How LISA dictionaries were built

The built-in LISA dictionaries were prepared before any user’s study was analysed. During construction, model-specific annotations rated how closely a gene set fitted a predefined biological category. lisaR now distributes the accepted gene-set-to-category assignments and applies them locally. An ordinary analysis does not ask a language model to classify the study.

This guide explains the recorded construction method and its limits. It does not describe a new analysis or a method for interpreting one study’s enrichment scores.

What the four collections mean

Collection	Biological scope during construction
`GOBP-C2`	Broad biological programmes. It combines GO Biological Process gene sets with selected curated pathway families from MSigDB C2:CP.
`GOMF`	Molecular activities represented by GO Molecular Function gene sets.
`GOCC`	Cellular locations, structures and complexes represented by GO Cellular Component gene sets.
`PATHWAYS`	A stricter signalling classification of the GOBP-C2 gene-set space.

HALLMARKS is separate. lisaR analyses Hallmark gene sets directly rather than assigning them through a LISA semantic dictionary.

The biological categories and their definitions existed before scoring. The inspected GOMF and GOCC annotation inputs contain the GO source identifier, gene-set name, example genes and category definitions. Their description fields are empty. The records therefore do not support a claim that every model received a full gene membership list or a prose description for every gene set.

Four recorded model streams

The saved construction table has separate score and source columns for Qwen3.5, Gemma4, Mistral3.2 and Phi4-14b. The selection code treats Qwen3.5 and Gemma4 as the anchor pair. Mistral3.2 and Phi4-14b provide supporting scores. This is not unrestricted majority voting.

The scores are ordinal judgements of biological fit. They are not probabilities, confidence percentages, votes, GSEA weights or measurements from a study.

The surviving records give the clearest collection-specific interpretation for GOMF and GOCC:

For GOMF, 0 means no relationship, 1 a weak or contextual relationship, 2 a supported secondary function, 3 a major function and 4 a primary or canonical function.
The GOCC production rubric allows 0, 3 or 4 and asks whether the category describes compartment identity. It does not score molecular function or pathway activity.

The records support a common selection rule for GOBP-C2, GOMF and GOCC, but they do not establish one identical production prompt for all three collections. Prompt optimisation records are historical calibration material, not proof of the final prompt used for every saved score.

Exact method for GOBP-C2, GOMF and GOCC

The ordinary GO-family rule was named anchor_support_consensus_v0_1. After any explicitly recorded imputation, the four scores were combined as follows:

A final score of 4 was assigned when both anchors were 4. It was also assigned when both anchors were at least 3, at least one anchor was 4, and at least one support model scored at least 3.
Otherwise, a final score of 3 was assigned when both anchors were at least 3. It was also assigned when one anchor was at least 3, the other was at least 2, and at least one support model scored at least 3.

Core retained final score 4. Expanded retained core and final score 3. OTHER_UNCLASSIFIED, MF_UNKNOWN and CC_UNKNOWN_OTHER were excluded. Lower-score outputs remained in construction and sensitivity tables rather than the default distributed dictionaries.

Missing GO-family scores

A derived GOBP-C2 layer filled a missing model score with the mode of the available model scores for that category. When modes tied, it chose the lower value. The saved consensus contains 3,649 gene-set and category pairs with an imputed value. Source fields distinguish these derived values from raw scores. The inspected builder applies this imputation layer to GOBP-C2, not to GOMF or GOCC.

An imputed value is not another independent model annotation.

Eleven recorded GOMF overrides

Eleven GOMF pairs have an explicit manual validation score recorded under David Olmeda’s review. That score replaced the ordinary rule result before tier selection.

Gene-set ID	Category ID	Rule score	Recorded override
`GOMF_ANAPHASE_PROMOTING_COMPLEX_BINDING`	`MF_CELL_CYCLE_REG`	3	4
`GOMF_ATP_DEPENDENT_PROTEIN_FOLDING_CHAPERONE`	`MF_STRESS_RESPONSE`	2	3
`GOMF_CHEMOREPELLENT_ACTIVITY`	`MF_DEVELOPMENT_REG`	3	3
`GOMF_CLATHRIN_ADAPTOR_ACTIVITY`	`MF_VESICLE_TRAFFIC`	3	4
`GOMF_CYCLIN_BINDING`	`MF_CELL_CYCLE_REG`	3	4
`GOMF_CYSTEINE_TYPE_ENDOPEPTIDASE_REGULATOR_ACTIVITY_INVOLVED_IN_APOPTOTIC_PROCESS`	`MF_APOPTOSIS_REG`	3	4
`GOMF_DOUBLE_STRANDED_RNA_BINDING`	`MF_VIRAL_HOST`	2	3
`GOMF_EXOGENOUS_PROTEIN_BINDING`	`MF_VIRAL_HOST`	2	3
`GOMF_HEAT_SHOCK_PROTEIN_BINDING`	`MF_STRESS_RESPONSE`	3	4
`GOMF_NOTCH_BINDING`	`MF_DEVELOPMENT_REG`	3	3
`GOMF_PEPTIDASE_ACTIVATOR_ACTIVITY_INVOLVED_IN_APOPTOTIC_PROCESS`	`MF_APOPTOSIS_REG`	3	4

An override that retains the same numerical score is still part of the recorded manual validation layer.

Exact method for the current PATHWAYS assignments

PATHWAYS was selected later under the historical v0.7 name-first method dated 3 July 2026. A pathway’s name had to support the biological assignment. Canonical genes could support a name-supported call, but gene overlap alone was not enough for a score of 3 or 4. Models could assign zero to every PATHWAYS category.

The current PATHWAYS core contains 921 pairs:

824 have Qwen=4 and Gemma=4.
97 have anchor scores 4 and 3 in either order, with both Mistral and Phi observed at 4.

The expanded PATHWAYS dictionary contains those core pairs and 134 additional pairs:

100 have anchor scores 4 and 3, both support scores at least 3, and exactly one support score equal to 4.
34 have both anchors at 3 and both support scores observed at 4.

A score of 4 from one anchor with the other below 3 was not admitted by these core or strict-expanded rules. A further 152 review-only rows are outside the production ledgers. The broad PATHWAYS sensitivity set was identical to the expanded set, not an additional permissive tier.

The distinction between observed and imputed support matters. The core 4/3 branch and the expanded 3/3 branch explicitly require observed support. The expanded 4/3 branch does not carry that same explicit source condition, so it must not be described as if it did.

Missing PATHWAYS scores

The PATHWAYS build records eight missing Phi gene-set responses, covering 576 category score cells. A derived layer filled those cells with the median of the available model scores. This is different from the lower-tied mode used for missing GOBP-C2 scores. The audit verified the source code and saved report for this step but did not replay every raw PATHWAYS record.

How historical selections reach current v1

The located source ledgers contain all assignments in the current package tables:

Tier	GO-family rows	PATHWAYS rows	Current total
Core	14,592	921	15,513
Expanded	22,171	1,055	23,226

The source-to-package comparison found the same seven-field rows and values, with no unmatched or duplicate projected rows. It was a multiset comparison, not a byte-for-byte replay of the historical build order.

The current first-publication resource identities are lisa_dictionary_core@1.0.0 and lisa_dictionary_expanded@1.0.0. The historical construction names v0_1 and v0_7 describe archived stages. They are not the version numbers of the current distributed resources. Older package resource pins also remain distinct historical identities.

Current runtime tables retain universe, gene-set identity and label, category identity and label, source identity, and tier. They do not retain LISA_score. The removal of that historical construction field did not change the accepted assignments or tiers.

Calibration evidence and limits

The manual calibration record describes 123 validated rows. Three unclassified rows were excluded, leaving 120. The saved summary reports 120/120 recovered by the expanded rule and 118/120 by the core rule. This was a development and calibration recovery check. It was not an independent held-out study, a whole-genome accuracy estimate or validation for every possible biological classification.

The available records do not support a complete historical replay. In particular:

The original MSigDB retrieval release was not recorded.
Complete production prompts, responses and successful execution receipts were not recovered for every GO model and gene set.
One aggregate raw GO score file named by the builder is absent at its recorded location, although the downstream consensus table and imputation layer are present.
Two exact historical hierarchy inputs named by the technical rebuild contract were not identified through the inspected construction chain.

These limits concern reconstruction of the historical annotation process. They do not make an LLM necessary for an ordinary lisaR analysis.

Sources available with lisaR

The installed package includes its current resource manifest, migration ledger and technical reconstruction notes. Locate them without relying on a private project path:

system.file(
  "extdata", "dictionaries", "RESOURCE_PROVENANCE.tsv",
  package = "lisaR"
)
system.file(
  "extdata", "dictionaries", "TECHNICAL_REBUILD_PROVENANCE.md",
  package = "lisaR"
)
system.file(
  "extdata", "dictionaries", "RESOURCE_MIGRATION_MANIFEST.tsv",
  package = "lisaR"
)

The detailed method above was checked against the archived records named build_lisa_dictionaries.py, lisa_consensus_scores_all_universes_v0_1.tsv, consensus_rule_scores_long.tsv, gobpc2_missing_model_scores_imputed_mode_long.tsv, derive_pathways_v0_7_core_extended_rules.py, pathways_v0_7_final_core_inclusions.tsv and pathways_v0_7_final_core_plus_extended_strict_inclusions.tsv. Those archive files are not currently distributed with lisaR. Their absence is why this page states the limits explicitly rather than claiming complete replay.

For ordinary resource selection, see Resources and provenance. To build and register your own classification, continue to Custom dictionaries and supercategories.