*2.6. Designing FNNs.*

Since an FNN can take epigenomic data as input in the same format as that of TargetFinder (i.e., TargetFinder-format data), we were interested to know if such a simple FNN could outperform TargetFinder or other CNNs. After tuning the model parameters through the validation dataset, we chose a 2-layer FNN shown in Figure 4.

An FNN can take both TargetFinder-format data and CNN-format data as input, and our analysis showed an FNN with the same structure performed better with the TargetFinder-format data than that with the CNN-format data (Table S3). Therefore, we further explored the performance of basic CNNs shown in Figure 2 and FNNs shown in Figure 4 in comparison to gradient boosting as implemented in TargetFinder [11]. We not only implemented neural networks for K562 cell line, but also expanded our evaluations to other cell lines (GM12878, HeLa-S3, IMR90). The training configuration was similar to that for CNNs.

**Figure 4.** TargetFinder-format (epigenomic) data and the feed-forward neural network (FNN) model structure.
