Data analysis

We first rescale the recorded frequency signals from each cantilever by applying a rough correction for the different sensitivities of the cantilevers. Cantilevers differing in only their lengths should have mass sensitivities proportional to their resonant frequencies to the power three-halves. Therefore we initially divide each frequency signal by its carrier frequency to the power three-halves such that the signals are of similar magnitude. To detect peaks, we first filter the data with a third order Savitzky-Golay lowpass filter³⁹, followed by a nonlinear highpass filter (subtracting the results of a moving quantile filter from the data). We find peak locations as local minima that occur below a user-defined threshold. After finding the peak locations, we estimate the peak heights by fitting the surrounding baseline signal (to account for a possible slope in the baseline that was not rejected by the highpass filter), fitting the region surrounding the local minima with a fourth-order polynomial, and finding the maximum difference between the predicted baseline and the local minima polynomial fit. Because this process occasionally makes errors (it sometimes detects noise, particles that got stuck in the cantilever, and particles that passed through the cantilever at the same time as another particle), for each cantilever we reject peaks that are very unlike the typical peak. We do this by first calculating the robust Mahalanobis distance for each peak in terms of a number of its estimated characteristics (baseline slope, time between sequential antinode peaks, minimum value between sequential peaks, and difference in heights of sequential peaks), and rejecting those with large distances above a user-specified threshold. We then identify the peaks corresponding to the calibration particles, and precisely estimate the mass sensitivity for each cantilever, such that the modal mass for the particles is equal to the expected modal mass according to the manufacturer’s datasheet. Finally we match up peaks at different cantilevers that originate from the same cell to extract single-cell growth information (more details in Supplementary Note 3). Figs. 2–5 show the mass accumulation rates of automatically-matched cells that were observed at least seven times. For E. coli (Fig. 6), because the cell concentrations were higher, we required both that the cell have been measured at least six out of ten times and that the root-mean-square-error of a linear fit (buoyant mass vs time) was less than 5 fg, as a greater fitting error suggests a possibly-incorrect matching.