Run @bklein7's consolidated network (db7) in GRNmap and Gephi

kdahlquist commented 6 years ago

In the meeting yesterday, @bklein7 showed a consolidated network that was generated by looking at common edges between db5 and rand1-rand31. @bklein7 should run this new network in GRNmap in two ways:

having the initial guesses for the network weights all = 1
having the initial guesses for the network weights be 1 or -1 based on his predictions as to whether that edge should be activating or repressing

@maggie-oneil will also take this network and run it in Gephi so that we have the corresponding graph statistics.

@maggie-oneil and @bklein7 need to coordinate to use the same input workbooks.

Let's call this new network "db7".

Please save the files in DahlquistLab > data > Fall2017

kdahlquist commented 6 years ago

This is running right now in Seaver 120.

bklein7 commented 6 years ago

The two db7 input sheets, containing either unweighted or weighted initial guesses, can be found here:

The two associated db7 output sheets can be found here:

Note that regardless of the weighting of initial guesses, both runs converged to a nearly identical solution for db7. However, the run starting with weighted initial guesses did require approximately 15,000 fewer iterations to reach this solution. The LSE:minLSE ratio for db7 (1.4471) was higher than that of db5 (1.4263).

bklein7 commented 6 years ago

db7 was derived by identifying common edges between db5 and successful db5-derived random networks. Ultimately, no new edges were introduced to db7. Edges and a single node (GCR2) from db5 were merely excluded when forming db7. Further, 1 edge appearing in three data-base derived networks (STB5->HAP4) did not appear in any db5-derived random networks and was therefore automatically excluded from db7.

This week, a comprehensive edge list was compiled for db1-db6 + the 5 random networks that outperformed db5 (166 edges total). Edges that appeared in two or more of the best performing random networks were checked for validity via the YEASTRACT database. If the regulatory relationship appeared in the database but not in the original six database-derived GRNs, it was nevertheless considered valid for use in new hypothesis networks. Four hypothesis networks were then constructed from this data (https://github.com/kdahlquist/DahlquistLab/blob/master/data/Fall2017/db5-random-networks_common-edges_BK20171106.xlsx):

"db7.2" hypothesis network: added STB5->HAP4, the single edge unrepresented among the db5-derived random networks to db7.
"node-based" db5 hypothesis network: the lowest average MSE for individual nodes was identified from among db5 and the 5 best performing db5-derived random networks. All valid inputs to the node were taken from the GRN exhibiting the lowest average MSE for that particular node. If no valid inputs were available, then inputs were taken from the GRN exhibiting the next lowest average MSE for that node.
"edge-based" db5 hypothesis network: an edge list for this GRN was compiled using three sources. (1) the 11 edges shared between db5 and the 5 best performing db5-derived random networks were listed. (2) the 12/29 valid edges shared between 2+ of the best performing db5-derived random networks were listed. (3) the two edges appearing in db5 as well as other database-derived networks that were not represented among the random networks, STB5->HAP4 & HMO1->HMO1, were included. Together this yielded a GRN with 15 nodes (same as db5) and 25 edges.
"db8", a meta-analysis hypothesis network: Out of the 166 regulatory relationships in the compiled edge list, only 28 edges appeared in 3+ networks were selected. Of these 28 edges, 3 appeared only in the best performing db5-derived random networks, and only 2 were valid per YEASTRACT. The remaining 27 valid edges were used to construct a 15 gene, 27 edge "meta network", named db8. The transcription factors included were near-identical to those in db5, except that ASH1 was replaced by MGA2. Of course, this network will be biased towards genes appearing in db5. Perhaps a similar "meta-analysis" could be included including data from random networks derived from db4, or other database-derived GRNs. Further, criteria other than frequency could be used to select promising edges from the complied list.

All four of these hypothesis networks are currently running in S120.

bklein7 commented 6 years ago

Results write up for hypothesis networks: https://github.com/kdahlquist/DahlquistLab/blob/master/data/Fall2017/hypothesis-network-testing_BK20171115.pptx

kdahlquist commented 6 years ago

@lkelly9 will run the network through Gephi on Tuesday.

kdahlquist commented 6 years ago

Need to double check if we have Gephi file for this.