mrjiaruiwang
commented
9 years ago Well transcriptional profile of a cell is tricky. Single-cell data is often noisy and replication errors are usually too high for clear results. Using multiple cell data is less problematic if you are willing to sacrifice resolution and look at clusters of cells for nodes instead of single cells, but at least the data ill be cleaner. Or we can just wait a few more years for single-cell data to improve.
One question that's come up in lecture is what are the 'fundamental'/'objective' properties of a graph node or edge that we can use to define and construct a brain graph. Do we care about morphology, neurotransmitter, location, NT receptors, etc etc? One metric that takes much of this into account is the transcriptional profile of a cell - the pattern of which genes (technically mRNAs) are expressed in a neuron. When I read about exchangeable graph models it seemed like they would be a good way to incorporate this information. In this model each node has a K-bit binary string associated with it which defines it and sets the probability of connections to other nodes. Could a transcriptome (or condensed version of one) be used as the K-bit string where each bit is whether a certain gene is on or off in that node (and each node is a neuron)? If uncondensed it would be 15-25K bits long. Am I understanding exchangeable graph models the right way? Would this approach be feasible / desirable?