Background & Summary

Multiple myeloma is a malignancy of plasma cells which are terminally differentiated B cells. It is the second most common hematological malignancy with a global incidence estimated at 176,4041 cases and a relative 5-year survival rate estimated at 61.1%2. Copy number alterations and complex genomic rearrangements involving translocations of immunoglobulin enhancers to growth promoting genes are key drivers of malignancy3. Despite significant advancements in treatment and a corresponding increase in average expected survival time over the past several decades4,5, this cancer remains incurable with current treatment strategies.

High-throughput small molecule screening assays provide an effective means to test a large set of candidate compounds against multiple cell lines6,7. To help facilitate the discovery of novel compounds with therapeutic potential in multiple myeloma, a large-scale small molecule compound library containing 1,912 compounds (the NCATS Mechanism Interrogation PlatE, “MIPE” 4.0) was screened against a set of 47 diverse human multiple myeloma cell lines (HMCL). This dataset was mined to identify potential drug combinations and perform preclinical assessments of their utility as potential therapies for multiple myeloma8. Briefly, robust regression analyses were performed with 178 drugs that produced a curve class of −1.1 and had an AC50 less than 2 μM in at least 25 cell lines. From the 178 drugs, 43 combinations were highly correlated in their responses across the myeloma cell lines with r2 values > 0.5. A secondary screen was performed to find 6 combinations that simultaneously increased the levels of the tumor suppressor CDKN2A (p16) and decreased levels of the MYC oncogene. Among these 6 combinations, three were found to be highly synergistic in their action against several myeloma cell lines. Common genetic pathways up- and down-regulated involved the Tgfβ/Smad signaling and cell cycle transition pathways, respectively. The drug screen dataset includes many additional drugs not considered in that study. The goal of this data descriptor is to provide a clear description of the full drug screen dataset9 including additional supporting information, while also adjusting for spatial biases subsequently detected in the drug screen.

Raw and processed forms of the dataset are provided, as well as supporting metadata and a summary of the salient features present at the drug, cell line, and drug plate levels. An emphasis is placed on transparency and reproducibility, with the entire pipeline used to process the data and generate the tables and figures in this manuscript provided in the form of a Snakemake workflow (Fig. 1).

Fig. 1
Fig. 1The alternative text for this image may have been generated using AI.
Full size image

Experimental workflow. High-level overview of data generation and processing pipeline.

Methods

Cell culture

Human MM cells were cultured and authenticated as described previously10. Briefly, cells were cultured in RPMI 1640 with GlutaMAX L-glutamine (Life Technologies) supplemented with 10% fetal bovine serum (Cambrex BioScience), 100 U/ml penicillin and 100 μg/ml streptomycin (Life Technologies), prior to assay. Cell lines were provided by the Keats lab; sources and reference publications can be found at https://www.keatslab.org/myeloma-cell-lines/hmcl-public-availability.

Small molecule screen

The National Center for Advancing Translational Sciences MIPE (Mechanism Interrogation PlatE) compound library of 1,912 small molecules (“drugs/metadata.tsv” in the Zenodo dataset) was screened in 47 MM cell lines from the same passage of cells that had been sequenced previously (Table 1)7. Cell lines were treated in 1,536-well plates (Fig. 2) and the response was measured after 48-hour drug exposure with CellTiter-Glo® (Promega) luminescent cell viability assay at 11 doses (a serial of 3-fold dilutions), each in single replicate; https://ncats.nih.gov/research/research-activities/compound-management/chemical-libraries. As previously described7, a Multidrop Combi dispenser (Thermo Fisher) was used to dispense a total of 10^3 cells/well in 5μl of media (described above) into 1536 well solid-bottom white Greiner Bio-one tissue culture plates. A Kalypsys pintool was used to dispense 23 nl of control compound (bortezomib) and library compounds to generate the dose response curves. The 1536 well plates were covered with Kalypsys stainless steel cell culture lids and incubated at 37°C with 5% CO2 and 95% humidity. Cells were incubated for 48 h, after which 3 μl of CellTiter-Glo® was added with a Bioraptor Flying Reagent Dispenser (Aurora Discovery-BD) and incubated for 15 min at room temperature. The CellTiter-Glo signal was measured using a luminescent filter and 10 s exposure on a ViewLux (Perkin Elmer).

Table 1 Cell line metadata.
Full size table
Fig. 2
Fig. 2The alternative text for this image may have been generated using AI.
Full size image

Plate map. Reference plate image depicting the locations of control wells and arrangement of treated wells for different drugs and concentrations. A raw L363 plate (T5612674) is used as an example.

Plates from 4 of the 47 cell lines were found to deviate significantly due to potential issues with control wells (KMS11 & KMS28BM) or edge effects (KMS21BM & Karpas417) and were filtered out prior to background plate construction. Raw data for the outlier cell lines are provided.

Normalization and background adjustment

Starting with the raw plate luminescence measurements, clipping was first applied within each plate, setting the small number of negative values present (<0.01%) to “0” and a maximum value equal to the 99.5% quantile. Normalized cell viability was computed as:

$${\rm{Viability}}\,( \% )=100\ast \,({\rm{well}}-{\rm{pos}})/({\rm{neg}}-{\rm{pos}})$$

Where “well” refers to the raw well luminescence values, and “pos” and “neg” refers to the median fluorescence measurements of the positive control wells (cells + bortezomib), and a negative control wells (cells + DMSO, dimethyl sulfoxide), respectively. Next, in order to adjust for spatial biases commonly encountered in high throughput screening assays11, a simple background adjustment approach was applied. A “background” plate matrix was constructed by taking the mean intensity for each well across all plates, and then subtracting the average within each specific drug concentration. The background plate was then subtracted from each raw plate matrix, and the resulting values were rescaled to the range [0, 100].

Dose response curve fitting

For each cell and drug combination tested, four-parameter log-logistic dose-response models were fit using the drc package for R12. In cases where drc failed to converge, likely due to atypical dose response curves, the corresponding result entries were set to “NA”.

Average cell line and drug behavior

For each cell line and each dosage increment, average viability was computed as the mean of the normalized viabilities across all 1,912 compounds tested. The resulting average response curves for each cell were combined into a single plot (Fig. 3a). Cells with higher average viability scores appear near the top, while those with lower average viability scores appear near the bottom of the plot. For each drug, median AC-50 values were computed, and the distribution of resulting AC-50 values visualized as a histogram (Fig. 3b, Table 2).

Fig. 3
Fig. 3The alternative text for this image may have been generated using AI.
Full size image

High-level view of differences in drug efficacy and cellular response to treatment. (a) For each cell line and dose increment, mean viability across all 1,912 compounds was computed and the resulting dose-response curves were plotted. (b) For each drug, median AC-50 values were computed and a histogram of the resulting averages was plotted.

Table 2 Summary of drug behavior.
Full size table

Computational pipeline

In order to facilitate transparency and reproducibility, a computational pipeline was created using the Snakemake workflow engine13, with code written primarily in the R programming language14. Figure 3 was generated by the pipeline using ggplot215 with colorblind friendly palettes from ggpubfigs.16. Data generated by the pipeline was packaged into a data package17 with relevant metadata for each element included.

The workflow code used for data processing and generation of supporting figures and tables can be found at https://github.com/khughitt/hmcl-drug-screen-pipeline.

Cell line mutations

Predicted mutations for 46/47 cell lines for which data are available is provided as a supporting table9. This is a modified version of the “HMCL69_Preliminary_Mutation_List” dataset available at https://www.keatslab.org/data-repository, with cell lines not present in the drug screen excluded and cell line identifiers harmonized. Briefly, the mutation calls originate from human whole exome data generated using the Agilent SureSelect V4 + UTR hybrid capture kit. Sequencing data was generated on an Illumina HiSeq 2000 using 83 × 83 paired-end sequencing. The data was aligned to GRCh37 using BWA followed by duplicate marking and base quality recalibration with GATK. Mutations were identified using SAMTOOLS and likely germline events were filtered out when seen in 1000 genomes, ExAC r0.3, or NHLBI esp6500si_v2_ssa137 and not seen in COSMIC v74. The consequences of the individual likely somatic mutations was determined with snpEFF.

These methods expand on descriptions from our previous work8.

Data Records

Raw and processed versions of the drug screen data described in this data descriptor are available at Zenodo9. Processed drug screen data is provided in several different forms, facilitating comparisons between drugs, cell lines, and plates.

In addition to the drug screen data itself, drug and cell line metadata tables are provided, as well as all tables and figures generated in the computational pipeline described above.

Data is provided in the form of a Frictionless data package (https://datapackage.org/), organized in subfolders relating to workflow steps. A top-level “datapackage.yml” file provides machine-readable descriptions of the main data elements including basic information such as file checksums, field types, and descriptions of each table.

The raw screening data, including measurements for all cell lines assayed, can be found at raw/data.tsv. Each row in the table describes a single well measurement, with fields corresponding to cell line (cell_line), experiment date (date), plate identifier (plate), a software parameter which can be used as a proxy for technical batch (layer_name), well row and column numbers (row and col), unadjusted well value (well_value), drug identifier (sample_id), and drug concentration in moles (concentration).

The file “drugs/drug_curves.tsv” contains the model fit parameters (slope, lower_limit, upper_limit, ac50, and lac50) and P-values (ac50_pval) corresponding to the fitted dose response curves for each cell line + drug pair, as well as the drug concentrations (conc_0..conc_10) and adjusted viability measurements (dose_0..dose_10) used as inputs for model fitting. The file “drugs/ac50.tsv” contains a matrix of AC-50 values, with cell lines arranged along rows, and drugs along columns. In cases where the curve fitting algorithm failed to converge, missing values may be present, represented as “NA”. Drug metadata can be found in the file “drugs/metadata.tsv” and is based off of a table of drug information provided by NCATS for the MIPE 4.0 screen, with updates based on metadata for the more recent MIPE 6.0 screen.

The file “manuscript/tables/Table 1.tsv” contains cell line metadata, and is the basis for Table 1 in the manuscript. The table is derived from a table of cell line information obtained from https://www.keatslab.org/myeloma-cell-lines/hmcl-characteristics, and has been modified to exclude cell lines not present in the drug screen and to harmonize identifiers. The folder “manuscript/figures/” contains the plots used to construct Fig. 3 in the manuscript, as output by the pipeline.

The “plates/” folder contains versions of the drug screen data at different stages of processing: “plates/raw.tsv”, “plates/raw_filtered.tsv”, “plates/normed.tsv”, and “plates/background adjusted.tsv”. In each case, data for each plate are stored in a single column, with the plate identifiers in the first row. The file “plates/concentrations.tsv” has the same structure, but contains the drug concentrations associated with each measurement. The file “plates/background.tsv” contains the background plate used to adjust for spatial biases, and mirrors the structure of a single 48 × 32 plate. Plate-level metadata including experiment date, cell line, and various quality assurance-related fields is provided at “plates/metadata.tsv”. The folder “plates/images/” contains image representations of each individual plate, as well as images for the “background” plate and average (“mean” and “median”) plates.

Predicted mutations for 46/47 cell lines for which data are available is provided as a supporting table at mutations/hmcl-predicted-mutations.tsv.

An earlier processed version of the drug screen dataset is also available on PubChem, with one record per cell line (e.g. https://pubchem.ncbi.nlm.nih.gov/bioassay/1918926). The data resource described here includes both the raw data and an alternative processing accounting for previously unappreciated spatial bias, as well as supporting datasets and workflow code.

Technical Validation

Drug screen validation and quality assurance

Each drug screen plate included four control columns which were used for quality assessment and normalization (Fig. 2). The first column contained only media (DMSO, dimethyl sulfoxide), the second and third columns contained cells treated with a proteasome inhibitor with known efficacy in myeloma cells (bortezomib), serving as a positive control, and the fourth column contained cells with DMSO, serving as a negative control. Each plate was assigned a heuristic quality score based on the median Pearson correlation of cell viability curves on the plate with a logistic curve representing an idealized sigmoid dose response curve, allowing potentially problematic plates to be flagged.

Plates from one experimental date and two cell lines (MM1S_ATCC and U266_ATCC) were found to have issues with the control wells and were flagged in the plate metadata table. Four additional cell lines (Karpas417, KMS11, KMS21BM, and KMS28BM) were found to have more widespread unexpected plate behavior, including visible edge effects, significantly lower correlations with other cell lines and lower average viabilities for nearly all doses. These were flagged as “outlier cell lines” in the plate metadata table. Plate images were constructed for each level of processing and manually inspected, helping to identify potentially problematic plates and cell lines as well as spatial biases described above and in the “background adjustment” methods section. These four cell lines were excluded from all downstream processing and figure generation steps.