets with clusterProfiler's gseGO, gseKEGG, gsePathway, and GSEA (fgseaMultilevel engine), and scores per-sample pathway activity with ssGSEA and GSVA. Covers why a GSEA result is a deterministic function of three implicit choices (the ranking STATISTIC, the weight exponent p, and which LABELS are permuted), why the input must be a NAMED vector sorted DECREASING by a signed variance-calibrated m...

Tests gene lists, ranked vectors, and fold-change vectors against KEGG pathways and modules with clusterProfiler enrichKEGG/enrichMKEGG (ORA), gseKEGG (GSEA), and SPIA/graphite (signed-topology perturbation) in R. Owns the third pathway-analysis generation because KEGG ships signed directed signaling topology (KGML). Covers why a KEGG result is a timestamped join against a live REST API (irreproducible unless pinned with a gson snapshot, not the stale 2012 KEGG.db), why enrichKEGG keyType is ...

Tests a gene list or ranked gene vector for over-representation or coordinated shifts in Reactome's curated, peer-reviewed, reaction-level pathways using ReactomePA's enrichPathway (ORA) and gsePathway (GSEA), reading the local reactome.db so a run is reproducible given the Bioconductor release. Covers why Reactome's atomic unit is the REACTION and pathways are nested containers so a parent and child enrich on the same genes and double-count one signal, why only human is curated and every oth...

Tests a gene list (ORA, enrichWP) or a ranked gene vector (GSEA, gseWP) against the WikiPathways community-curated pathway collection with clusterProfiler and rWikiPathways. Covers why a WikiPathways result is a snapshot of a live, monthly-updated database (enrichWP/gseWP/gson_WP silently pull data.wikipathways.org/current/), why reproducibility requires pinning a dated GMT via downloadPathwayArchive(date=, format='gmt'), why the WP GMT is Entrez-keyed so symbols and Ensembl silently overlap ...

Imputes untyped genotypes against a phased reference panel with Beagle, Minimac4, or IMPUTE5 (array data) or from genotype likelihoods with GLIMPSE2, QUILT2, or STITCH (low-coverage WGS), producing per-variant dosages (DS) with a self-estimated quality (Beagle DR2, Minimac R2, IMPUTE INFO). Covers why the honest output is a dosage posterior not a hard call, why GWAS regresses on DS, why the quality metric is an ESTIMATE of r2 from posterior spread (not validation against truth), the DS/GP/HDS...

Estimates haplotype phase from population linkage disequilibrium with SHAPEIT5, SHAPEIT4, Eagle2, or Beagle - turning unphased genotypes (0/1) into phased haplotypes (0|1) for imputation input, compound-heterozygote calls, HLA typing, or population genetics. Covers why statistical phase is an INFERENCE (not a measurement) whose error concentrates at rare variants, why a genome-wide switch-error rate hides catastrophic rare-variant error and must be reported MAC-stratified, the SHAPEIT5 common...

Assesses and filters phasing/imputation output - the quality metrics (Beagle DR2, Minimac R2 and EmpRsq, IMPUTE/GLIMPSE INFO), MAF-stratified filtering, true accuracy by masking, the differential-imputation confound, dosage-based downstream usage, and phasing switch-error QC. Covers why every routine quality score is an ESTIMATE of r2 from the posterior spread (not validation against truth), why it is confounded with MAF so a flat INFO>=0.3 cutoff is a hidden rare-variant filter, why concorda...

Selects and prepares the reference panel that phasing/imputation copies haplotypes from (1000 Genomes, HRC, TOPMed, HGDP+1kGP/gnomAD, CAAPA), matching panel ancestry to the target, reconciling genome build and chromosome naming, and running the strand/allele harmonization gate. Covers why ancestry-match beats panel size (imputation can only copy haplotypes the panel contains), why palindromic A/T and C/G SNPs flip strand without erroring, why liftover is a strand-flip generator in between-bui...

Frames Bayesian phylogenetics as approximating a posterior distribution over trees conditioned on data AND priors via an MCMC that must be proven to have converged, using MrBayes, BEAST2, RevBayes, and PhyloBayes-MPI. Covers why convergence (ESS, PSRF, ASDSF, topology vs scalar) is the load-bearing claim, why posterior probabilities are systematically higher than bootstrap and overconfident under model misspecification, why the default branch-length prior inflates tree length, why the harmoni...

Build model-corrected evolutionary distance matrices and distance trees (NJ, BIONJ, FastME, UPGMA) with Biopython Bio.Phylo plus R ape/phangorn/FastME. Covers why a distance is a model-corrected estimate of substitutions per site that undercounts raw because of multiple/back/parallel hits (saturation); why the matrix discards the per-site information ML keeps; the LogDet/paralinear fix for compositional heterogeneity; the UPGMA molecular-clock trap; and the Bio.Phylo landmine that DistanceCal...