CacaoIQ doesn't invent new statistics. It builds on peer-reviewed methods and widely used open-source engines, chooses sensible defaults, and documents every step in a methods manual that ships with the software.
No proprietary statistics licenses are required: every engine below is open source and bundled with CacaoIQ.
| Stage | Method | Engine |
|---|---|---|
| Phenotype cleaning & BLUPs | Linear and generalized mixed models with spatial and design terms; ordinal models for scored traits | lme4, ordinal |
| Genomic relationships | VanRaden genomic relationship matrix with SNP quality control | NumPy |
| Heritability | Bayesian animal model on genomic relationships, liability scale for binary traits | MCMCglmm |
| Association mapping | Mixed-model GWAS (EMMAX) with kinship correction, Bonferroni thresholds and genomic control | rrBLUP |
| Linkage mapping | Phase and recombination-fraction estimation for full-sib (CP) families | Lep-MAP3, OneMap |
| QTL mapping | Interval mapping for outcrossed full-sib families, LOD-support intervals, variance explained | fullsibQTL |
| Combining ability | Factorial GCA/SCA mixed model with confidence intervals and narrow/dominance ratios | lme4 |
| Diversity analysis | Relationship-matrix PCA, LD decay, neighbor-joining trees | NumPy, Biopython |
| Candidate genes | Gene models from the Matina 1-6 v1.1 cacao reference, with UniProt and InterPro functional annotations bundled for offline use | Bundled reference |
| Biology summaries | Local language model constrained to the annotations shown | llama.cpp |
A pipeline is only useful if it gets the right answer. CacaoIQ's analyses are checked against the results of published studies, using the studies' own data.
Using the public genotype and phenotype data from a Colombian cacao collection (Osorio-Guarín et al., 2020,G3), CacaoIQ's association analysis for witches' broom found genome-wide significant hits within about 45 kb of both chromosome 2 loci reported by the original authors, with well-controlled test statistics (genomic inflation factor 0.90).
On a published cacao factorial mating trial, CacaoIQ's open-source GCA/SCA model reproduced the original commercial mixed-model software's estimates for every parent and every family (correlation 1.00), with heritability and dominance ratios inside the original study's reported uncertainty.
Quality checks report duplicates, outliers and impossible values for you to review. Nothing is silently dropped or imputed.
Candidate genes are only reported around loci that clear the significance threshold, so a gene list never looks like a finding when it isn't.
Language models help explain results. They never produce the numbers, and every summary is grounded in annotations you can see.