Explaining Drug-Drug Interactions with LINCS L1000 Signatures

Abstract

Drug-drug interactions (DDIs) are a major cause of adverse drug events, but pharmacovigilance signals rarely say which biology links two drugs to a shared harm. This article describes Use Case 1 ("Y1") of the NIH Common Fund Data Ecosystem (CFDE) Playbook Partnership, a workflow that proposes mechanistic explanations for DDI-associated adverse events using only public gene-set data. For a triple of Drug 1, Drug 2 and an adverse event, the workflow builds gene sets for the event from human and mouse phenotype ontologies (HPO, MGI-MP), literature co-mentions (Geneshot) and the GWAS Catalog, and gene sets for each drug from the up- and down-regulated genes of its LINCS L1000 consensus signature. The seven sets are compared in a SuperVenn diagram, and the most surprising intersections are passed to enrichment analysis to yield candidate pathways. We place the approach against a text-mining study in which only 42.7% of DDIs with clinical pharmacodynamic evidence alone could be predicted from known drug-gene interactions, and discuss its limits: single-drug signatures, cell-line context, noisy gene sets and uncertain input signals.

Why explain a drug-drug interaction?

The Y1 use case proposes a way to explain why two drugs taken together cause a side effect: compare the genes each drug moves with the genes already tied to that side effect, and look for surprising overlaps. It is Use Case 1 of the CFDE Playbook Partnership and is published as a reusable workflow in the Playbook Workflow Builder.¹

DDIs are a major cause of adverse drug events, and their incidence rises with polypharmacy.² Pharmacovigilance can flag that a pair of drugs is associated with an event. It rarely says which biology connects them. Y1 tries to fill that gap with public gene-set data, so a researcher gets a short list of candidate genes and pathways instead of a bare statistical signal.

Where the interactions come from

Every run starts from one triple: Drug 1, Drug 2, and the adverse event they are reported to cause together. The use case names three places to get those triples: the literature, the FDA Adverse Event Reporting System (FAERS), and SIDER, a database of drug side effects drawn from labels.

For literature-derived input, the use case points to the supporting tables of Translational Knowledge Discovery Between Drug Interactions and Pharmacogenetics (PMID 31863452).^2,3 That study text-mined about 25 million MEDLINE abstracts and sorted DDI evidence into three kinds:

The gap that motivates Y1 sits in the last row. Of 13,012 pairs with only clinical PD evidence, just 42.7% could be predicted from known drug-gene interactions, versus 94.8% of pairs backed by all three evidence types.² Those PD-only pairs, an observed harm without a known enzyme story, are the natural candidates for a gene-expression explanation.

The workflow at a glance

1. Inputs: an adverse event, Drug 1 and Drug 2.
2. Seven gene sets: three for the event (HPO + MGI-MP phenotypes, Geneshot literature genes, GWAS Catalog genes) and two for each drug (L1000 up genes, L1000 down genes).
3. Overlap: all seven sets are compared in one SuperVenn diagram, and intersections are scored for how surprising they are.
4. Mechanism: the most surprising shared genes go to enrichment analysis, producing a candidate mechanism to test.

Branch one: genes behind the adverse event

The adverse event is turned into gene sets from three independent lines of evidence:

• Human and mouse phenotypes. The Human Phenotype Ontology (HPO) links clinical features to the genes that cause them in people; MGI Mammalian Phenotypes (MGI-MP) does the same for knockout and mutant mice.
• Literature gene mentions. Geneshot ranks genes by how often they co-appear with a search term in PubMed, catching associations not yet curated into an ontology.
• GWAS gene associations. The GWAS Catalog contributes genes near variants statistically linked to the trait in human populations.

In the published workflow, each source is searched through Enrichr for gene-set labels that contain the phenotype term, and all matching sets from one library are merged into a single union gene set. The libraries used are GWAS Catalog 2019, MGI Mammalian Phenotype Level 4 2019 and Human Phenotype Ontology.¹ The Geneshot branch of the original design does not appear in that published version.

Branches two and three: what each drug does to cells

Each drug is described by its transcriptional footprint from LINCS L1000. L1000 is a low-cost assay that measures about 1,000 landmark genes and infers the rest, across many compounds, doses, time points and cell lines. A consensus signature collapses those many profiles of one drug into the genes it most consistently pushes up and the genes it most consistently pushes down.

The workflow keeps the two directions apart, so each drug contributes two gene sets: an up set and a down set. They come from Enrichr's LINCS L1000 Chem Pert Consensus Sigs library, the same signature collection served by SigCom LINCS.¹ Keeping direction matters: two drugs that both suppress the same gene tell a different story from two that pull it in opposite directions.

Bringing it together: SuperVenn and surprising overlaps

All seven gene sets meet in a single SuperVenn diagram, and the interesting result is an overlap that should not be there by chance.

A classic Venn diagram becomes unreadable past four sets. SuperVenn instead draws each set as a row and each block of shared genes as a column, so every intersection among seven sets stays visible and countable. In the published workflow, the phenotype unions and the four drug signatures are assembled into one gene-set file and passed to this view.¹

The use case notes that unexpected entries can be scored by enrichment results and overlap levels. In practice the most informative columns are:

• genes moved in the same direction by both drugs that are also phenotype genes, which suggests an additive or synergistic effect;
• genes moved by only one drug but central to the phenotype, which points to that drug as the driver;
• genes moved in opposite directions, which can flag an antagonistic interaction.

Overlap size alone is misleading, because large sets overlap by chance. Ranking intersections by how much larger they are than expected, given set sizes and the genome background, is what separates a surprising overlap from a big one.

Enrichment analysis: from genes to mechanism

The final step turns the most surprising overlap into a hypothesis. The genes shared by, say, both drugs' up sets and the phenotype set are submitted to enrichment analysis against pathway, ontology and cell-type libraries. A shared gene list that lands squarely in one pathway gives a concrete, testable mechanism.

Take an illustrative (hypothetical) case: two drugs reported together with muscle injury. Suppose both drugs' down sets share genes with the HPO and MGI myopathy sets, and those genes are enriched for mitochondrial function. The hypothesis would be that the two drugs jointly impair muscle mitochondria, independent of any metabolic-enzyme interaction. That hypothesis could then be tested in cells.

This complements the pharmacogenetic route in the supporting paper. There, most explained DDIs ran through CYP enzymes: CYP3A was involved in 87.8% of the 986 pairs with all three evidence types, and CYP2D6 in 48.8%.² Y1 looks for the pharmacodynamic explanations that an enzyme-centric view misses.

Caveats and outlook

Y1 generates hypotheses; it does not prove them. Several limits shape how its output should be read:

• Each drug is profiled alone. L1000 consensus signatures come from single-compound treatments, so the combination's own effect is inferred from two separate footprints, not measured.
• Cell lines are not patients. Most L1000 profiles come from cancer and other cultured lines, which may not express the target tissue's biology.
• Gene sets are noisy. Term matching can pull in loosely related sets, and union sets grow large, which inflates chance overlaps unless they are scored against a background.
• The input signal can be wrong. FAERS reports are spontaneous and confounded, and text-mined DDIs carry extraction error; the supporting paper needed manual review to confirm its associations.²

The appeal is reach. Every step uses public data and runs without code in the Playbook Workflow Builder, which the NIH Common Fund presents as a way for researchers without programming experience to chain datasets into reproducible, citable workflows.⁴ Run across the thousands of PD-only interactions in the literature, it could prioritize which ones deserve a laboratory experiment.