Fig. 5.

Comparison of different transfer learning strategies. (a) Detailed overview of the training scheme and prediction generation procedure using three transfer learning strategies implemented by NetTIME, Catchitt and AgentBind. $[t,c]$ and $[p,q]$ denote the particular TF (t or p) and cell type (c or q) combinations. S refers to the set of all conditions in our training dataset. T is the set of 10 leave-out conditions. C_p is the set of cell types that satisfy $\forall r\in C_p,[p,r]\in S$. $x_{[t,c]}$ is the input data from $[t,c]$. $Z_{[p,q]}$ is the per-base-pair binding probability predictions for $[p,q]$. F and M can be any machine learning models. To avoid performance differences introduced by the model architecture, we use NetTIME model with TF and cell-type embeddings (F) and with no embedding (M). (b) Transfer predictions using different transfer learning strategies are generated for 10 leave-out TF and cell-type combinations within the training panels of TFs and cell types