Here there be dependencies: Putting cancers’ vulnerabilities on the map

Cancer cells can harbor a whole gamut of genetic errors, from small mutations to wholesale swaps of DNA between chromosomes — sometimes thousands of molecular flaws that should leave them dead. But when an error impacts a critical gene, a cancerous cell will compensate by adjusting other genes’ activity — increasing expression of another member of the same pathway, for instance.

From a researcher’s perspective, these adaptations — which allow the tumor to persist — represent dependencies: vulnerabilities that provide deeper insight into cancer biology and might serve as targets for designing new therapies, or for repurposing existing ones.

“Much of what has been and continues to be done to characterize cancer has been based on genetics and sequencing. That’s given us the parts list,” said William Hahn, an institute member in the Broad Cancer Program and an oncologist at Dana-Farber Cancer Institute. “Mapping dependencies ascribes function to the parts and shows you how to reverse engineer the processes that underlie cancer.”

That reasoning underlies the Cancer Dependency Map, a joint effort bringing together researchers from the Cancer Program’s Project Achilles, the Broad’s Genetic Perturbation Platform (GPP), and other teams across the institute. The team has spent nearly 15 years conducting genome-wide RNA interference (RNAi) screens on a growing number of cancer cell lines, probing thousands of genes individually for possible vulnerabilities.

In a study conducted as part of the Slim Initiative in Genomic Medicine for the Americas and reported in Cell, the Dependency Map team describes a major set of findings: 769 strong dependencies unique to cancer cells uncovered through RNAi screens of 501 cell lines representing a range of tumors. The list reveals intriguing themes in cancer cells’ survival strategies, and may also open new avenues for cancer drug development.

Fruition long coming

The data in the Cell paper represents an effort that reaches back to the earliest days of the Broad.

“In the early 2000s we worked out how to do pooled RNAi screens in mammalian systems well,” which gave researchers the tools to run genome-wide screens on many cell lines at once, said GPP director and institute scientist David Root, who, with Hahn, is one of the study’s co-senior authors. “That led us to doing genome wide screens on a dozen cancer cell lines,” work that he and his colleagues published in 2008.

RNAi effectively silences genes using small pieces of RNA called small interfering RNAs (siRNAs). These RNA tidbits bind to and call for the destruction of messenger RNAs (mRNAs) transcribed from individual genes, perturbing their expression. To run a genome-wide RNAi screen, researchers expose cells to pools of siRNAs, track the cells’ behavior, and work back to see which genes were silenced.

“The simplest thing one can do with perturbed cells is allow them to keep growing over time and see which ones thrive,” Root explained. “If cells with a certain gene silenced disappear, for example, it means that gene is essential for proliferation.”

Even those first dozen cell lines held revelations. For instance, tumor cells depended heavily on genes active in their original tissues (e.g., blood cancer cells needed blood-lineage genes, lung cells needed different genes). Other relationships were specific to individual cell lines, like one between cells from a chronic myelogenous leukemia (CML) line and ABL, a known CML driver gene.

But the team knew even then that they were not close to seeing the whole picture. “A dozen cell lines was far too few to really probe the breadth of dependencies,” Root said.

Watching tumors express themselves

Over the following years, Root, Hahn, and their collaborators — including the Broad Cancer Program’s Paquita Vazquez, Aviad Tsherniak, Cancer Program associate director Jesse Boehm, and Broad chief scientific officer and Cancer Program director Todd Golub — continued to systematically screen additional cell lines until they had comprehensive RNAi data (available via a dedicated online portal) from 501 lines curated by the Broad-Novartis Cancer Cell Line Encyclopedia (CCLE), representing multiple cancer types.

“Few places have tried to collect this kind of of data at this scale,” Hahn said. “But we felt that it was important to go after this many cell lines because it would give us a more comprehensive view.”

The total dataset revealed some striking patterns in the genes and pathways cancer cells come to depend on. Many dependencies were cancer-specific, in that silencing them each affected only a subset of the cell lines. However, more than 90 percent of the cell lines had a strong dependency on at least one of a set of 76 genes, suggesting that many cancers rely on a relatively few genes and pathways.

Using a set of molecular features (e.g., mutations, gene copy numbers, expression patterns) from each cell line, the team also generated biomarker-based models that helped explain the biology behind 426 of the 769 dependencies. Most of those biomarkers fell into four broad categories:

• mutation(s) of a gene

• loss of a copy or reduced expression of a gene

• increased expression of a gene

• reliance on a gene functionally or structurally related to another, lost gene (a.k.a., a paralog dependence)

Surprisingly, more than 80 percent of the dependencies with biomarkers linked to changes (up or down) in a gene’s expression. Mutations (often used as the grounds for pursuing a gene as a drug target) accounted for merely 16 percent.

Encouragingly, 20 percent of the dependencies the team discovered linked back to genes previously identified as potential drug targets.

“We can’t say we’ve found everything, but we can say that the genes we’re seeing fall into a relatively small number of bins, some of which are familiar, some less so,” Hahn said. “That initial taxonomy is a great starting point for building a full map.”

“Our results provide a starting point for therapeutic projects to decide where to focus their efforts,” said Vazquez, a study co-first author and a Cancer Dependency Map project leader. She added that while there was still much to do to validate the list, “it’s becoming increasingly easier to triangulate data and generate hypotheses as more genome-scale systematic datasets, like those from the CCLE, Genotype-Tissue Expression, and The Cancer Genome Atlas projects, become available.

“Bringing of all the data together,” she continued, “will help us generate a truly comprehensive cancer dependency map.”

Harvesting real hits

To make sure their RNAi screening results were accurate, the team had to overcome a significant obstacle: false positive results due to seed effects, which often complicate RNAi-based experiments.

Seed effects are a byproduct of how the RNAi process works. siRNAs use so-called seed sequences use to home in on their mRNA targets. But seed sequences are so short that siRNAs often inadvertently silence mRNAs that match the seeds but which are actually irrelevant.

“People sometimes take a dim view of RNAi because seed effects make the data so noisy,” said Tsherniak, a study co-first author and leader of the Cancer Program’s Data Science group. “And in fact, most of the signal in our screens before processing was driven by seed effects. It’s a big problem.”

To weed out the false positives, Tsherniak engineered a computational tool dubbed DEMETER (named for the Greek goddess of the harvest). “DEMETER models gene knockdown and seed effects within the data, and computationally subtracts the seed effects,” he explained. “It cleans up the data and helps you find true dependencies.”

Looking past mutations

According to Root, the work behind this paper represents movement towards a larger goal — creating a comprehensive map of all exploitable dependencies with cancer cells. But, he cautioned, there is a risk of looking at the map and focusing only on individual genes in individual tumors, rather than taking a broader view.

“These data are helping us realize the genetic networks behind human cells’ proliferation and viability requirements,” he said. “And for each dependency, we can ask, how specific is it to a particular tumor? Do we have ways to safely target that gene in people?”

Hahn added that the data argue that the time is ripe to pay more attention to the broader landscape of functional aspects of cancer, in addition to focusing on protein-coding gene mutations and variations.

“I think we’re close to the end of finding genes that are mutated or focally amplified in cancer,” he said. “To me, that’s a huge opportunity, because it means we have many heretofore untapped avenues for understanding cancer.”