FDA Workshop on Pathogen-Specific Antibiotics - Part 2

To continue our discussion of the FDA Workshop on the development of pathogen-specific antibiotics, I’ll start briefly with a discussion of statistical issues and approaches that concluded day 1 of the meeting.  Honestly, since it was a fairly authoritative and intense set of presentations that left me n the dust, I have little to say other than to refer you to the slides. What was clear is that there are a number of Bayesian methods that would allow for a more comprehensive analysis of trials where the numbers are small.  Enough said.

The second day of the workshop was the meat of the meeting and was set-up by our discussion on the first day.  John Rex and colleagues invented a fictional drug they called X-1 that was active only against Pseudomonas aeruginosa.  They then set about trying to design a non-inferiority trial to compare the new drug to a comparator agent – in this case, meropenem.  Why, I asked myself, did he start with a non-inferiority design. In trying to read John’s mind, I thought the following. (1) NI trials are the most reliable way to antibiotic approval and have been used for the approval of new agents (like ceftazidime-avibactam). (2) The exercise would expose in a very quantitative way the costs and risks of such a program. (3) The discussion would stimulate thinking about new approaches to the NI design or to thinking about other designs.  You can find all the details about X-1 in the slide packages provided for the meeting.

John started with a few assumptions and limitations.  The most important was that the trial could not exceed 1000 patients for cost and time reasons. Given the numbers of potentially enrollable patients with these infections, mostly HAP/VAP and IAI, under the usual circumstances of such a trial, you cannot meet the 1000 patient limitation. The proportion of Pseudomonas infections among enrolled patients is just too small. So, to meet this goal required several manipulations.  First, a wide NI margin of 30% is required for HAP/VAP and 25% for IAI.  He was able to justify this in that for HAP/VAP he used the FDA endpoint of 28 day all cause mortality in the microbiologically documented population and for IAI he used cure in the microbiologically documented population as endpoints based on the FDA’s own guidance documents.  The development of agents for patients with unmet needs specifically states that wider NI margins could be considered and the margins chosen lie within the crude M1 calculations provided by the FDA. The design involved the use of X-1 + ertapenem (erta is not active vs. Pseudomonas) vs. meropenem.  Investigators were allowed to add amikacin to each arm for up to 4 days initially (a potential confounder). Even with these wide NI margins, to reduce the trial size, John had to invent a rapid, bedside test for the diagnosis of Pseudomonas in respiratory secretions or abdominal culture swabs to increase the chances of culture-positivity. This led to a total of almost 1900 screened patients in the two indications and 915 enrolled. Then came the results.  Based on John’s impeccable math, 175 patients with Pseudomonas across the two indications were treated divided in a 2:1 randomization schedule. You can see that the numbers are going to be small.  The table shows numbers achieving the endpoint over the total treated for each treatment group and each indication. These fit within the prescribed margins. But what if one or two patients are moved or removed on various sides of this table?  It can all rapidly fall apart.  What if there is no diagnostic test or if the test used fails to predict culture positivity?  The numbers then become really small. In all cases, the conclusions based on the trial, whether positive or negative, are very fragile.

	X-1	Mero
NP	37/48	19/24
IAI	55/69	27/34

 I was convinced from the beginning, and even more so after all of John’s marvelous mathematical scenarios and contortions, that such a trial would be too risky, too expensive and would expose too many patients to an unknown drug that might well still be unknown at the end of the study.  I made the statement that no one would run such a trial.  Jeff Lowtit from the Medicines Company then promptly jumped up and said that he would run such a trial! Nevertheless, it was clear to all in the room what the costs and risks of such a trial would be and what the prospects for funding and for return on investment in such a scenario would be.  It was not a pretty picture. One caveat – could a Bayesian approach substantially de-risk the NI venture?

I suggested that we should abandon the NI trial design concept for such a drug altogether and go towards a superiority design approach as I have argued consistently in the past.  Even here, there are major problems. (1) External controls are probably required as well as some sort of within-trial validation set to show that control levels of response are real. (2) Even with such controls, the trial might not be inferential at the P=.05 level – although possibly it would be for P=.1 or .2.  The trial would probably have to include amikacin for several days in both arms thus failing to avoid this potential confounder.  This puts the trial at higher risk and the addition of amikacin would have to be factored in to the external controls one would use. As I have discussed, such control levels of response could be established pre-trial – so the calculations could be more informed.

Now imagine the situation where infections are even less frequent such as would be the case for Acinetobacter infections.

Paul Ambrose kept coming back to the same theme – and not without justification.  If the trials are not inferential, wouldn’t a strong pharmacometric argument be a strong rationale for approval?  In the case of the externally controlled superiority trial, 30 patients enrolled, receiving the new therapy (X-1 in our hypothetical case), where PK is performed, could provide evidence of response-exposure relationships and these could be placed in the context of the preclinical and phase 1 target attainment data available prior to the start of the “pivotal” trial. I agree with Paul – but would the regulatory authorities consider the more robust PK data substantial evidence in combination with everything else to allow approval? I’m not sure – but I think maybe the FDA is not sure either.

The EMA responded by again pointing out their tools including conditional approval with regular re-review of ongoing data for key products like X-1. Although this tool does not exist for the FDA per se, the FDA could approve a drug, require post-market studies and convene additional advisory committees or take other action based on emerging post-market data.

While the workshop was unable to establish a clear pathway forward for pathogen-specific antibiotics, it was clear that the usual pathways were going to be challenging to say the least.  It was also clear that post-approval data collection was going to be an essential piece of any approval for a pathogen-specific product for a Gram negative pathogen. And all of these considerations are now, thanks to the FDA, in the public domain.