Fig 1.
Pipeline for analyzing regulatory grammar learned by ResNet models trained on simulated regulatory sequences.
(a) Regulatory sequence and negative sequence simulation. We designed twelve regulatory grammars, including five homotypic clusters, five heterotypic clusters, and two enhanceosomes as prototypes for simulated regulatory sequences. Then, to reflect that regulatory regions active in a cellular context may have multiple grammars, we defined twelve regulatory sequence classes, each with two different grammars. Finally, we generated two sets of negative sequences: k-mer shuffled and TF shuffled versions of the simulated positive sequences. (b) Classification tasks. ResNets are trained on simulated regulatory sequences and the negative sets in three increasingly realistic scenarios. (c) Regulatory grammar reconstruction framework.
Fig 2.
ResNet trained on simulated regulatory sequences and TF-shuffled negatives accurately models the regulatory grammar.
(a) Example saliency map for a simulated regulatory sequence from class 6. Class 6 sequences harbor instances of homotypic cluster 3 and heterotypic cluster 3. The saliency map shown is computed with respect to neuron 1 in the penultimate layer. The red dashed lines show simulated TFBSs in their respective regulatory grammars. (b) The saliency values of the binding sites of each TF in a specific regulatory grammar with respect to neuron 1 in the penultimate layer. (c) Heatmap of the median saliency value of the binding sites of each TF in a specific regulatory grammar (x axis) across neurons of the penultimate layer (y axis). The order of x and y axis labels are determined by hierarchical clustering. The color bars on the side indicate the group label assigned by hierarchical clustering. (d) Actual labels of simulated regulatory grammar of the TFBS overlaid on t-SNE visualization of TFBS saliency values across neurons. Correct prediction of the regulatory grammar for a TF (the predicted label agrees with the actual label) is represented by a dot. Incorrect prediction of the regulatory grammar of a TF is indicated by an “x”. (e) The sensitivity (TP/(TP+FN)) of the regulatory grammar predictions.
Fig 3.
ResNet trained on simulated regulatory sequences against 8-mer shuffled negatives accurately models the regulatory grammar.
(a) The performance of five different ResNet models trained on simulated regulatory sequences against different k-mer shuffled negatives at predicting the regulatory class of the simulated regulatory sequences vs. TFs-shuffled negatives test dataset. (b) Actual labels of simulated regulatory grammar of the TFBS overlaid on t-SNE visualization of TFBS saliency values across neurons. (c) The sensitivity of predicted labels in (b) of the ResNet model trained on the simulated regulatory sequences against 8-mer shuffled negatives.
Fig 4.
Regulatory grammar can be learned by ResNet despite heterogeneity in the regulatory sequences.
(a) Actual labels of simulated regulatory grammar of the TFBS overlaid on t-SNE visualization of TFBS saliency values across neurons. (b) The sensitivity of predicted labels in (a) across regulatory grammars.
Fig 5.
Regulatory grammar can be learned by ResNet when TFBSs are outside of regulatory grammars and there is heterogeneity in the regulatory sequence categories.
(a) Sum of saliency values for TFBSs in each regulatory grammar across neurons in penultimate layer. (b) Actual labels of simulated regulatory grammar of the TFBS overlaid on t-SNE visualization of TFBS saliency values across neurons. (c) Actual labels of simulated regulatory grammar of the TFBS filtered to only those in the top 10% sum of saliency values across neurons in penultimate layer overlaid on the t-SNE visualization. (d) The sensitivity of predicted labels in (c) across regulatory grammars.
Table 1.
Simulated heterogenous regulatory sequence classes with multiple regulatory grammars that can distinguish one class from another.
Fig 6.
The ResNet model fails to learn the correct representation of individual grammars when there are multiple regulatory grammars that can distinguish one heterogenous regulatory class from another.
For this simulation, we created three heterogeneous regulatory sequence classes with no overlap among their grammars (Table 1) and applied our interpretation approach. a) Actual labels of simulated regulatory grammars of the TF binding sites overlaid on t-SNE visualization of TFBS saliency values across neurons. The TFBSs do not separate according to their grammars. b) Sum of saliency values for TFBSs in each regulatory grammar across neurons in the penultimate layer.
Fig 7.
ResNets identify known heart heterotypic cluster when trained on mouse enhancers.
We trained a ResNet on developmental mouse enhancers from 12 tissues identified from histone modifications (S6 Fig) and applied our saliency map approach to interpret the trained network. a) Pipeline for identifying regulatory grammar in mouse developmental heart enhancers. b) t-SNE visualization of clustered TFBS saliency maps from top scoring heart enhancer sequences. Clusters determined by k-means with k = 9 are indicated by color (S6C Fig). Instances of NKX2-5, TBX5, and GATA4 motifs are labelled with shapes. These factors form an essential heterotypic cluster during heart development and are significantly enriched in cluster 4 (S5 Table).