Fig. 4

Application of ExplaiNN in predicting chromatin accessibility in the mouse immune system. A Performances (average PCC; y-axis; green) and number of well-predicted sequences (secondary y-axis; gray) for ExplaiNN models (solid lines) with increasing numbers of units (x-axis) and AI-TAC (dashed lines) trained on OCRs in 81 immune cell types from 6 different lineages [9]. B Pairwise comparison of the individual performances (PCC) of the OCRs from the previous dataset between the ExplaiNN model trained using 300 units (y-axis) and AI-TAC (x-axis). The Pearson correlation coefficient (R) of the individual OCR performances between the two methods is shown at the lower right corner. C Heatmap of the final linear layer weights of the ExplaiNN model trained using 300 units, with columns representing units with assigned biological annotations based on their Tomtom [27] similarity to known TF profiles from the JASPAR database [18] and rows representing the 81 immune cell types colored by lineage: stem cells (navy blue), B cells (turquoise), alpha/beta (forest green) and gamma/delta T cells (olive green), innate lymphoid cells (pink), and myeloid cells (purple). The logos derived from visualizing the filters of selected lineage-specific units are shown at the right. D, E Visualizations of importance scores colored by lineage for two units from the previous model annotated as CEBP and PAX, revealing their importance to the prediction of chromatin accessibilities in myeloid and B cell lineages, respectively. The logos of the filters of these units are shown at the top. OCR, open chromatin region; PCC, Pearson correlation coefficient; TF, transcription factor