Finding small-molecule mimickers of a single-gene CRISPR knockout in LINCS L1000: a protocol and SMAD3 worked-example design (Use Case 31)

Abstract

Pharmacological phenocopying asks whether a small molecule can reproduce the transcriptional effect of a genetic loss of function. The LINCS L1000 resource profiled thousands of CRISPR single-gene knockouts and tens of thousands of small molecules across many human cell lines, so it can in principle answer this question for any profiled gene. Here we describe a reproducible workflow (LINCS Use Case 31) that takes one human gene, checks whether L1000 holds knockout signatures for it, builds a consensus knockout signature and per-cell-line signatures, scores chemical perturbation signatures against them to find mimickers, and reports the results as tables, a bar plot, and a list of consensus compounds supported by both the consensus and multiple individual cell lines. We propose SMAD3 as the example gene, because Smad3-null mice show accelerated cutaneous wound healing, so mimickers of its knockout are hypotheses for wound-healing therapeutics. This version is a protocol only. The L1000 queries have not yet been run, and no results are reported. Results will be added in a later revision.

Introduction

A genetic knockout reveals what a gene does, but it is rarely a practical therapy. A small molecule that produces a similar transcriptional response is a candidate for achieving the same effect pharmacologically. The L1000 assay of the Library of Integrated Network-Based Cellular Signatures (LINCS) program has measured gene expression responses of 248 human cell lines to 33,621 small molecules and 7,508 single-gene CRISPR knockouts, according to the L2S2 resource paper.¹ Search engines built on these data, including L2S2 and SigCom LINCS, allow a knockout signature to be compared against chemical perturbation signatures, including a consensus search that ranks perturbations across cell lines, time points and concentrations.¹

SMAD3 is a useful example because its knockout phenotype is known and clinically interesting. Smad3-null mice show accelerated cutaneous wound healing, with a faster rate of re-epithelialization and reduced local monocyte infiltration, compared with wild-type mice.² Later work reported that loss of Smad3 also improved healing of wounds in previously irradiated skin, with enhanced epithelialization and fewer neutrophils and myofibroblasts.³ Small molecules that mimic a SMAD3 knockout signature are therefore candidates for follow-up as modulators of wound repair, subject to the caveats below.

Methods

Workflow overview

The workflow (Figure 1) has five stages.

Figure 1. Use Case 31 workflow.

Human gene
   |
   v
Check whether KO signatures are available in L1000
   |                                  |
   v                                  v
Get consensus signature      Get individual cell-line signatures
   \                                  /
    \                                /
     v                              v
              Get mimickers
        /          |            \
       v           v             v
  Results as   Results as    Identify
   tables       bar plot    consensus drugs

Stage 1: Input and availability check

The input is one human gene symbol (SMAD3 in the worked example). The first step is to determine whether LINCS L1000 contains CRISPR knockout signatures for the gene and, if so, in which cell lines. If it does not, the workflow stops and reports that outcome. A pre-specified fallback gene list can be used instead, and any substitution must be stated in the paper.

Stage 2: Knockout signatures

Two kinds of signature are retrieved. The consensus signature condenses the knockout signatures across cell lines, time points and concentrations into a single ranked up- and down-regulated gene signature. Individual cell-line signatures are retrieved separately for every cell line in which the knockout was profiled.

Stage 3: Mimicker search

Each signature is used to query the chemical perturbation signatures in L1000 for small molecules whose signatures are most similar to the knockout (mimickers). The search engine, release version, similarity metric, and any score or significance thresholds will be recorded exactly as used. The query is run once for the consensus signature and once for each cell-line signature.

Stage 4: Outputs

• Tables: compound, similarity score, cell line (or "consensus"), and rank.
• Bar plot: the top-ranked mimickers by score.
• Consensus drugs: compounds that rank as mimickers in the consensus search and also in more than one individual cell-line search. The minimum number of cell lines will be fixed before the analysis is run.

Data source and reporting

The only data source is LINCS L1000. The dataset version and access date will be recorded for every query. Every value in the Results section will be taken directly from the retrieved output, with no manual edits.

Results

Pending. The L1000 queries for SMAD3 have not been run in this version. In particular, it has not yet been confirmed that SMAD3 has CRISPR knockout signatures in L1000, which cell lines are covered, or which compounds are mimickers. A subsequent revision will report:

1. Availability of SMAD3 knockout signatures and the cell lines covered.
2. The consensus signature summary and per-cell-line signatures.
3. Mimicker tables and the bar plot.
4. The consensus drug list.

Discussion

What the eventual results can and cannot show

The wound-healing phenotype comes from knockout mice, not from L1000 cell-line data.² A compound that matches the SMAD3 knockout signature in cultured human cell lines is only a hypothesis. Several gaps lie between a transcriptional match and improved healing:

• L1000 cell lines may not reflect keratinocytes, monocytes, or fibroblasts in a wound.
• A CRISPR knockout is complete and permanent, while a drug acts partially, transiently and often on several targets.
• Similarity of signatures does not establish that a compound acts through the TGF-β/SMAD pathway.
• Cell-viability effects can drive signature similarity independently of the target, so toxic compounds may appear as spurious mimickers.

Intended interpretation

Once results are available, the top mimickers will be examined for pathway plausibility, for example compounds that affect TGF-β receptor signaling, and for consistency across cell lines. Any compound proposed for follow-up would need testing in wound-relevant cell types and in vivo models.

Limitations of this version

This protocol has no results. It does not demonstrate that the workflow runs end to end for SMAD3, and it does not demonstrate that SMAD3 is present in L1000.