Extended Data Fig. 4: Track error rates.
a) The predicted likelihood (after marginalization) of links in the tracking solution against the measured likelihood of being correct based on the manual data (top panel). The line does not overlap the dotted line, because not all graph information available to the flow-solver can be integrated in the error prediction. Key is that they remain above the dotted line so that the error predictions are conservative. The histogram (bottom panel) shows the distribution of likelihoods of correct (green) and incorrect links (red). Because the linking error rate is so low, incorrect instances cannot be seen in the histograms a) and b). The black vertical line indicates the threshold of 99% chance of being correct. b) The likelihood of a pair of links both being correct can be calculated by combining their constituent probabilities by simple multiplication. It does not matter if the links are subsequent (blue line) or unconnected (gray line). It is thus not so that a link being true is informative of the truth of the subsequent link, beyond its predicted error probability. c) The predicted error rates for tracks of arbitrary length. The probability that a cell can be correctly traced back to its last division (red dotted lines) is predicted for every cell at every timepoint. The probability of the track being correct is calculated by multiplying all the constituent probabilities. The tracks are compared to the ground truth and deemed correct if they recapitulate it exactly, yielding a similar calibration curve to a). d) The same as in c) but now only with tracks that span the full cell cycle, again producing a similar calibration curve as in a). Error detection works efficiently for full tracks spanning the complete cell cycle, only one incorrect track is above the 99% certainty cut-off.
