comparing genome sequences using LSTM recurrent neural network

• format 1: xxxxxx6yyyyyy1
• format 2: xyxyxyxyxyxy1
• format 3: xxxxxxyyyyyy1

• The randomness of the data make significant differences. For fair comparison, the data set must be fixed. ( See format1_leng20_err0_500_100/*.png )

• Comparing the DNA sequence encoding format, format 2 > format 1 > format 3 ( See formatMOD_leng20_err0_500_100/err*png )

• LSTM can classify the sequences, even if the match is not 100% ( See formatMOD_leng20_err0_500_100/err1*png )

• 2-neuron LSTM can deliver quite good performance on format 2 data. However, the performance of 8-nerual LSTM is bad on format 1 data and format 3 data. ( See format1_leng20_err1_500_100__different_RNN/*.png )

• Using the full length data string in format 2, the 16-neural LSTM and 8-neuron LSTM gives good result. ( See formatMOD_leng65_err1_500_100/*.png )