The PRISM Assay

PRISM is a cutting-edge technology for multiplexed cancer cell line viability screening to evaluate anti-cancer molecules or other perturbagens in hundreds of cancer models simultaneously. A unique DNA barcode is stably integrated into each cell line. Cell lines are then pooled, verified for cell line and barcode identity, and confirmed free from mycoplasma. The pools are then plated and treated with compounds for five days. Cell lines are lysed and mRNA is isolated and amplified and detected using the Luminex detection system.

Detection of barcode abundance in lysates allows for the generation of relative cell line sensitivity signatures by comparing treatment conditions to control conditions. PRISM data is processed through a series of quality control (QC) checks to evaluate the technical performance of signal detected in each well and the performance of control compounds.

The PRISM sensitivity profiles are used in conjunction with genomic information to generate predictive models to identify features that correlate with sensitivity. The data incorporated are from the Broad Institute’s Cancer Dependency Map genetic screens (CRISPR, shRNA), published PRISM compound screens (drug repurposing), and published Cancer Cell Line Encyclopedia (CCLE) data (cancer type and subtype, mutations, mRNA/protein expression, copy number, methylation).

The current list of datasets used for each of the biomarker analyses can be found on our GitHub repository here. They are available for download in the DepMap portal.

PRISM technology is compatible with screening a range of compound library sizes, from a single high-value compound to chemical libraries of tens of thousands of novel compounds in a short period of time.

Generation of Barcoded Models for PRISM Screens

Over 930 genomically characterized CCLE cell lines have been barcoded with a DNA barcode. All cell lines are tested for mycoplasma, verified with SNP fingerprinting, and the barcode identity is confirmed. Cell lines are then mixed together in assay ready pools according to doubling time.

PRISM Viability Assay

Pools of cells are treated for 5 days with compounds, then cells are lysed and mRNA is isolated. The barcode sequences are then amplified by PCR and detected by a Luminex scanner. The quantity of each barcode remaining after treatment serves as a readout to generate cell line sensitivity signatures for each compound.

Predictive Modeling

Sensitivity signatures from PRISM data are run through predictive modeling algorithms, such as random forest in order to identify biomarkers using CCLE genomic characterization data, Repurposing drug viability data, and Dependency Map loss-of-function genetic perturbation data.

Example Data Quality

PRISM provides QC data for each cell line in the screen. PRISM quantifies the separation between negative control (vehicle-only treatment) and positive control (a high dose of a pan-cytotoxic compound, generally bortezomib) measurements, using metrics such as error rate and Strictly Standardized Mean Difference (SSMD), and filters out cell lines that do not pass our standard from the final dataset. We also filter out cell lines that do not show significant growth (at least one doubling) from a day 0 timepoint to the endpoint.

This dataset’s quality is consistent with our typical screen. The majority of cell lines and replicates pass our quality standards (error rate < 0.05), with a few cell lines performing poorly in the screen.

PRISM Viability Assay

PRISM includes validation compounds in each screen with known MOAs. These compounds are used to ensure that, for known compounds, expected sensitivities and biomarkers are observed in each screening run. Here we show PRISM data for two compounds we use frequently as controls: AZ628 (a BRAF inhibitor) and Imatinib (an inhibitor of PDGFRA and the BCR-ABL1 fusion protein).

Melanoma cell lines harboring the hotspot BRAFV600E mutation are more sensitive to AZ628 than BRAF wild-type cell lines. CML cell lines, which harbor BCR-ABL1 fusions, are selectively killed by Imatinib, as expected.

Predictive Modeling

Predictive models are run on available feature sets including CCLE genomic characterization data such as gene expression, mutation and copy number, as well as loss-of-function genetic perturbation data from the Dependency Map. In addition to these datasets, viability data from the PRISM drug repurposing project is used as a feature set for univariate analysis.

PRISM identifies the expected biomarkers for validation compounds. Shown here: BRAF is identified as a top dependency from the Dependency Map CRISPR dataset, correlating with the AZ628 sensitivity signature. PDGFRA is identified as a top correlated gene expression feature for Imatinib.