Table 2. Evaluation of the DirectInt model with various cooperativity parameters.

Model	# Pars	Avg. CC	#(CC>0.65)	CVCC (STDEV)
No Coop	13	0.547	16	0.400 (0.02)
Neg Ctrl No Coop	13	0.211±0.076	7.76±1.6	0.02±0.083
Bcd Coop	14	0.577	22	0.428 (0.01)
Cad Coop	14	0.553	21	0.428 (0.02)
Gt Coop	14	0.557	22	0.428 (0.03)
Hb Coop	14	0.552	20	0.328* (0.02)
Kni Coop	14	0.565	20	0.458 (0.02)
Kr Coop	14	0.550	16	0.441 (0.02)
All TF Coop	19	0.603	25	0.418 (0.03)
Bcd & Kni Coop	15	0.587	24	0.460 (0.02)
Neg Ctrl Bcd & Kni Coop	15	0.214±0.08	8.04±1.86	0.027±0.077

The models examined include those without self-cooperative DNA binding (“No Coop”), with cooperative binding by one of six different TFs (“Bcd Coop”, “Cad Coop”, etc.), with cooperative binding by all six TFs (“All TF Coop”), and with cooperative binding by Bcd as well as Kni (“Bcd&Kni Coop”). For each model, the number of free parameters used is shown (“# Pars”), along with average correlation coefficient (“Avg. CC”) between model prediction and true readout over all 37 CRMs in the data set, the number of CRMs where the average CC was above 0.65 (“#(CC>0.65)”), and the average correlation coefficient under a 10-fold cross-validation scheme (“CVCC”). Also shown are evaluation results for negative controls (“Neg Ctrl”) corresponding to the “No Coop” model and the “Bcd&Kni Coop” model. A negative control involved re-training a model with randomly permuted PWMs; shown are the average and standard deviation (of each evaluation metric) over 25 independent replicates of such a control.

*Note that the CVCC values depend upon how the data set got partitioned in the cross-validation exercise. The values in the last column are from one such partition (same across all rows). The Hb Coop model shows lower CVCC than the “No Coop” model in this partition. CVCC values from 6 additional partitions are shown in Table S2, and the Hb Coop model performs better than the No Coop model in five of those six cross-validation exercises.