Perturbation modeling 
¶
Cross-study drug-perturbation transcriptomics on Tahoe-100M + LINCS Level 2 (and appended DRUG-seq): harmonize compounds and gene symbols onto a shared pert_compound label and gene panel, train a modlyn linear classifier on a MappedCollection, then enrich and interpret top genes per perturbation.
Reusable logic lives in the perturbation_modeling library. Thin tracked scripts on the instance import it (build_collection, append_dataset, train_feature_selection, enrich_top_genes, build_evidence_report) instead of copying pipeline code.
A key pattern: append a new study to an existing collection and retrain. After Tahoe + LINCS were harmonized, DRUG-seq was aligned to the same pert_compound label and gene-symbol panel (including when var_names are Ensembl IDs), versioned into the collection, and the classifier was retrained on the expanded MappedCollection — without rebuilding the upstream library.
Instance: laminlabs/perturbation-modeling · Repo: laminlabs/perturbation-modeling
Harmonize & append¶
Tahoe compounds are kept when they appear in any LINCS phase (not the intersection of every phase). Outputs use the demo/ key prefix.
Step |
Transform |
Outputs |
|---|---|---|
Build Tahoe + LINCS collection |
collection · compounds · Tahoe |
|
Append DRUG-seq |
Modeling & interpretation¶
Retrain on the updated collection after each append so new compounds and cells enter the feature-selection model. The interpretation script loads those tables, calls build_evidence_report for the evidence block, and puts compound notes in extra_sections.
Step |
Transform |
Outputs |
|---|---|---|
Train / retrain SimpleLogReg (modlyn) |
||
Gene-module enrichment |
||
Interpretation report (agent) |