The NCI-60 Human Tumor Cell Line Screen: A Catalyst for Progressive Evolution of Models for Discovery and Development of Cancer Drugs

A well-known aphorism—attributed to British statistician George Box—states, “All models are wrong, but some are useful.” Another—credited to French philosopher Voltaire—opines, “Perfect is the enemy of good.” These words of wisdom are valuable when selecting experimental models of human cancer, and especially applied to the tough challenge of discovering new cancer drugs. If asked the question, “What is the best cancer model?,” I reply, “It depends on the purpose of the research and question you're asking.” For example, whether it is fundamental research to discover a pathogenic mechanism, or alternatively various stages of translational research, the most suitable model—or more likely the best combination of models—will probably differ. No single model can possibly be perfect, and major factors to consider include the best representation of the human malignancy of interest balanced with the practicality of application (including time, cost, reproducibility, etc.). Here, we focus on the rationale, implementation, and legacy of the “NCI-60” human tumor cell line screen applied to small-molecule drug discovery.

In their landmark article published in Cancer Research in 1988, Alley and colleagues (1) provided the initial conceptual and technological basis for NCI-60 human tumor panel screening (2, 3). They demonstrated the feasibility of testing, at unprecedentedly high-throughput, vast numbers of synthetic compounds and natural products, evaluated at multiple concentrations against >60 cancer cell lines, aiming to identify small molecules exhibiting selective cytotoxicity against different cancer histologies. Cited 4,289 times at the time of this writing, the vision and impact of the paper and overall initiative cannot be overestimated.

For a modern audience to fully appreciate the significance of this landmark paper and NCI-60 project (2, 3), we must follow another French philosopher, Auguste Comte, and review the history to understand the profound difference between 1955 and 1985 and today—both in the drugs used to treat patients with cancer and also the models used to discover them (Fig. 1A). We must recall first that in those days clinical treatment was dominated by cytotoxic chemotherapy agents that act predominantly by killing rapidly dividing cells and exhibit only modest selectivity (therapeutic index) between malignant and healthy cells in sensitive normal tissues, like bone marrow and gut mucosa (4). And second, although it seems extraordinary to us now, we must recognize that far less drug discovery was done in industry then, and most of the extensive antitumor testing was performed by the NCI; moreover, this involved systematic primary triage screening of huge compound numbers in two transplantable in vivo syngeneic mouse leukemias, the L1210 and P388 models (5). During that period, there was no real appreciation that the anticancer activity of a potential new agent could be dependent on the cell lineage or histologic subtype of a given cancer, let alone be dictated by the specific genomic abnormalities driving each malignancy, which were mainly unknown.

Using mouse leukemia models as primary screens—with subsequent testing in a few transplantable syngeneic rodent solid tumor models and later a small number of human tumor xenografts—had certainly been valuable in discovering and developing many clinically important cytotoxics, for example, paclitaxel, cisplatin, and topoisomerase II inhibitors (3, 5). However, there was considerable frustration about the limited number of cancer drug approvals for the common solid cancers, the long times from discovery to marketing, and importantly the inability of the available preclinical models to predict clinical success in cancers overall, and especially in any particular histologic subtype (2, 3, 5).

This exasperation encouraged what was certainly a radical and ambitious change in direction pioneered by Dr. Mike Boyd and colleagues in the NCI Developmental Therapeutics Program (DTP). They advocated abandoning the longstanding primary in vivo mouse leukemia models and their replacement by screening for differential in vitro sensitivity across a panel of 60 human tumor cell lines, representing leukemia, melanoma, and cancers of the lung, colon, brain, ovary, breast, prostate, and kidney. The new NCI-60 panel approach was referred to as “disease-oriented” as distinct from the previous, so-called “compound-oriented” screening, reflecting the underlying hypothesis that compounds showing promising patterns of in vitro activity against cell lines from cancers of a particular histology would translate more efficiently into drugs with activity against the corresponding tumors in patients.

Previously, the human tumor stem cell assay, an in vitro soft-agar colony-forming assay using tumor cells collected directly from patients, had been evaluated for compound screening, but proved unsuitable for the large-scale required by the NCI (2, 5). A more successful forerunner had been the use of large human lung cancer cell line panels in vitro. (2)

The initial landmark paper demonstrated the practical feasibility of a high-throughput, 96-well, plastic plate–based screen, requiring large-scale culture of authenticated cell lines, automated liquid handling, miniaturized microculture assay, use of a colorimetric tetrazolium-based metabolic dye reduction endpoint after several days’ incubation with drug, and demonstration of stable and reproducible sensitivity patterns over time, together with methods for high-volume data analysis. Without doubt, a great technological tour de force.

Following detailed discussions with advisers and technical refinements (including switching to the sulforhodamine B protein stain as the cell mass endpoint), the new strategy was approved in 1984 by NCI's Division of Cancer Treatment Board of Scientific Counselors (https://cancerletter.com/the-cancer-letter/19841026_1/). After a technical build-up phase, the NCI-60 in vitro screen ran full tilt from 1990, reaching its zenith during the 1990s, testing hundreds of compounds/day, ∼10,000 compounds/year. A 1990 article in Time magazine carried the title “Giving Up on Mice” and quoted Boyd saying, “It's a high-risk venture, but this is a gamble worth taking.” (https://content.time.com/time/subscriber/article/0,33009,971146–2,00.html). Dr. David Korn, Chair of the NCI's Advisory Board and Dean, Stanford University School of Medicine, commented, “All it will take is one smashing winner. Then everyone will say it was worth it.”

So, let us take stock of the achievements arising from the NCI-60 experiment. For this, I suggest using two success measures: (i) the hoped for increase in innovative, disease-oriented cancer drugs arising directly via initial testing in the screen and (ii) the unexpected gains in understanding the molecular underpinnings of drug response and the catalysis of a subsequent progressive evolution to even larger-scale molecularly annotated tumor cell panels for cancer drug research and development.

Many interesting new compounds were discovered through NCI-60 (Fig. 1B; refs. 2, 6), including ellipticinium analogs exhibiting selectivity in CNS-derived tumor cell lines, related to drug transport, with activity confirmed in orthotopic human glioblastoma xenografts, although not progressed to patients, and halichondrin B, from a marine sponge, as a new type of tubulin binder, leading to the synthetic analog eribulin gaining approval in breast cancer, liposarcoma, and leiomyosarcoma. An especially noteworthy development was the simple but inspired mean-graph data representation and powerful COMPARE pattern recognition algorithm, whereby unique NCI-60 tumor cell line activity “fingerprints” could indicate either a known mechanism of action (MOA) or a novel, unknown (COMPARE-negative) MOA (7). Arguably the drug benefiting the most was the proteasome inhibitor bortezomib. Exhibiting a unique COMPARE-negative profile, this was then active in the newly introduced “hollow fiber,” rapid in vivo prioritization assay (2, 5) and then in human tumor xenografts, and was subsequently approved in multiple myeloma, despite this tumor type not being represented in NCI-60. Additional COMPARE-negative compounds progressed to the clinic, although these did not lead to clinical activity in predicted cancer types (Fig. 1B).

COMPARE use frequently provided valuable mechanistic pointers, especially when coupled with detailed molecular annotation of NCI-60, positioning it as a “smart screen” (6). This is exemplified by the correlation of response to compounds across NCI-60 with multidrug resistance protein expression, supporting development of MDR1 inhibitors, and the concordance of response to anthrax lethal factor with that to inhibitors of the MEK/ERK pathway, enabling full mechanistic elucidation involving cleavage of MEKs (2).

Further important information science contributions included correlating NCI-60 response with gene expression microarray data and the use of neural networks and clustered heat map visualization, especially for MOA hypothesis generation (8). There were also undoubted technical benefits of NCI-60 for the natural product field and spin-off technical applications to anti-HIV compound screening in virus-containing cells by NCI, leading to the approval of lamivudine and abacavir (2, 3).

However, further developments were afoot in the 1990s. An extensive review of the NCI DTP in 1997 (https://deainfo.nci.nih.gov/advisory/bsa/bsa_program/bscdevtherprgmin.pdf) recommended major changes, especially in view of our rapidly burgeoning postgenomic knowledge of the molecular basis of cancer, many new drug targets emerging, and extensive industrial efforts to pursue molecular target–based drug discovery rather than random screening for cancer cell proliferation. Thus, the best role for NCI was seen as providing information, tools, and mechanistic hypothesis-generation capability for extramural researchers. Support for using NCI-60 for primary random screening was reduced and a prescreen with three tumor cell lines (breast, large-cell lung, and glioblastoma) was introduced using a single-test concentration to increase efficiency (2). However, it was recommended that NCI-60 be retained for access by external researchers for mechanistic studies, exploiting the powerful database of COMPARE profiles for 100,000 compounds already screened. Ongoing enhancements to NCI-60 include increased automation, microsizing to 384-well plates, and a switch to the commonly-used CellTiter-Glo luminescent cell viability readout (2).

Overall, the courageous NCI-60 random screening experiment did not lead directly to an increase in innovative cancer drug approvals or better preclinical prediction of clinical efficacy in particular classes of solid cancers. However, it did change our way of thinking, still operates for mechanism prediction, and delivered a range of valuable tools and resources now freely available online in the NCI DTP databases. Most important in my view, is NCI-60’s enormously important conceptual and technological legacy in the evolution to much larger panels of several hundred human tumor cell lines, with detailed molecular annotation, providing broader representation of clinical intertumor heterogeneity—as developed by the Sanger Institute/Massachusetts General Hospital and Broad/Novartis teams, with their pioneering papers published in 2012 (9). This led in turn to such widely used community resources as PRISM for compound screening and biomarker identification, Dep Map for RNAi and CRISPR loss-of-function screening for target and biomarker discovery, and the canSAR integrated knowledgebase platform for broader therapeutic research. Further evolution of models by NCI and others now includes large panels of patient-derived organoids and xenografts to better represent the complex tumor microenvironment as well as next-generation engineered mouse models with enhanced relevance to human cancer (10). When used as part of modern molecular target–based drug discovery (Fig. 1C), these various models now greatly empower modern biomarker–led precision oncology—with drug design prospectively focused on a clear line of sight to particular patient groups stratified by patient selection biomarkers that are specified on the drug label.

How drug discovery and models used have changed since the 1950s and even the 1990s! Still not perfect, more evolution continues, but increasingly useful.

P. Workman reports personal fees and other support from Alterome Therapeutics, Black Diamond Therapeutics, CHARM Therapeutics, NextechInvest, and STORM Therapeutics; nonfinancial support from CV6; other support from AstraZeneca and Epicombi Therapeutics; grants and personal fees from Astex Therapeutics and Merck KGaA; grants, personal fees, and other support from Nuvectis Pharma; personal fees from Vividion Therapeutics; and nonfinancial support from Vivan Therapeutics outside the submitted work; in addition, P. Workman is a nonfunded Executive Director of the nonprofit Chemical Probes Portal. No other disclosures were reported.

The author thanks Drs. Rosemarie Aurigemma, James Doroshow, Mathew Garnett, Barry O'Keefe, William Sellers, Bob Shoemaker, and Beverly Teicher for helpful discussions, and John Caldwell for drawing the chemical structures. He acknowledges recent and current support for his research from Cancer Research UK (CRUK Program Grants C309/A31322 and C309/A11566; Strategic Award C35696/A23187; and Infrastructure Award C309/A27413), Wellcome (Biomedical Resource and Technology Development Grant 212969/Z/18/Z to support the Chemical Probes Portal), Chordoma Foundation, Mark Foundation, Bone Tumour Research Trust, CRIS Cancer, and The Institute of Cancer Research. P. Workman is a CRUK Life Fellow.