Fig. 4

Module substitution shows that Aggregator learns similar SRF activities when trained with FMs or AMs. In this experiment, we demonstrate that the AM model still produces motif sites that carry the same semantic meaning as the original FM sites. To this end, we take the AM motifs and send them to the FM aggregator. If the AM model were producing non-motif information, this would lead to a degradation or at best no change in performance. However, we observed an improvement, from ~68 to ~71% accuracy, demonstrating that the AM models are better at predicting motif binding sites, even provided to a model trained to work with FM predictions. a–c Stages of the module substitution process. The FM Motif Model and Aggregator are in blue, while those corresponding to the AM model are in red. a Both models are trained as usual. b The models are binarized by applying a binary layer and then the FM model’s Aggregator is retrained to handle the new binary motif outputs. c The new FM Aggregator is used in conjunction with the binarized AM motif model, with sparse layers and binarization layers as shown. d Results of the combination experiment