Speed SPYing: adaptive clinical trials hit the gas

The I-SPY 2 breast cancer clinical trial paves the way for multi-arm adaptive learning trials.
Haddon Township, NJ, USA.

Search for this author in:

The multi-armed I-SPY 2 breast cancer clinical trials are starting to fulfill the promises of smarter, more efficient, biomarker-driven trials long touted by reformers (Nat. Biotechnol. 25, 287–292, 2007). I-SPY 2 has now sped six drug candidates or combinations to phase 3 and has inspired a generation of Bayesian adaptive platform trials beyond breast cancer applications.

An eighteenth-century theologian, Thomas Bayes, came up with a formula for determining conditional probability that is increasingly used in the life sciences. Credit: The Picture Art Collection / Alamy Stock Photo

Donald Berry, a biostatistician at the University of Texas MD Anderson Cancer Center, an architect of I-SPY 2 (Investigation of Serial Studies to Predict Therapeutic Responses with Imaging and Molecular Analysis 2) and founder of the Bayesian statistical consultant group Berry Consultants, discussed trial outcomes at the 2019 Biotechnology Innovation Organization International Convention on 6 June 2019 in Philadelphia. Puma Biotechnology’s Nerlynx (neratinib), AbbVie’s veliparib, Genentech’s Perjeta (pertuzumab), and Merck’s MK-2206 and Keytruda (pembrolizumab) have moved on to phase 3 trials in breast cancer subtypes, as has a combination of Perjeta and Genentech’s Kadcyla (ado-trastuzumab emtansine) (Table 1).

Table 1 | Select Bayesian-platform trials



Year started



Breast cancer


Phase 2

Precision Promise

Pancreatic cancer

2016 (announced; enrolling patients 2019)

Phase 2/3

Glioblastoma Adaptive Global Innovative Learning Environment (GBM AGILE)


2015 (announced; enrolling patients 2019)

Phase 2/3

European Prevention of Alzheimer's Dementia Consortium (EPAD)

Alzheimer’s disease

2015 (announced)

Phase 2

Partnership for Research on Ebola Virus in Liberia (PREVAIL) II



Phase 2

Rapid Administration of Carnitine in Sepsis (RACE)



Phase 2

Berry says that I-SPY 2 launched in 2010 and received mixed enthusiasm from the community (Nat. Biotechnol. 28, 383–384, 2010). But the tide began to turn in 2016 when researchers published the results of the first two arms to graduate—Nerlynx (N. Engl. J. Med. 375, 11–22, 2016) and veliparib (N. Engl. J. Med. 375, 23–34, 2016)—in the New England Journal of Medicine. “People read the papers and observed there was this strange thing going on—there were no P values, it was Bayesian probabilities,” says Berry. “The reaction has moved from being really skeptical of these lunatics that are trying to do this to ‘this is transformative.’”

I-SPY 2 uses magnetic resonance imaging and ten biomarkers to direct patients to the right trial arm, monitor their progress and identify promising results as data become available, by using Bayesian statistical modeling to learn and adapt the trial rather than assembling the data afterward. Unlike randomized controlled trials, which must be designed from scratch, I-SPY 2 can add in new experimental agents or combinations to its ongoing trial without ramping up patient accrual, by using a single control arm for all the experimental arms. Its Bayesian algorithm incorporates pre-existing and constantly accruing data into probability predictions about efficacy, thus enabling faster decisions. When the algorithm predicts that a drug candidate’s efficacy is above 85% in a molecular subtype of breast cancer, the drug graduates and can move on to phase 3 studies. The result is a more nimble trial that can be started faster and can generate results more quickly: I-SPY 2 drugs can spend as few as 12 months in phase 2.

“I-SPY was truly groundbreaking,” says Victoria Manax Rutson, Chief Medical Officer for Pancreatic Cancer Action Network (PanCAN), adding that master protocols allow researchers to ask multiple questions simultaneously and to carve out the right niche for new drugs.

Inspired by I-SPY 2’s underpinnings, other similarly structured platform trials have since cropped up in a host of disease areas. One example is Precision Promise, a trial that PanCAN will launch this year in an attempt to accelerate development of drugs for pancreatic cancer.

Janet Woodcock, director of the US Food and Drug Administration (FDA) Center for Drug Evaluation and Research (CDER), predicts that platform trials will gain popularity once more of them are up and running, because of how quickly a company can get an arm started for an investigational drug. “All the infrastructure is done,” she says. “Compared with designing a protocol de novo, and then getting recruiting sites and then training the sites, this is much faster.”

Woodcock notes that some companies may still opt to skip platform trials if they are unwilling to cede control, because a platform’s design may not answer the same questions as protocols built around a single compound. “They’ve invested hundreds of millions of dollars in this, and often they want to fully control the parameters under which something is tested. A platform trial is really mostly for the benefit of the patients, where an individual trial is very investigational-agent centered.”

Laura Esserman, principal investigator for I-SPY 2 and director of the University of California San Francisco Carol Franc Buck Breast Care Center, says the trial adopted the Silicon Valley “fail early, fail fast” mantra. “And you want to learn early, and learn fast, so you can build on the drugs in your pipeline that are really successful. That’s actually good for everybody,” she says.

One example of the trial’s speed is in the I-SPY 2 tests for Keytruda in triple-negative (lacking estrogen receptor, progesterone receptor and the HER2/neu receptor) breast cancer (TNBC). The trial tested Keytruda in combination with standard therapy of paclitaxel followed by doxorubicin and cyclophosphamide, and surgery to remove the tumors. On the basis of the patients’ magnetic resonance imaging scans and biomarker signatures alone, the trial’s algorithm had already predicted that Keytruda combinations would have a greater than 85% probability of success in phase 3 trials—the threshold for graduation. That prediction, however, was possible after only one patient completed the full treatment course through surgery, thus creating a “sticky” situation, according to Berry. “I couldn't go to Merck and say, ‘We just graduated your drug with one patient,’” he says. So the trial remained open for an additional month, during which 11 of the next 12 TNBC patients had pathologically complete responses at the time of surgery.

This case exemplifies how a shift in mindset may be necessary for industry to adopt Bayesian recommendations. “The way we measure success is different from the way other people measure,” Berry adds. Instead of measuring how many patients enrolled in the trial and which patient subsets could move on to phase 3, “we look at how fast we do things.” Keytruda graduated from I-SPY 2 in less than a year. On 29 July, in a result that validates the signal seen in I-SPY 2, Merck announced Keytruda met a phase 3 endpoint of pathologic complete response in TNBC patients when used as a neoadjuvant and adjuvant.

Another way in which I-SPY 2 hits the gas is by using a single control arm instead of setting up 20 concurrently randomized arms for the 20 compounds that have entered the program. Berry calls the technique a statistical “time machine” because it allows researchers to save both time and money through updating the control by using data collected since the trial began, skipping duplicative efforts.

The cost savings could prove considerable. Running I-SPY2 is much less expensive than 20 separate trials conducted by 20 companies, each with its own control arm. Since its launch, Esserman estimates that the trial has spent approximately $100 million. I-SPY 2 was originally funded by donations, but, after complete response became an approvable endpoint, companies began paying a fee to cover the cost of their arm.

Since I-SPY 2 began, Bayesian designs for clinical trials have grown in popularity, thus further underlining the success of this approach. Adaptive learning trials have sprung up for pancreatic cancer, glioblastoma, Alzheimer's disease, Ebola, sepsis and Duchenne’s muscular dystrophy (Table 2).

Table 2 | I-SPY 2 graduates

Compound or combination



Months in I‑SPY 2


Nerlynx (neratinib)

Puma Biotechnology




Veliparib + carboplatin









HR, HER2+ and HR HER2+

Perjeta (pertuzumab)




HER2+, HR HER2+, HR+ HER2+

Perjeta + Kadcyla (ado-trastuzumab emtansine)




HER2+, HR HER2+, HR+ HER2+

Keytruda (pembrolizumab)





At least two adaptive learning trials, including Precision Promise, have gone a step further than I-SPY 2 by seamlessly incorporating a phase 3 trial into their design. Manax Rutson said that PanCAN has the backing of the FDA to add a registrational trial stage. “That’s one of the things that can be appealing to pharma: if your drug is performing well, and it does graduate, you have an opportunity to expedite your drug development.” Woodcock thinks that more such trials are on the way. “That’s where the companies are most interested, and patients are very interested in getting all the way to an approved product that benefits the disease,” she says.

Esserman says that she hopes to eventually graduate combinations to phase 3 within the trial as I-SPY continues to innovate. The US National Institutes of Health awarded a P01 grant to Esserman to adapt the therapeutic course within trial participants (that is, patients within a subgroup who share particular characteristics), such that patients within a trial arm may move to sequences of therapies that hold promise for that subgroup. “We’ve had several conversations with FDA about that modification to I SPY 2—they want to do this kind of thing, but they recognize, as we do, that it’s really a difficult route to drug approval,” said Berry.

Berry thinks that running complex multi-arm trials will require a neutral arbiter with no intellectual-property stake in the middle. For I-SPY 2, it is trial sponsor Quantum Leap Healthcare Collaborative, but government support will probably be critical for the proliferation of such adaptive learning trials. According to Woodcock, the FDA works closely in the planning stages for all the platform consortia. Europe’s Innovative Medicines Initiative (IMI) has also stepped to the fore, supporting the European Prevention of Alzheimer’s Dementia Consortium trial starting in 2015.

Bayesian trials were encouraged in the US under the Prescription Drug User Fee Amendments of 2017 and 21st Century Cures Act in 2016, which created the Complex Innovative Trial Designs Pilot Program at the FDA. Woodcock says that the FDA received several applications from companies interested in developing adaptive Bayesian trials since the program was launched last year. And in January, Wave Life Science announced that the FDA approved its application for a planned phase 2/3 trial of suvodirsen in patients with Duchenne’s muscular dystrophy amenable to exon 51 skipping; the trial will leverage historical data to decrease the size of its control group.

As adaptive learning trials take root, Woodcock thinks that I-SPY 2’s influence will continue to be felt. “Because it’s been successful, it’s shown people you can evaluate a lot of biomarkers for prediction and also prognosis. You can learn about multiple things at once, rather than just do a single trial about every single intervention. I think it’s showed the field this can be done.”

doi: 10.1038/d41587-019-00021-8

Nature Briefing

An essential round-up of science news, opinion and analysis, delivered to your inbox every weekday.