Fig. 4: Adaptation of previously trained networks to a novel smFRET system with Transfer Learning by AutoSiM.

a Schematic depiction of TL. Networks previously trained on a large database of smFRET traces are adapted to a new system by retraining only the final layers of each network using a small training set from the new experimental system. b Schematic of a previously characterized Mn²⁺ riboswitch system⁴² that is used here as a test case for TL, since data from this system were not present in the original training set for the LSTM networks. c Three representative time traces from the Mn²⁺ riboswitch system illustrating its diverse molecular behaviors: static low-FRET (top), static high-FRET (middle), and dynamic (bottom). d Representative traces illustrating True Positive, False Negative, and False Positive classification and segmentation results from TL. e FRET histograms for the Mn²⁺ riboswitch following classification and segmentation by TL or manual selection, with two-peak Gaussian fits and corresponding population estimates for the low- and high-FRET states. The number of traces (N) included is shown in the upper right corner of each histogram.