IAAI Articles

14 AI MAGAZINE

Despite enormous strides in scientific understanding andmedical treatment, cancer still kills millions of peopleeach year. This article will explore why cancer has

proven to be such a formidable adversary, and how AI andmachine learning (ML) can save lives by helping individualpatients beat the odds.

Modern molecular biology supports the hypothesis that can-cer is actually hundreds or thousands of rare diseases, and thatevery patient’s tumor is, to some extent, unique (Pleasance etal. 2010). Although there is a rapidly growing arsenal of target-ed cancer therapies that can be highly effective in specific sub-populations,1 especially when used in rational combinations toblock complementary pathways, the pharmaceutical industrycontinues to rely on large-scale randomized clinical trials thattest drugs individually in heterogeneous populations. Such tri-als are an extremely inefficient strategy for searching the com-binational treatment space, and capture only a small portion ofthe data needed to predict individual treatment responses. Onthe other hand, an estimated 70 percent of all cancer drugs areused off label in cocktails based on each individual physician’sexperience, as if the nation’s 30,000 oncologists are engaged ina gigantic uncontrolled and unobserved experiment, involvinghundreds of thousands of patients suffering from an undeter-mined number of diseases.2 These informal experiments couldprovide the basis for what amounts to a giant adaptive searchfor better treatments, if only the genomic and outcomes datacould be captured and analyzed, and the findings integrated anddisseminated.

Toward this end we are developing Cancer Commons,3 a fam-ily of web-based rapid-learning communities in which physi-cians, patients, and scientists collaborate to individualize cancertherapy (Shrager, Tenenbaum, and Tavers 2011). The goals ofthis initiative are to (1) give each patient the best possible out-come by individualizing his or her treatment based on theirtumor’s genomic subtype, (2) learn as much as possible fromeach patient’s response, and (3) rapidly disseminate what islearned. The key innovation is to run this adaptive search strat-egy in “real time,” so that the knowledge learned from onepatient is disseminated in time to help the next.

Cancer Commons provides an exciting and important

Copyright © 2011, Association for the Advancement of Artificial Intelligence. All rights reserved. ISSN 0738-4602

Cancer: A Computational Disease That AI Can Cure

Jay M. Tenenbaum and Jeff Shrager

n Cancer kills millions of people each year.From an AI perspective, finding effective treat-ments for cancer is a high-dimensional searchproblem characterized by many molecularlydistinct cancer subtypes, many potential targetsand drug combinations, and a dearth of high-quality data to connect molecular subtypes andtreatments to responses. The broadening avail-ability of molecular diagnostics and electronicmedical records presents both opportunities andchallenges to apply AI techniques to personalizeand improve cancer treatment. We discuss thesein the context of Cancer Commons, a “rapidlearning” community where patients, physi-cians, and researchers collect and analyze themolecular and clinical data from every cancerpatient and use these results to individualizetherapies. Research opportunities include adap-tively planning and executing individual treat-ment experiments across the whole patient pop-ulation, inferring the causal mechanisms oftumors, predicting drug response in individuals,and generalizing these findings to new cases.The goal is to treat each patient in accord withthe best available knowledge and to continual-ly update that knowledge to benefit subsequentpatients. Achieving this goal is a worthy grandchallenge for AI.

IAAI Articles

SUMMER 2011 15

domain for the application of AI and ML. Achiev-ing its goals will require developing models for pre-dicting individual responses to cancer therapiesand using them to plan and run thousands ofadaptive treatment experiments. The genomic andclinical response data from each patient can beanalyzed to continuously improve the predictivemodels, as well as our understanding of cancerbiology and drugs. Each patient is treated in accordwith the best available knowledge, and that knowl-edge is continually updated to benefit the nextpatient.

In the next section, we review the current para-digm for translational cancer research, and its lim-itations, as we enter the new era of genomics andpersonalized medicine.4 Then we introduce CancerCommons as a new, open-science paradigm forreal-time translational research, with no daylightbetween lab and clinic. The fourth section exploresthe opportunities and challenges at the intersec-tion of Cancer and AI research. Finally, we con-clude with an invitation to participate in the grandchallenge of curing cancer.

Cancer Research in the Era of Genomics and

Personalized Medicine

It is unlikely that a single cure will be found forcancer—it is not a single disease. To a first approx-imation, cancer is what happens when a normalcell mutates in such a way that it can reproducebut no longer control its growth. As genomesequencing becomes less expensive and morewidespread, scientists are finding a spectacularvariety of mutations that lead to cancerous growth(Pleasance et al. 2010). Consensus is building that,because cancer is an individualized disease, it mustbe addressed through individualized treatment(see, for example, McDermott, Downing, and Strat-ton [2011]).

Developing and justifying individualized cancertreatments is a very difficult problem. First, eachorgan-specific cancer involves many rare genomicalterations, thus leading to highly heterogeneousdisease populations. This heterogeneity flowsdirectly from the underlying biological and genet-ic complexity of cancer. The mutation history ofeach tumor cell is also unique—the result of cas-cading and compounding errors in the cell’smachinery for maintaining the integrity of itsDNA, as well as evolutionary adaptations that helpit evade therapy.

Genetic mutations typically manifest them-selves through cell signaling pathways, chains ofprotein-catalyzed reactions that govern basic cel-lular functions such as cell division, cell move-ment, how a cell responds to specific external stim-

uli, and even cell death. The proteins in a pathwayoften communicate by adding phosphate groups,which act as “on” or “off” switches, to a neighbor-ing protein. When one of the oncogenes that codesfor such proteins is mutated, the resulting proteincan be stuck in the “on” or “off” position, leadingto uncontrolled cell growth, and thus cancer.

In contrast to traditional chemotherapy, whichtargets all rapid cell growth, modern targeted can-cer therapies attempt to block specific molecules inthe signaling pathways. Targeted therapies may bemore effective than current treatments and lessharmful to normal cells. However, a given patientmay require a custom cocktail of targeted therapiesto block both the pathways directly affected by anoncogene as well as secondary pathways that mayenable the cell to evade therapy.

Large-scale clinical trials are problematic ingenomic diseases like cancer because they rely onpopulation statistics rather than individualresponses. A drug that works on 50 percent ofpatients tested may or may not be better for anygiven patient than one that works on 20 percent.This inability to account for individual responsesmay explain why so many late-stage trials fail todemonstrate statistical efficacy, even though a fewindividuals do respond. What disease did theseresponders have, that used to be called “breast can-cer” or “melanoma,” for which there is now aneffective drug? How many other cancer patientshave the same genomic disease? Unfortunately,we’ll never know because when a trial fails, thedrug is, as a rule, abandoned.

One happy exception to this rule involvesGleevec (Imatinib), originally developed in 1998by Novartis for chronic myelogenous leukemia(CML). Gleevec was among the first of a new gen-eration of targeted therapies developed to block aspecific enzyme—in this case, BCR-ABL—in a can-cer signaling pathway. Shortly after its develop-ment, it was experimentally observed that Gleevecalso blocked a structurally similar enzyme called c-KIT, which was a primary cause of gastrointestinalstromal tumors (GISTs), a rare form of sarcoma(Demetri 2002). Gleevec was subsequently testedon many forms of cancer, including a 2003 trial formelanoma that failed (Kim et al. 2008). Therewere, however, a handful of responders in that tri-al, all of whom had two things in common: theirprimary tumors were located not on their skin, butin places where the sun don’t shine (for exampleinside the mouth), and their tumors all had amutated c-KIT gene. In 2008, results of a definitivetrial testing Gleevec on patients whose tumors hadbeen pretested for c-KIT mutations demonstrated amajor response (Hodi et al. 2008). A randomizedtrial would have needed perhaps a million patientsto demonstrate statistical significance for a drugthat affected at most 3 percent of melanoma

IAAI Articles

16 AI MAGAZINE

they pertain to patients who urgentlyneed new options. Over time, theanswers should lead to MDM modelswith increasingly specific molecularsubclasses, linked to increasingly effi-cacious therapies.

Cancer Commons’ rapid learningloop is illustrated in figure 1. It appliesthe same continuous improvementphilosophy to drug development andclinical trials that management guruslike W. Edwards Deming used in the1980s and 1990s to slash time and costsin automobile manufacturing andmany other industries. It is also relatedto the concept that Andy Grove, theformer CEO of Intel, had in mind whenhe advocated, in an influential JAMAeditorial, that the pharmaceuticalindustry should strive to achieve thefast “knowledge turns” achieved bychipmakers (Grove 2005). According toThe Economist, “A knowledge turn […]is the time it takes for an experiment toproceed from hypothesis to results, andthen to a new hypothesis—around 18months in chipmaking, but 10–20 yearsin medicine.”5

In addition to accelerating clinicalresearch, Cancer Commons also accel-erates communication of the results andsubsequent collaborations. The MDMsserve as living review articles, main-tained online and continuously updat-ed by top cancer experts. Clinicians andresearchers can post peer-reviewed clin-ical observations and data, anecdotalcase reports, rational treatmenthypotheses, and actionable researchfindings to the relevant MDM sub-type—preliminary results that are tooearly for formal presentation in journalsor at conferences but that may be help-ful to late-stage patients running out ofoptions. Physicians and patients whochoose to act on these postings canthen report clinical outcomes and sideeffects from their “N of 1” experimentsthrough the web. Treatment hypothesesthat prove successful in a few patientscan be generalized to a specific molecu-lar phenotype and validated in proof-of-concept studies or small clinical trials.Hypotheses that hold up can ultimatelybe incorporated into the guidelines asthe new standard of care. Hypothesesand observations that don’t hold up canbe quickly rejected or revised based onthe data points from human subjects.

patients. There’s more to this story, andwe’ll return to it several times.

As we enter the age of personalizedgenomic medicine, the Gleevec experi-ence will become the norm. At themolecular level, cancer is hundreds ifnot thousands of unique diseases. Giv-en that there are more than 800 thera-pies in development targeting specificmutations, and that they will oftenhave to be used together in cocktails toensure a durable response, there simplyisn’t enough time, money, patients, orspecimens to test treatments individu-ally on large heterogeneous patientpopulations. We need a smarter way tosearch for effective treatments, onethat utilizes the complete molecularand clinical profile of every patient toefficiently decide which drugs are like-ly to work best in a specific patient.

Community oncology practices,where the vast majority of patients aretreated, provide a promising way for-ward. Every day, thousands of patientswho have exhausted the standard ofcare are treated with off-label drugs andcocktails. These treatment decisions arebased largely on the judgments andexperience of individual physicians,and the results are seldom reported.What if we could coordinate thesethousands of ad hoc “N of 1” experi-ments, capture their results, and inte-grate them with the evidence fromlarge-scale controlled trials? CancerCommons provides the platform fordoing just that, transforming theeveryday practice of oncology into agiant adaptive search for better treat-ments.

Cancer Commons Cancer Commons is a family of web-based rapid learning communitieswhere physicians, scientists, andpatients collaborate to individualizetherapies based on the molecular pro-file of each tumor. Cancer Commonsprovides physicians with the knowl-edge and resources to treat each patientin the best possible way, and to dis-seminate the resulting data and knowl-edge as efficiently as possible to helpsubsequent patients. Cancer Commonsis, in essence, a huge adaptive clinicaltrial, whose goal is to continuouslyimprove patient outcomes, rather than

to demonstrate the efficacy of a specificdrug for regulatory approval.

Cancer Commons is being devel-oped one cancer at a time, beginningwith melanoma. At the core of eachCommons is a molecular disease mod-el (MDM). MDMs are expert-curatedreference models enumerating theknown molecular subtypes of a cancer.Each subtype is linked to relevant path-ways, diagnostic tests, approved andexperimental (targeted) therapies, andclinical trials. At each point in time,oncologists can use the appropriateMDM to treat patients with the bestavailable therapies for their tumors’molecular subtype. The subtypes andassociated therapies can be continuallyrefined based on how patients respond.For example, subtypes can be split cor-responding to responders and nonre-sponders, and new subtypes added toaccommodate previously unseentumor types. Conceptually, all patientswithin Cancer Commons are partici-pating in a huge, continuously run-ning, adaptive clinical trial that is con-stantly testing and refining both theMDM subtypes and their potentialtreatments.

The power in the Cancer Commonsapproach lies in the rapid feedbackloop that is generated when a patient isrationally treated based on his or hermolecular subtype but does notrespond as predicted. Armed with stateof the art genomics technology,researchers can attempt to uncoverwhy that individual failed to respondand apply the findings to the nextpatient with the same molecular sub-type. Studying response outliers can beparticularly informative. Earlier, we dis-cussed the example of a handful ofmelanoma patients who responded toGleevec in a trial that failed to meet itsclinical end points. The flip side is thesubset of patients who fail to respondin a successful trial. For those patientsthat were directed to the trial throughCancer Commons, one can go back tothe scientists on whose research themolecular disease model was based andask them to revise the model based onthe new molecular and clinical datafrom the subset of nonrespondingpatients. What disease subtype didthey have and how should it be treat-ed? These questions are not academic;

IAAI Articles

SUMMER 2011 17

A Melanoma Molecular Disease Model (MMDM) Figure 2 depicts a portion of the current MDM formelanoma, created by CollabRx, Inc. (Vidwans etal. 2011) This expert-curated model summarizeswhat is currently known about each molecularsubtype of melanoma—which biochemical path-ways are disrupted, their prevalence, their geneticmarkers, and, critically, how they respond to ther-apeutic interventions.6 Figure 2a illustrates themajor signaling pathways implicated inmelanoma, including the map kinase (MAPK)pathway (on the left) and the AKT/PI3K pathway(on the right), which regulate cell growth, cell pro-liferation, and cell death. There is a lot of crosstalkbetween these pathways and their downstreameffectors. Figure 2b is a table summarizing the sub-types of melanoma and linking them to thesepathways. Each subtype is defined based on themutational status of a key oncogene or tumor sup-pressor gene in a given pathway (such as BRAF forsubtypes 1.1 to 1.4 and c-KIT for subtype 2.1). Eachsubtype is then further divided by mutations thatmay coexist in genes that play a supportive role(such as PTEN, AKT, and CDK4 in the case of sub-types 1.2, 1.3, and 1.4). The subtypes are orderedby their prevalence and potential for therapeuticintervention. All told, the model currently distin-guishes eight actionable subtypes. Melanomas thatdo not fit into these subtypes are parked in subtype9 (TBD) pending further research.

The MMDM represents the current state ofknowledge about melanoma subtypes and theirassociated molecular diagnostics and targeted ther-apies. It connects information about pathways, tar-gets, tests, drugs, and trials previously scatteredacross hundreds of publications and databases, notto mention information that previously existedonly in the heads of a few experts. It is clearly awork in progress. As oncologists use the MDM totreat patients, the subtypes and their associatedtherapies will be continually refined based on howindividual patients respond. Subtypes will be splitcorresponding to responders and nonresponders,and new subtypes added to accommodate previ-ously unseen tumor types. Over time, it would notbe surprising to see the model grow to encompassdozens of subtypes and sub-subtypes with distincttherapeutic options.

The MMDM appeared in a peer-reviewed scien-tific journal (Vidwans et al. 2011). The PDF, con-taining many tables, charts, literature references,and supporting data, is also available online,7 host-ed on a semantic media wiki (SMW) platform.8

SMW is an extension of MediaWiki—the wikiapplication best known for powering Wikipedia.The semantic extensions transform the MMDMinto a hyperdocument that can be rapidly navigat-ed to find all information relevant to a given topic

(such as subtype, test, drug, or trial) as well as addi-tional information on the web. Moreover, theSMW adds semantic annotations that make thecontent of the MMDM understandable by a com-puter. The MMDM can thus function as a collabo-rative database, whose contents can be accessedthrough the semantic web.

While valuable to medical and AI researchers,few practicing physicians or patients are likely touse the MMDM in either of the above forms toguide treatment decisions. What is needed areuser-friendly web-based applications that trans-form the knowledge in the MMDM into personal-ized, actionable information.

The Cancer Commons Platform and ApplicationsCancer Commons participants can interact withthe molecular disease model through web-basedand mobile applications (apps), each designed todo one task well (see figure 3). The first such app isa targeted therapy finder (TTF) for melanoma. Itwas developed by CollabRx, Inc.,9 and released atthe end of 2010 (see figure 4). The CollabRx TTFguides advanced melanoma patients and theirphysicians through a series of molecular tests todetermine their tumor’s subtype, and thendescribes experimental therapies based on the bestavailable knowledge. Additional apps, under devel-opment, will support outcomes capture and analy-sis, adaptive treatment planning, and collaborativeupdating of the MMDM based on the latest clini-cal and laboratory findings.

Learn Test

TreatAnalyze

Figure 1. Cancer Commons “Rapid Learning” Loop.

IAAI Articles

18 AI MAGAZINE

NRAS

PTEN

PI3K

Growth, metastasis

GI/S progression

BAK

BAX

AKTMEK

BRAF

GNA11GNAQ

ERK Bcl-2, Bcl-xL, Mcl-1

NOXA, PUMABIM, BID, BAD

DNA damage,cellular stresses

MITFProliferation

Apoptosis,chemosensitivity

p16CDK4/6

Cyclin D

p53

MDM2

p14ARF

mTOR

c-KIT

ARF/INK4A

Detailedsubtypes

1.1

1.2-1.3

1.2-1.3

1.4

2.1

3.1

3.2

4.1

5.1 MITF MITF Copy number HDAC inhibitors

BRAF inhibitors

NRAS NRAS

GNA11 Targeted sequencing MEK inhibitors

GNA11GNAQ/GNA11

c-KIT c-KIT

BRAF/CDK4

BRAF/AKT

BRAF/PTEN

BRAFMAPK

Targeted sequencing &Copy number/CGH

Targeted sequencing &Copy number

(BRAF inhibitors) ANDCDK inhibitors)

(BRAF inhibitors) AND (AKTinhibitors or mTOR inhibitors)

(BRAF inhibitors) AND(AKT inhibitors or mTOR

inhibitors)

Targeted sequencing

Targeted sequencing

Targeted sequencing & IHC

Targeted sequencing

Gleevec & other inhibitors

MEK inhibitors

Targeted sequencing MAPK & PI3K pathway inhibitors;Famestyl transferase inhibitors

1.4

Pathway(s) Key gene/biomarker(s) Diagnostic technologies Potentially relevant therapeutics

Figure 2. A Portion of the Current MDM for Melanoma, Created by CollabRx, Inc.

(a) A highly simplified depiction of cell signaling pathways known to be active in melanoma. The MAPK pathway communicates a signalfrom growth factor receptors on the surface of the cell to the DNA in the nucleus of the cell, triggering cell division. A defect in this pathwaycan lead to uncontrolled cellular proliferation. The PI3K pathway is involved in a range of cellular functions implicated in cancer, includingcell growth, proliferation, differentiation, motility, and survival. Interactions between the MAPK and PI3K/AKT pathways lead to prolifera-tion and survival of melanoma cells. The CDK pathway, shown in blue, regulates progression through the cell cycle, which includes stagesfor DNA replication, inspection (checkpoints), and repair. The P53 pathway plays several critical anticancer roles, initiating DNA repair whenDNA has sustained damage, and apoptosis (programmed cell death) if the damage proves to be irreparable. One or more of these pathwaysare mutated in the vast majority of melanoma cells. (b) A subset of the table relating the biochemical pathways to subtypes of melanomaand their possible treatments.

IAAI Articles

SUMMER 2011 19

The Cancer Commons application platform hasopen APIs for third-party developers to createsophisticated medical decision support apps that,like the TTF, combine the science in the MMDMwith relevant clinical and personal considerations(for example, the patient’s state of health andresponse to prior treatments, the availability ofreimbursement). Since the platform is web based,such apps can also access a rich ecosystem of med-ical and research resources, including moleculardiagnostics laboratories, specimen and data repos-itories, and high-throughput drug screening andanimal testing facilities.

The Cancer Commons platform—encompass-ing the MDMs, apps, and services ecosystem—pro-vides the computational framework for the rapidlearning loop illustrated in figure 1. Experthypotheses about subtypes and their potentialtreatments, codified in the MMDM, can be testedin the clinic as appropriate patients present.Hypotheses can also be tested in the laboratory onappropriate cell lines or animal models. The result-ing clinical and laboratory observations may giverise to new hypotheses, which can be vetted bythe experts and added to the model for future val-idation.

The Cancer Commons NetworkCancer Commons is being developed one cancerat a time in partnership with leading professionaland patient advocacy organizations, pharmaceuti-cal companies, medical centers, and health infor-matics companies. These disease commons form anetwork (figure 5) that will exploit the significanteconomies of scale and opportunities for crosslearning (for example, common pathways anddrugs) that come from rethinking cancer as amolecular disease.

Following the process developed in melanoma,each new Commons begins by recruiting a panel ofexperts. The panel develops an MDM for that can-cer. The knowledge in the MDM is then represent-ed in a semantic media wiki, so it can be accessedthrough apps. Several new commons are currentlyin various stages of development, including sarco-ma, breast, and lung cancer.

Linking these Commons into a knowledge net-work occurs naturally through shared conceptssuch as subtypes, pathways, targets, tests, drugs,and trials. For example, while many cancer path-ways have been identified, a dozen or so appearover and over again, including the MAPK, PI3K,

AppsPlatform

Adaptive Therapy PlannerResearch Apps Outcomes Analyzer Targeted Therapy Finder

PayersPharma’sLabsSpecimens Hospitals Clinical Trials

Figure 3. Schematic of the Cancer Commons Platform.

At the center is an applications platform and knowledge hub that draws from, and feeds, the semantically represented MDM (center right).Specific applications (and example users) are depicted at the top, and examples of services are depicted along the bottom.

IAAI Articles

20 AI MAGAZINE

CDK, c-KIT, and P53 pathways. These play a promi-nent role in melanoma and many other solidtumors. When cancers share a common pathwayor subtype, much of the information, experimen-tation, and hard-won research results can poten-tially be leveraged across them. In particular, tar-geted therapies that prove successful in one ofthese cancers might also work in the other. Theaforementioned example of Gleevec is an excellentcase in point; yts effectiveness against c-KIT muta-tions was first demonstrated on GISTs (a form ofsarcoma), and it was subsequently successfullyapplied to c-KIT mutated melanomas. The hopeand expectation is that many such synergies willemerge from Cancer Commons’ knowledge net-work—slashing years off the decade-long process ittook to establish the efficacy of Gleevec—first forGISTs, and then for melanoma.

Many experts believe that cancers will one daybe characterized as molecular diseases (for exam-ple, c-KIT or BRAF disease) rather than organ-basedones. The Cancer Commons network could accel-erate that future.

AI Opportunities and ChallengesFrom an AI perspective, finding effective treat-ments for cancer is a very high-dimensionality

search problem, guided by our rapidly increasingknowledge of cancer biology and drugs. The num-ber of potential hypotheses about the causes andassociated treatments of the disease is huge, espe-cially when complete genomic profiles and combi-national therapies are considered—exponentiallylarger than the number of patients that could rep-resent each possible combination of factors. It istherefore essential to use knowledge to focus thesearch on promising opportunities, learn as muchas possible from every patient, and generalize thatknowledge to benefit future patients.

Cancer Commons is an ideal framework for cre-ating human–machine knowledge systems to effi-ciently conduct this search for better treatments.Think of the networked MDMs as a web-basedblackboard and the apps as knowledge sources. TheMDMs organize the world’s knowledge about can-cer subtypes, pathways, tests, drugs, and trials, andthat knowledge is made available in a semanticweb–compatible format. The human–machineknowledge system that is Cancer Commons gener-ates and tests hypotheses, creates plans, and learns.This shared human-machine approach is essentialbecause although the complexity of the searchdemands computational management, it is ulti-mately human medical specialists who interactwith patients and who are responsible for carryingout treatment plans. It is imperative that the physi-cians and patients understand the general ration-ale behind all decisions and inferences.

The overarching challenges for the AI commu-nity are to use the Cancer Commons framework to(1) organize the world’s collective knowledge ofcancer and use it to predict treatment responses fora given patient; (2) use these predictions to devel-op treatment plans that provide high expected val-ue for each patient, while simultaneously maxi-mizing the potential for new learning; and (3)learn as much as possible from each patient’smolecular and clinical data to improve future pre-dictions while advancing the community’s knowl-edge of cancer. The remainder of this section focus-es on AI opportunities and challenges in thesethree areas.

Knowledge ChallengeOrganize the world’s collective knowledge of cancerand transform it into personalized, actionable infor-mation that can be used to guide treatment deci-sions.

The molecular disease model approach taken byCancer Commons posits that a great deal is actual-ly already known about how to treat many genom-ic subtypes of cancer, but this knowledge is scat-tered across hundreds of thousands ofpublications, clinical and biological datasets, andknowledge bases. Medicine has traditionally reliedupon the publication of review papers or clinical

Figure 4. Screen Shot of the Melanoma Targeted Therapy Finder App10

Developed by CollabRx, Inc., Based on the Melanoma MDM.

Selecting characteristics of a patient’s specific tumor and pressing “SEARCH”will report clinical trails and literature that are possibly appropriate for thisindividual patient.

IAAI Articles

SUMMER 2011 21

guidelines for dissemination of important results,but these compilations have two important limita-tions: they quickly become dated, especiallyregarding the latest undocumented results thatmay only be known to a few experts; and theirresults are not computationally accessible and socannot be efficiently utilized.

MDMs are a step forward. As previously noted,they are living review articles, continuously updat-ed by top cancer experts, and their results are avail-able to computers as Resource Description Frame-work (RDF) triples.11 However it’s a safe bet thatmuch of what is currently known about cancer ascodified in MDMs is incomplete, inconsistent,uncertain, or downright wrong. Cancer Commonsparticipants can post case reports, laboratory find-ings, news articles, hypotheses, and the like thatsupport or refute specific MDM hypotheses andrecommendations. MDMs can serve as a commonpoint of reference at scientific meetings and inpublications to facilitate collaborative research.

Still, organizing the world’s knowledge of canceris easier said than done. The genomics revolutionand associated technologies are driving exponen-tial increases in all medical knowledge and data,far outpacing the ability of physicians and patientsto keep up. A few examples will illuminate thescale of the problem. PubMed, the definitive med-ical information retrieval service run by theNational Library of Medicine indexes about 20 mil-lion medical abstracts. Approximately 750,000new articles are published annually in medicaljournals. ClinicalTrials.gov lists about 100,000active clinical trials. From 80 to 100 trial results arereported daily. The Gene Expression Omnibus(GEO) database includes some 450,000 data setsfrom gene array experiments. Approximately 25percent of all these figures involve cancer.

The American Society for Clinical Oncology(ASCO) holds its annual meeting each June inChicago’s McCormick Hall. It is the only venuelarge enough to accommodate the 35,000 oncolo-

Sarcoma

Breast

Colorectal

Melanoma

Lung

Prostate

Figure 5. Figurative Depiction of the Cancer Commons Network.

Tripartite icons represent knowledge hubs (for example, MDMs; see figure 3). Circles represent applications and other services that bothdraw from and feed into the hubs.

IAAI Articles

22 AI MAGAZINE

the next patient. This analysis must bedistributed across all patients. A keychallenge is balancing these planningobjectives given the inherent tensionbetween treating patients (do every-thing you can to make this patient bet-ter) and populations (learn as much asyou can now to help better treat futurepatients). The planning challenge sub-divides into separate challenges in theareas of clinical treatment andresearch.

Treatment Planning. Treatment plan-ning for an individual is a classic prob-lem of planning under uncertainty andconstraints. The basic treatmentoptions could come from standardtreatment guidelines, an MDM, or oth-er predictive models. Ideally, eachoption would be accompanied by a pri-ori likelihoods of success that dependon the molecular subtype of thepatient’s tumor. Usually, there will alsobe constraints, including cost, avail-ability (for example, can the patient getinto a trial), and general health consid-erations that might contraindicate atreatment. Sometimes the planner canrefine the odds by ordering additionaltests, but these carry their own costsand constraints—for example, can thepatient afford to defer treatment, arethere suitable tumor specimens avail-able for testing.

The planning process must includenot just the science, as reflected intoday’s MDMs, but also qualitativeclinical judgments and patient prefer-ences. Even among drugs that “work,”there are still judgments that go intodecision making that aren’t universal.For example, many of today’s single-agent targeted therapies have a highlikelihood of response in the shortterm, but the patients are prone torelapse. Immune therapies (for exam-ple, Interferon or a cancer vaccine), onthe other hand, typically have less thana 20 percent response rate, but thoseresponses tend to be durable. Applyingwhat is known scientifically to individ-ual patients is challenging.

The planning process becomes evenmore interesting, and (a lot!) more com-plicated, when it includes combinationand sequential therapies. Where morethan one drug is involved questionsarise concerning dosing and sequenc-ing, for which answers might not be

gists and cancer researchers whoattend. Over the course of 4 days, some4000 abstracts are presented, summa-rizing the results of cancer clinical trialsconcluded in the past year. About adozen of those abstracts are deemedimportant enough to be practicechanging and are presented in plenarysessions. One can often read aboutthese in the New York Times. The resteffectively never see the light of day. Itis likely, however, that many otherabstracts, while not practice changing,might be life saving for the rightpatients; even failed trials often havesome responders, just not enough forstatistical significance. If only we couldfigure out what clinical or molecularfeatures characterized these subpopula-tions and get the lifesaving abstracts infront of those patients and their physi-cians who might benefit. These are cen-tral objectives of Cancer Commons,and AI is essential for achieving them!

Medical advances are also stymiedby the serendipity of scientific commu-nications. It can take decades to con-nect the dots among what is alreadyknown to infer a new treatment. Thestory leading up to the successful test-ing of Gleevec on c-KIT melanomasactually began in 1991 when Ruth Hal-aban, a melanoma researcher at Yale,discovered that c-KIT mutations wereresponsible for the uncontrolledgrowth of a few percent of melanomas.At the time, there was nothing anyonecould do about it. She published apaper in Cancer Metastasis Review andmoved on (Haliban 1991). In 1998, ateam led by Brian Druker at the OregonHealth Sciences University demonstrat-ed that an experimental signal trans-duction inhibitor known as STI571 waseffective in treating CML, a blood can-cer caused by a single oncogene (BCR-ABL) (Mauro and Druker 2001). In2001 Druker and others published anabstract at the ASCO Annual Meetingdemonstrating that STI571 was alsoeffective in controlling gastrointestinalstromal tumors expressing mutated c-KIT (Blanke et al. 2001), a protein struc-turally similar to that produced byBCR-ABL. There was now a promisinginvestigational drug for c-KIT tumors.Three years later, an Italian researchernamed Giammaria Fiorentini heardabout using STI571 to treat GISTs at an

Italian oncology meeting. He recalledHalaban’s 1991 paper associating c-KITwith some melanomas and published aletter to the editor describing a few iso-lated cases of c-KIT mutatedmelanomas that they had successfullytreated using Gleevec (Fiorentini et al.2003). Five years later, Stephen Hodiand colleagues at the Dana Farber pub-lished a definitive paper in the Journalof Clinical Oncology titled “MajorResponse to Imatinib Mesylate(Gleevec) in KIT-Mutated Melanoma”(Hodi et al. 2008). It took 17 years toconnect three dots from Halaban toHodi, and it will likely take 17 morebefore this result is widely disseminat-ed to community oncologists.

A human expert like Fiorentini canconnect a few dots and experience a“Eureka!” moment. But this onlyinvites the question of how many oth-er effective treatments might be foundif one were to systematically (that is,computationally) follow similar con-nections through the millions ofPubMed abstracts and other onlineknowledge resources. How many sub-types of cancer might already be cured,or at least manageable, if we only knewwhat we know? AI can help! Table 1outlines some important AI opportuni-ties in data and knowledge manage-ment for Cancer Commons.

Planning ChallengeAdaptively plan individual treatmentprotocols to achieve optimal outcomeswhile maximizing the learnings forother patients and cancer research.

From a planning perspective, thefundamental question is: How do youbest treat the patient in front of youbased on facts about that patient (suchas medical history, genomic informa-tion), the state of knowledge in thefield (results of clinical trials, publishedfacts, and so on), domain expertise(experience of the physicians), andprospective goals (planning of patient’streatment and optimizing learningacross the field)? This problem hasboth retrospective and prospectivechallenges, the former being classifyingthe patient and choosing the appropri-ate therapy, and the latter being theproblem of determining what knowl-edge and data are missing that wouldhelp optimize the treatment plan for

IAAI Articles

SUMMER 2011 23

readily available because the drugs were never test-ed together in a trial. Indeed, if either of the thera-pies is still experimental, it may not be available foruse in a cocktail. This situation arises frequentlywith targeted therapies for melanoma and is highlyfrustrating in the face of compelling evidence thatmultiple mutated pathways, for which experimen-tal drugs exist, need to be blocked.

Most cancer patients require a sequence of treat-ments and multiple treatment modalities (such assurgery, radiation, chemotherapy). A treatmentmay work for a while, but ultimately will have tobe augmented or replaced by a second- or third-line therapy when the cancer develops resistance.Good treatment plans maximize “shots on goal.”Picking one treatment can preclude future options(for example, taking drug A can exclude a patientfrom a trial for drug B, so to preserve options, drugB should be tried first). There is thus a need to con-sider potential interactions and plan ahead. Oneshould also take into account responses to previ-ous treatments as they can be important clinicalpredictors of how a patient will respond to a futuretherapy. Doctors find this kind of optimization dif-ficult, but it is a simple matter for a constraint-

based planner, given the right constraints or theability to learn them.

Research Planning. Treatment planning for indi-viduals takes place within a larger scientific con-text. Here the goal is to coordinate planning acrossall patients to efficiently explore treatment alter-natives and fill in knowledge and data gaps. Plan-ning at this level is complicated by the ambiguityand noisy data inherent in medical experiments. Itis also strongly constrained by ethics and the inter-ests and incentives of individual physicians.

Balancing treatment and research planninginvolves the classic exploration-exploitationdilemma from the reinforcement-learning com-munity (Sutton and Barto 1998) but with someimportant new (ethical) constraints: one wouldlike to do as much exploration as possible whilemaximizing the outcome for the current patientand respecting that patient’s treatment prefer-ences. It is an important open question as towhether the degree of exploration possible underthese constraints will be adequate.

In principle, one could coordinate treatmentplanning over a population to optimize informa-

Modeling

Build or learn models of cancer biology and care that are capable of representing both causal and probabilistic knowledge aboutdiseases, pathways, targets, and drugs.

Use these models to organize the copious information about cancer now scattered across the web for ef!cient use by people and computers (including data, scienti!c literature, social media, case reports, electronic medical records, and even speculative hypotheses). Given the state of medical knowledge, the models must be capable of encompassing inconsistent and contradictory information.

Knowledge Acquisition

Populate the models by mining major cancer information sources (for example, journals, websites, blogs, and databases). These miners will identify relevant content and tag it with metadata, thereby linking the content to elements of the disease models (for example, subtypes, pathways, targets, drugs, and trials).

Use domain-speci!c natural language processing (NLP) to extract knowledge from published content and informal communications (for example, transform a paper on pharmacogenomics into a set of probabilistic assertions about how various mutated genes alter the chemosensitivity of drugs; extract inclusion and exclusion criteria from clinical trial announcements; mine case reports and associated discussions for evidence to support alternative treatment hypotheses).

Search and Question Answering

Use the models to guide deep, vertical search for cancer-related information (for example, !nd and organize all the information a patient might need in order to evaluate the safety and ef!cacy of a drug or trial).

Provide direct answers to frequently asked questions (for example, “!nd the best trials for my cancer subtype” should return a table of ranked candidates).

Inference

Develop domain-speci!c inference engines that “connect the dots” across knowledge sources to discover new knowledge (for example, discover new uses for existing targeted therapies by searching for subtypes of other cancers that are structurally similar targets; discover new cancer mechanisms by linking two previously disjoint cancer signaling pathways that share a common protein).

Use probabilistic inference to rank targets, drugs, and trials by integrating evidence pro and con from experimental data, the literature, and the Cancer Commons community.

Collective Intelligence

Develop crowd-sourcing approaches that augment automated NLP (for example, community tools for submitting, tagging, ranking, reviewing, commenting on, and recommending content).

Develop tools that make it easy for experts to edit the models and to add or annotate content (for example, suggest new hypotheses and link them to supporting evidence in the models).

Build interactive web forms that enable authors to codify the key concepts in their papers and postings and connect them to concepts in the models. Also, enable them to correct errors introduced by others (for example, NLP and crowd sourcing).

Build tools for collective decision making (for example, enabling physicians to collaborate on a treatment plan that synthesizes their individual experiences and knowledge from the model).

Table 1. Data and Knowledge Management Opportunities.

IAAI Articles

24 AI MAGAZINE

tion gain. Imagine, for example, running a defini-tive set of treatment experiments to systematicallyrule out an entire family of drugs or drug targets asviable therapeutic candidates, rather than foreverrevisiting the same dead ends (as is all too often thecase today). In practice, it could be hard to con-vince a group of physicians, with individual inter-ests and incentives, to go along with such a plan.Table 2, summarizing the opportunities for auto-mated planning in Cancer Commons, suggestssome ethical strategies for exploiting the daylightbetween individual benefit and group knowledgeacquisition.

Learning ChallengeUse data from thousands of individual treatmentexperiments to infer the causal mechanisms oftumors and drugs, and develop predictive modelsfor individualizing therapy. Generalize the resultsso they can be applied to new cases.

Cancer Commons is quintessentially aboutlearning. So far, the focus has been on humanlearning—tapping the collective intelligence ofphysicians and patients. However, machine learn-ing can complement human learning by seekinghigher-level patterns in the clinical and genomicdata that Cancer Commons will collect. Addition-al insights can be gleaned by mining the massiveamounts of drug response data that are becomingavailable from laboratory experiments.12

Knowledge about cancer exists at many otherlevels of abstraction, such as causal models oftumorigenesis and pathway signaling. Connecting

the dots between these causal models and empiri-cal models of drug response—and filling in themissing knowledge gaps—is hard to do directly. Ithelps to decompose the overall search into man-ageable subproblems using intermediate-levelmodels. The molecular disease models at the heartof Cancer Commons are an interesting case inpoint: MDMs link molecular subtypes of a cancer,which are rooted in causal pathway models, topromising drugs and clinical trials, often selectedon the basis of empirical data. MDMs can be usedto seed ML algorithms that classify tumors intomolecular subtypes (for example, through cluster-ing), as well as algorithms that predict therapyresponse based on clinical and laboratory data.

There are many other models that are common-ly used in cancer research, ranging from the causalmodels of cancer pathways used in systems biolo-gy13 to clinical guidelines. The challenge is first touse machine learning to improve these modelsindependently, and then to chain them togetherto obtain a consistent understanding of cancer atmultiple levels. If these challenges can be met, itmay one day be possible to use a tumor’s biology topredict drug response in individual patients.

Table 3 proposes some relevant machine-learn-ing projects for Cancer Commons. In addition,there is a rich literature of AI and machine learningapplications in cancer biology and medicine thatwould be interesting to integrate into Cancer Com-mons. We invite potential collaborators to contactus regarding these and other opportunities.

Cancer Commons provides a rich context for

Treatment Planning—Develop an incremental planning method that optimizes treatment plans (that is, a sequence of tests, therapies, and trials) for individual patients, based on evolving information and constraints, including:

The patient’s clinical history, including diagnoses, treatments and responses.

Evolving clinical guidelines for standards of care, and MMDM recommendations (based on the tumor’s subtype) for patients beyond the standards of care.

Patient preferences (for example, chemotherapy versus immunotherapy, geographic location, risk tolerance).

Availability of drugs (for example, through a trial, or through compassionate use); trial exclusion criteria may dictate the order in which treatments are tried.

Availability of tumor specimens for molecular testing (to predict therapy response).

Financial considerations, including insurance coverage for tests and drugs.

Complicating factors include, for example, dosing and sequencing of combination therapies, prioritizing trials to maximize shots on goal.

The planner should support shared decision making by the patient and physician.

Research Planning—Coordinate individual treatment plans to ef!ciently !nd the best therapy for each cancer subtype. Overcome ethical con"icts that may arise in treating individual patients by:

Randomizing patients to trials or drugs that are deemed equivalent, or whose ef!cacy is unknown; as evidence mounts, adaptively direct patients toward the more effective therapies.

Trying alternative therapies when the best choice is excluded by externalities such as the patient’s state of health, personal preferences, prior treatments, or drug availability (for example, the patient can’t get into a trial because of brain metastases).

Utilizing historical controls and observational data to assess trials and drugs that are under-represented in existing Cancer Commons “N of 1” experiments.

Globally coordinating the search to ensure adequate coverage of all major drug classes and targets, while minimizing redundant testing of the most popular ones.

Table 2. Planning Opportunities.

IAAI Articles

SUMMER 2011 25

ML projects like those proposed in table 3. Therewill be opportunities to cross-validate results fromhuman and machine learning, as well as fromcausal and empirical models. There will be “trans-fer learning” opportunities across different organ-based cancers through the Cancer Commons net-work—using an approved cancer drug off label totreat a different cancer with a similar moleculartumor profile. The goal is to generalize the findingsfrom a few patients to a much larger group ofpatients like them. Opportunities like these willdrive research in both ML and cancer and, wehope, save lives.

ConclusionCancer Commons is the harbinger of a newpatient-centered paradigm for cancer research inwhich every patient receives individualized thera-py based upon the best available knowledge andwhere the resulting data contributes as efficientlyas possible to improved treatment for each subse-quent patient. AI and machine learning will beessential to realizing this vision.

Cancer, as a domain, presents the AI and MLcommunities with extraordinary challenges and,hence, opportunities. These include (1) managingthe sheer quantity and heterogeneity of the dataand knowledge involved—encompassing millionsof medical records, genomewide datasets, and doc-uments; (2) planning thousands of complex, mul-tistep treatment strategies that ethically balance

the needs of the individual with those of science;(3) capturing and analyzing the results of thesetreatment experiments in diverse causal andempirical models of cancer biology and drugresponse; (4) continuously testing and refiningthese models to account for new clinical and labo-ratory findings; (5) generalizing the models acrosspatients and cancers and integrating them toimprove decision making; and (6) integratinghuman and machine planning, learning, and deci-sion making to exploit their respective strengths.

Taken together, these challenges add up to thegrand challenge of curing cancer. This challenge isa worthy heir to the succession of grand chal-lenges—from chess to unmanned vehicles—thathave driven so much progress in AI over the years.While AI may not be able to cure cancer any timesoon, it may soon make life or death differences inoutcomes for individual patients. Please join usand the Cancer Commons community in rising tothis challenge.

Grand ChallengeDevelop AI applications that tap the world’s collec-tive knowledge and data on cancer to improvepatient outcomes and save lives.

To learn more about this grand challenge, pleasevisit www.cancercommons.org.

AcknowledgementsThe authors are indebted to our colleagues at Col-labRx and on the Cancer Commons melanoma

“N of 1” Experiments—Given clinical and genomic data from a series of patients with a given type of cancer.

Apply hierarchical clustering techniques to determine subtypes and validate them against those in the expert-curated MDM.

Determine the optimal therapy for each individual, based on previous clinical outcomes and therapy responses for patients with the same subtype.

Replace clustering with more sophisticated approaches to tumor subtyping, such as using genomewide expression and sequence data to infer dysfunctional upstream pathways and then clustering patients with the same aberrant pathways.

Replace subtyping with a similarity metric that matches patients directly, based on their molecular tumor pro!les and responses to previous therapies; use this metric to recommend future therapies, based on what worked in similar patients.

Clinical Trials—Reanalyze clinical and molecular data on individual patients in population-based trials.

Analyze the biological differences between responders and nonresponders to discover predictive biomarkers. Use them to salvage drugs that were effective in subpopulations. Meta-analyze data from multiple trials to enhance statistical power.

Design “intelligent” adaptive Bayesian trials14 that use AI techniques such as reinforcement learning (RL) to ef!ciently test molecular diagnostic and targeted therapy candidates. Unlike traditional trials, which test one or two drugs on a population, the goal here is to predict responses and maximize outcomes for individuals for a continuously evolving set of drug and biomarker candidates.

Laboratory Studies—Given results of high throughput drug screens on cancer cell lines.

Build a predictive model of drug response based on the molecular pro!les of the cells (for example, from microarray data).15

Use the model to recommend therapies for patients, based on the drugs that worked best for genomically similar tumor cells in the laboratory. If resources permit, validate the prediction in vitro or in vivo as appropriate.

Use laboratory and clinical results to continually re!ne the predictive model.

Literature Studies—Given a corpus of PubMed articles reporting on various interactions among genes, proteins, and drugs.

Infer network (pathway) models of cancer biology by integrating fragments of evidence on gene-protein and protein-protein interactions from the literature or from actual experiments (for example, Bonetta [2010]).

Meta-analyze the extensive literature on pharmacogenomics (the effect of genes and gene expression on drug response and drug-drug interactions) to create predictive models for drugs (that is, what are the optimal drugs for a patient whose tumor cells express particular genes and proteins) (for example, Mocellin et al. [2010] and Theobald, Shah, and Shrager [2009]).

Test predictions in the laboratory and ultimately in patients.

Table 3. Potential Machine-Learning Projects for Cancer Commons.

IAAI Articles

26 AI MAGAZINE

G. R.; Bignell, G. R.; Ye, K.; Alipaz, J.; Bauer,M. J.; Beare, D.; Butler, A.; Carter, R. J.;Chen, L.; Cox, A. J.; Edkins, S.; Kokko-Gon-zales, P. I.; Gormley, N. A.; Grocock, R. J.;Haudenschild, C. D.; Hims, M. M.; James, T.;Jia, M.; Kingsbury, Z.; Leroy, C.; Marshall, J.;Menzies, A.; Mudie, L. J.; Ning, Z.; Royce, T.;Schulz-Trieglaff, O. B.; Spiridou, A.; Steb-bings, L. A.; Szajkowski, L.; Teague, J.;Williamson, D.; Chin, L.; Ross, M. T.; Camp-bell, P. J.; Bentley, D. R.; Futreal, P. A.; andStratton, M. R. 2010. A Comprehensive Cat-alogue of Somatic Mutations from a HumanCancer Genome. Nature 463 (14 January):191–196.

Shrager, J.; Tenebaum, J. M.; and Travers, M.2011. Cancer Commons: Biomedicine inthe Internet Age. In Collaborative Computa-tional Technologies for Biomedical Research,ed. by S. Elkins, M. Hupcey, and A. Williams.New York: Wiley.

Sutton, R., and Barto, A. 1998. ReinforcementLearning: An Introduction. Cambridge, MA:The MIT Press.

Theobald, M.; Shah, N. H.; and Shrager J.2009 Extraction of Conditional Probabilitiesof the Relationships between Drugs, Dis-eases, and Genes from PubMed Guided byRelationships in PharmGKB. Paper present-ed at the AMIA Summit on TranslationalBioinformatics, San Francisco, March 15–17.

Vidwans, S.; Flaherty, K. T.; Fisher, D. E.;Tenebaum, J. M.; Travers, M. D., andShrager, J. 2011. A Melanoma MolecularDisease Model. PLoS ONE 30 6(3): e18257.doi:10. 1371/journal.pone.0018257.

Jay M. “Marty” Tenenbaum, Ph.D., waseducated at the Massachusetts Institute ofTechnology and Stanford University in the1960s. He spent the 1970s doing artificialintelligence research at SRI, the 1980s man-aging computer science research forSchlumberger, and the 1990s pioneeringInternet commerce. He’s currently focusedon using AI and the web to transform med-icine.

Jeff Shrager, Ph.D., is the CTO of CollabRxand a consulting associate professor in theSymbolic Systems program at Stanford Uni-versity. His research focuses on understand-ing how science works and how computerscan facilitate scientific discovery. Shragerhas cofounded several scientific-computingstartups and is the inventor of BioBike, aweb-based biological knowledge platformenabling biologists to develop, run, andshare symbolic analyses of genomic infor-mation.

editorial board, especially Smruti Vid-wans, Allan Schiffman, David Fisher,MD, and Keith Flaherty, MD, for theircontributions to the melanoma molec-ular disease model and therapy finderapplication. We also acknowledge thethoughtful input we received from col-leagues in machine learning and com-putational biology, particularly AtulButte, David Haussler, and DaphneKohler who compose the Cancer Com-mons computational advisory board,and Michael Littman and SriraamNatarajan who helped organize a work-shop and team proposal on AI and can-cer.

Notes1.  See www.cancer.gov/cancertopics/fact-sheet/Therapy /targeted.

2. See www.alternet.org/health/140234.

3. See www.CancerCommons.org.

4. “Translational research” is defined as thetranslation of scientific discoveries intopractical applications (commonfund.nih.gov/clinicalresearch/overview-translation-al.aspx).

5. See www.economist.com/node/14299624.

6. The MMDM was curated by a panel of 11melanoma experts led by cochief editorsDavid Fisher, MD, head of dermatology, andKeith Flaherty, MD, head of experimentaltherapeutics, at MGH.

7.  See mmdm.cancercommons.org/smw/index.php/A_Melanoma_Molecular_Dis-ease_Model.

8. See semantic-mediawiki.org.

9. See collabrx.com.

10. See therapy.collabrx.com.

11. Resource Description Framework (RDF)triples are used in the semantic web toencode knowledge in the form of subject-predicate-object expressions.

12. See, for example, sagebase.org.

13. See, for example, Ingenuity.com.

14. See, for example, www.sciencenews.org/view/generic/id/58401/title/BATTLE_trial_personal izes_lung_cancer_treatment,http://www.ispy2.org.

15. See www.broadinstitute.org/cmap.

ReferencesBlanke, C. D.; von Mehren, M.; Joensuu, H.;Roberts, P. J.; Eisenberg, B.; Heinrich, M.;Druker, B.; Tuveson, D.; Dimitrijevic, S.; Sil-berman, S. L.; and Demetri, G. D. 2001.Evaluation of the Safety and Efficacy of anOral Molecularly-Targeted Therapy, STI571,in Patients (Pts) with Unresectable orMetastatic Gastrointestinal Stromal Tumors

(GISTs) Expressing c-KIT (CD117). Proceed-ings of the American Society of Clinical Oncol-ogy 20: (abstr 1). Alexandria, VA: AmericanSociety of Clinical Oncology.

Bonetta , L. 2010. Protein–Protein Interac-tions: Interactome Under Construction.Nature 468: 851–854.

Demetri, G. D. 2002. Targeting the Molecu-lar Pathophysiology of GastrointestinalStromal Tumors with Imatinib. Mecha-nisms, Successes, and Challenges to Ration-al Drug Development. Hematology OncologyClinics of North America 16(5): 1115–24.

Fiorentini, G.; Rossi S.; Lanzanova G.;Bernardeschi, P.; Dentico, P.; and De Giorgi,U. 2003. Potential Use of Imatinib Mesylatein Ocular Melanoma and LiposarcomaExpressing Immunohistochemical c-KIT(CD117). Annals of Oncology 14(5): 805.

Grove, A. 2005. Efficiency in the HealthCare Industries: A View from the Outside.Journal of the American Medical Association(JAMA) 294(4): 490–492.

Haliban, R. 1991. Growth Factors and Tyro-sine Protein Kinases in Normal and Malig-nant Melanocytes. Cancer Metastasis Review10(2): 129–40.

Hodi, F. S.; Friedlander P.; Corless, C. L.;Heinrich, M. C.; Mac Rae, S.; Kruse, A.;Jagannathan, J.; Van den Abbeele, A. D.;Velazquez, E. F.; Demetri, G. D.; and Fisher,D. E. 2008. Major Response to ImatinibMesylate in KIT-Mutate Melanoma. Journalof Clinical Oncology 26(12): 2046–51.

Kim, K. B.; Eton, O.; Davis, D. W.; Frazier, M.L.; McConkey, D. J.; Diwan, A. H.;Papadopoulos, N. E.; Bedikian, A. Y.; Cama-cho, L. H.; Ross, M. I.; Cormier, J. N.; Ger-shenwald, J. E.; Lee, J. E.; Mansfield, P. F.;Billings, L. A.; Ng, C. S.; Charnsangavej, C.;Bar-Eli, M.; Johnson, M. M.; Murgo, A. J.;and Prieto, V. G. 2008. Phase II Trial of Ima-tinib Mesylate in Patients with MetastaticMelanoma. British Journal of Cancer 99(5):734–740.

Mauro, M. J., and Druker, B. J. 2001. STI571:Targeting BCR-ABL as Therapy for CML. TheOncologist 6(3): 233–238.

McDermott, U.; Downing, J. R.; Stratton, M.R. 2011. Genomics and the Continuum ofCancer Care. New England Journal of Medicine364(4): 340–350.

Mocellin, S.; Shrager, J.; Scolyer, R.; Pasquali,S.; Verdi, D.; Marincola, F. M.; Briarava, M.;Gobbel, R.; Rossi, R.; and Nitti, D. 2010. Tar-geted Therapy Database (TTD): A Model toMatch Patient’s Molecular Profile with Cur-rent Knowledge on Cancer Biology. PLoSONE 5(8): e11965. doi:10.1371/journal.pone.0011965

Pleasance, E. D.; Cheetham, R. K.; Stephens,P. J.; McBride, D. J.; Humphray, S. J.; Green-man, C. D.; Varela, I.; Lin, M.-L.; Ordóñez,