Fig. 2: Inferring the effectome using standard and Bayesian approaches on simulated data.
a, We used the connectome to set the effectome weights for a whole-brain simulation using all 121,327 connected neurons (neurons with no incoming or outgoing connections, above a threshold of >5 synapses, were not included). b, Synaptic weights were set proportional to synapse count, with positive (negative) sign for excitatory (inhibitory) synapses. c, IV estimates of postsynaptic weights of a single example neuron. Most of the error falls along the vertical line where true weight equals 0 (which is the majority of the weights, owing to the sparsity of the fly connectome). d, Mean ± 2 s.d. of an independent Gaussian connectome prior on each weight in the effectome. We set the prior mean to be proportional to the signed synapse count of the connectome and variance equal to the absolute value of the mean plus a small constant, so that the prior width is non-zero between neurons with no known synapses. e, An IV–Bayes estimator shows smaller error than the raw IV estimator in b. f, The error of estimated weights (residual sum of squares, RSS) decreases with the number of samples for both estimators, but IV–Bayes gives several orders of magnitude faster convergence (orange below blue line; mean ± s.d.; n = 10 simulations). Mean squared error will decrease indefinitely for both estimators because they are consistent (that is, they converge to ground truth as the number of samples goes to infinity). Horizontal lines show error level where the R2 of the recovered weights is zero (long dash) and 0.9 (short dash), respectively.
