Are a Large Number of Biological Replicates Ever an Acceptable Replacement for Technical Replicates?

Question

[JMallory]

I know that technical replicates are never acceptable replacements for biological replicates. How about the reverse? Are a large number of biological replicates ever an acceptable replacement for technical replicates?

In my own experience in personalized medicine/biomedical big data, there is often a push to say that we have "catalogued over X unique transcriptomes/proteomes/etc. from Population Y (e.g., cancer, autism, multiple sclerosis)". If there is money to sequence more samples, it is a fair bet that the priority is not to run quadruplicates. The attitude of several generalist data scientists/machine learning specialists I have worked with is that, as the data becomes larger, worry about the quality of any single sample becomes less.

Is this an issue that contemporary biostatisticians have an opinion on? To the extent that I am seeing increasing numbers of guys who have more background in building deep ANNs to identify cats than transcriptomics cavalierly dive into this field, this issue worries me a bit. What does this imply for the field?

Answer 1

[Michele Busby]

Yes.

Biological replicates will contain both biological variance and technical variance. Therefore, when you run biological replicates you are often "blocking" on both biological noise and technical noise. Since these are usually uncorrelated, the total variance in the system is additive.

Where this would not be true (in bioinformatics) is where samples are grouped in a way that makes this untrue. For example, if you have 192 samples on two 96 well plates there will be inter-plate variability that may not be well-blocked in this experiment, for example if you put your controls on one plate and your test condition on the other plate. But if you scramble your samples on the plates (and don't drop your plate and mix them up!) you should be properly blocked for both biological and technical noise.

I have a longer blog post about it here as it pertains to RNA seq.

http://michelebusby.tumblr.com/post/26913184737/thinking-about-designing-rna-seq-experiments-to

If you have a proper setup technical replicates are much less valuable than biological replicates because they don't wash out any biological noise. But they are sometimes invaluable as standards, i.e. a technical replicate in each of 20 plates in a huge experiment.