3.7 Barcode
Counting
1. Sequence the cleaned-up pooled sequencing library (see step 3
under Subheading 3.6) using an NGS platform such as Illumina MiSeq or HiSeq 2500.
2. Several software packages exist, such as FASTX Toolkit (http://
hannonlab.cshl.edu/fastx_toolkit), HTSeq [14], and Bartender [15], to count combinations of strain, plate quadrant,
and well barcodes from FASTQ files or raw sequencing data.
Using Python, R, or other scripting language, these barcode
counts can then be merged with a predefined conversion table
or dictionary in order to map the counts to strains, hypomorphic gene products, wells, and compounds.
3.8 Fitness Inference
The custom R package ConcensusGLM (https://github.com/bro
adinstitute/concensusGLM) enables experimental batch correction and fitness inference from barcode count data. The script
exec/concensusGLM.R contained in the package allows for common use cases, and is run from the command line with the following options:
--data Input count table CSV, one line per lane
per strain per well
--meta CSV of handling records
--neg Name of negative control compound,
e.g., DMSO or none or untreated
--pos Name of positive control compound,
e.g., rifampin or BRD-K01507359-001-19-5 or BRD-K01507359
--checkpoint Save checkpoints to allow restart in case of
failure
--parallel Analyze strains in parallel (needs multiple cores)
--sge Use Grid Engine cluster with template
to analyze strains in parallel
--no-count-threshold Do not discard plates or strains based on
low counts
3.9 Mechanism
of Action
Interpretation
Where compounds with well-characterized targets exist, their
CGIPs can be used to annotate the mechanisms of action of compounds whose targets are unknown using supervised machine
learning. Lasso classification is a fast, interpretable supervised
learning method with reasonable predictive performance on new
data. Conveniently, the Lasso classification model is implemented
in the R package glmnet [16]. An outline to training and deploying a Lasso model in R follows, although more elaborate machine
learning methods could be used if desired.
1. Read fitness data into an R dataframe and add a column for
known mechanism of action labels.
338
Eachan O. Johnson and Deborah T. Hung
Counting
1. Sequence the cleaned-up pooled sequencing library (see step 3
under Subheading 3.6) using an NGS platform such as Illumina MiSeq or HiSeq 2500.
2. Several software packages exist, such as FASTX Toolkit (http://
hannonlab.cshl.edu/fastx_toolkit), HTSeq [14], and Bartender [15], to count combinations of strain, plate quadrant,
and well barcodes from FASTQ files or raw sequencing data.
Using Python, R, or other scripting language, these barcode
counts can then be merged with a predefined conversion table
or dictionary in order to map the counts to strains, hypomorphic gene products, wells, and compounds.
3.8 Fitness Inference
The custom R package ConcensusGLM (https://github.com/bro
adinstitute/concensusGLM) enables experimental batch correction and fitness inference from barcode count data. The script
exec/concensusGLM.R contained in the package allows for common use cases, and is run from the command line with the following options:
--data
per strain per well
--meta
--neg
e.g., DMSO or none or untreated
--pos
e.g., rifampin or BRD-K01507359-001-19-5 or BRD-K01507359
--checkpoint Save checkpoints to allow restart in case of
failure
--parallel Analyze strains in parallel (needs multiple cores)
--sge
to analyze strains in parallel
--no-count-threshold Do not discard plates or strains based on
low counts
3.9 Mechanism
of Action
Interpretation
Where compounds with well-characterized targets exist, their
CGIPs can be used to annotate the mechanisms of action of compounds whose targets are unknown using supervised machine
learning. Lasso classification is a fast, interpretable supervised
learning method with reasonable predictive performance on new
data. Conveniently, the Lasso classification model is implemented
in the R package glmnet [16]. An outline to training and deploying a Lasso model in R follows, although more elaborate machine
learning methods could be used if desired.
1. Read fitness data into an R dataframe and add a column for
known mechanism of action labels.
338
Eachan O. Johnson and Deborah T. Hung