DNA sequence classification I have a database of 3190 instances of DNA consisting of 60 sequential DNA nucleotide positions classified according to 3 types: EI, IE, Other.
I want to formulate a supervised classifier.
At the moment I apply the 60 nucleotides as features and in addition 64 features from a 2nd order Markov Transition Matrix.
The nucleotides A,C,G,T are encoded simply as 1,2,3,4
I'm running a Neural Net and achieving about 90% accuracy.
Next thing to try is a SVM.
My Question is: 
1. Is there a better way to encode the feature data?
2. Is there a better way to approach the problem given that the sequential nature of the data is relevant.
 A: I would also suggest looking at recurrent neural networks for their natural application to sequence data. Check out this tensorflow tutorial on machine translation for ideas about network specification: https://www.tensorflow.org/tutorials/seq2seq. Also this nice explanation on Long Shirt Term Memory (LSTM) RNNs: http://colah.github.io/posts/2015-08-Understanding-LSTMs/. I'm not sure there is a "better" way to encode the data, but there are definitely more compressed ways, such as packing multiple binary variables into a byte. There is also chas game representation, and the related universal sequence mapping (USM) which encodes lengths of sequences into pairs of 32 or 64 bit floats or ints (essential just bit packing on the quaternary code), both of which have been used recently for NNs on DNA sequences.
A: Interesting questin. Since extracting relevant features from DNA sequences could be non-trivial it could be a good idea to look  at convolutional neural networks where the network learns which features to extract.
Cool example:
http://www.nature.com/nbt/journal/v33/n8/full/nbt.3300.html
A: I assume all the instances have equal length. This is why you can try neural network or SVM.
As you mentioned, neural network and SVM is not explicitly considering the dependencies on the evens of the sequence.
For sequential model, Markov Chain or Hidden Markov model are widely used and can be used to model variable length sequences.
