To develop a molecular classifier for cancer diagnostic, we need numbers of samples to train the classifier, many researchers stated that: the sample size is still very small (within 200).
for example, in http://www.broadinstitute.org/mpr/publications/projects/Leukemia/Golub_et_al_1999.pdf , the sample size is smaller than 100.
Questions: are there more samples available now? Or there are still limited samples can be used for training classifiers (due to the issues like cross-lab evaluation)?
You can look for an integration method that can combine the sample thereby allowing you to increase sample size, this will increase the features that you will use to train your classifier. for instance if you a looking at microarray data this method will combine gene expression values regardless of the lab which generated the data
The launch data portal of the cancer genome atlas may be a good starting point:
http://cancergenome.nih.gov/
The sample size for any given study is usually driven by the statistical needs of the study and resources available to it. For instance, research into rare cancers makes it difficult to get one's hands on enough distinct samples from that population. Conversely, if the are large effects on distinctions between the phenotypes that are being studied, the number of samples used may be limited because more aren't needed to show significance.
The Cancer Genome Atlas is definitely a good starting point for getting data for initial studies and hypothesis generation (training classifiers is a form of hypothesis generation), but for testing you may need to either split that dataset, find other datasets from resources like the Gene Expression Omnibus or ArrayExpress, or gather new data yourself.
Ultimately, the researcher is responsible for gathering and evaluating the data they hope to use for their investigation. If they can reuse existing data, that's great, but if it's not there and your purpose is to investigate X, you're going to need to gather your data yourself. I hope you share it.
The short answer is that it's almost certain that there are currently inadequate samples to train a classifier for cancer "in general".
There are two problems at work here:
The first is that historically, cancers haven't been particularly well phenotyped. Numerous cancers that probably have distinct etiology have all been lumped together diagnostically, and from a sample-collection standpoint, labeled the same thing (this isn't unique to cancer, it's pretty much the norm with most diseases). When you have what are perhaps thousands of distinct diseases lumped together into the same data set, you need a lot more than 200 samples to distinguish them.
The second is that the penetrance of any given measurable in terms of causing cancer is probably low. That is, for example, if some particular genotype predisposes a person to having cancer X, that genotype may only contribute 2% to their relative risk of having that cancer. As a result, numerous people without that genotype will still have cancer X, and numerous people with that genotype won't have cancer X. Numerous people with the genotype who don't have cancer X, will, completely independently, have some other cancer Y. As a result, for many of the measurables that you will need to use to distinguish between cancers X, Y, and Z, you're going to need quite a lot of samples to tease out their predictive effects.
How many samples is enough, is really, at this point, anyone's guess. Anyone who tells you that N samples is sufficient, is blowing smoke out of his butt. Once a sufficiently-well-curated collection of samples is generated for unique cancer X, it might turn out that that cancer has completely distinct biomarkers, making it unambiguously identifiable with only a few samples - for some cancers this will almost certainly be the case. On the other hand, many others are likely to be quite complex and require enormous sample sizes to pin down.
- Badges
- Science topic
- Similar topics
- Bioinformatics
- Bioinformatics and Computational Biology