CANCER COMMONS: BIOMEDICINE IN THE INTERNET AGE - Collaborative Computational Technologies for Biomedical Research - page 171

this is much more easily said than done. The profi ling step is relatively simple
in these days of microarrays that measure everything from gene expression to
chromosomal rearrangements, but choosing a treatment based upon such pro-
fi les has not even begun to become practical and is unlikely to progress very
rapidly because of the high dimensionality of the problem.
The Cancer Commons approach to this problem combines N - of - 1 treat-
ment (as defi ned above in cancer) with adaptive trials in a particular way that
begins to address the high dimensionality of the problem. We call this approach
N - of - 1 adaptive trials . The idea is to genomically profi le (to the degree possi-
ble) every patient and then treat each of them with the best possible treatment.
Of course, it is often not possible to determine the best possible treatment,
either because there are multiple treatment options available that have not
been tested head to head, or because no effective treatment is currently
known. The latter case is dealt with by personalized virtual biotech, described
above. Here we focus on the former case, where there are a range of potential
treatments. In this case, we would like to essentially run an adaptive experi-
ment over a set of patients, trying multiple treatments, collecting and analyzing
the response data as effi ciently as possible, and then integrating the results into
the molecular disease model to tune the treatment regimens for these patients
(if possible) and for subsequent patients. It is important to note that, on an
individual basis and leaving the genomics out of the picture, treating a particu-
lar patient, seeing how things go, and adjusting the treatment accordingly are
what physicians do all the time and have done through time immemorial. What
is new here is effi ciently collecting the results of a large number of such N - of - 1
experiments and using them in a rapid learning loop to adjust the treatment
regimens for the patients being treated and for subsequent patients.
As plausible as this might seem, the N - of - 1 adaptive approach holds hidden
dangers [7] and must be carefully thought out to avoid inappropriate chan-
nelization into a less effective treatment based on an apparent early success
due, say, to a misclassifi cation of the disease or perhaps noisy data [8]. To take
an extreme case, say that, when one drug appeared slightly better in one patient
all other patients were immediately switched to it, it would be the end of the
trial. One should, in theory, wait until one has statistically strong evidence
before switching everyone onto another treatment, but because there is a large
statistical price to pay for our “peeking” at the data at every step, it may well
be that such a trial requires many more subjects than a classical trial. Moreover,
if we continue treating everyone with one of the two treatments at random
(more precisely, in accord with randomized trial protocols), then if it requires
more patients to see an effect in the adaptive trial, we may have harmed the
excess number of patients assigned to the poorer treatment over the classical
trial by virtue of not having treated them with the best available treatment as
soon as would have been possible in the classical model. Of course, in powering
the study to begin with, one can tell how many subjects one is likely to need,
but the adaptive approach will always require, in the worst case, more subjects
than a classical trial. (“Whopping” effects in either direction are, of course, also