It is always fun when the research you are working on is considered one of ultimate importance.
This is how a recent Scientific American article on Liquid Biopsy ended:
A high rate of false positives or negatives or a tendency to detect cancers that are slow growing or trivial will not be useful. These companies have to prove that they can detect early cancer and, more important, that the early detection can have an impact on cancer survival. That is the holy grail of the holy grail.
This brings me to our (with Kyra Gan, Su Jia and Andrew Li) recent paper:
From its introduction:
Among the most important open problems in cancer research today is the development of an effective approach for the detection of cancer, particularly at its earliest stages. Unfortunately, although monitoring certain “warning signs” occasionally yields early diagnoses, cancer screening is in general notoriously difficult, and existing approaches fall short.
Because of these difficulties, there has always existed a dream within the medical community of developing a liquid biopsy, i.e. a blood test for cancer. This test would naturally be minimally invasive, and ideally would have the same accuracy as a traditional biopsy. This paper addresses a set of active learning problems that occur in the development of liquid biopsies.
The main idea: if an individual has a tumor, some portion of their cell-free DNA will contain mutations which signal the existence of that tumor. So, performing the liquid biopsy simply involves extracting cell-free DNA (a relatively easy task), and sequencing it in search of these mutations. There is no purely biological reason this approach should fail.
Instead, the constraint faced today is cost – human DNA consists of three billion addresses, but the cost of DNA sequencing means that any reasonably-priced test can only include approximately 10 thousand of those addresses.So the challenge is to design a liquid biopsy using just a panel of 10 thousand pre-identified addresses.
Cutting through the chase, from our concluding paragraph (scroll right on top of the post to see numerical results):
To illustrate the applicability of our proposed method to the liquid biopsy problem, we conducted numerical studies on the COSMIC dataset. We found that our algorithms outperform the existing state-of-art benchmarks by large margins. Furthermore, our algorithms consider the priors for having different cancer types explicitly when constructing the action sequences, yielding superior performances compared to under non-uniform priors.
More can be done. Indeed, we have embarked on testing out a couple of new ideas that will further speed up and increase the accuracy of liquid biopsy. Those interested in having us run our algorithms on your data, you know where to reach us – at the Mecca of Machine Learning — Carnegie Mellon University!