“…We employed the encoder structure of the seq2seq model [24] here as the instance feature extractor. The embedding layer [25] was employed to represent bases (15 (A, T, G, C, N, H, B, D, V, R, M, S, W, Y, K) → 4 (representative dimension)) ∘ The encoder used a bi-directional RNN structure, which given equal attention to the head and the tail of the instance, and the output was a context vector [26] to represent the feature of the instance. And subsequently, through the MIL layer, the features of all instances were scored and aggregated jointly to determine the type of the bag [20, 21, 27] .…”