Machine Learning with Kernel Methods (MVA 2021) - DNA sequence classification from scratch using kernel methods and ML algorithm
Transcription factors (TFs) are regulatory proteins that bind specific sequence motifs in the genome to activate or repress transcription of target genes. Genome-wide protein-DNA binding maps can be profiled using some experimental techniques and thus all genomics can be classified into two classes for a TF of interest: bound or unbound. In this challenge, we will work with three datasets corresponding to three different TFs, and predict whether or not a sequence bound with a TF or not.
Exemple: ATGCCGATCG -> 0 TGCCAGTGCA -> 0 TGCCAATCGA -> 1
To generate a submission file, please start the script: svm_generate_predictions.py. Feel free to try different parameters by modifying the "parameters" section in the script. It is provided here with the parameters that gave our best submission.
Note that it may take a long time to compute the sum kernel of our best submission (several hours). If you want to quickly have correct baselines, you can change to a mismatch kernel with k=10 and m=1 (very fast), or a mismatch kernel with k=12 and m=2 (around 15 minutes to compute the kernels).
The file kernels.py contains all the kernels that have been implemented (Linear, Polynomial, Gaussian, Spectrum, Mismatch, Sum, etc.).
In the folder classifiers you will find 3 types of classifiers: Kernel Logistic Regression, Kernel Ridge Regression used as classifiers, and Kernel SVM.
The mismatch kernel use an auxiliary dict to store the kmers and their neighbors (up to m mismatches), when run for the first time for given values of k and m, the mismatch kernel creates the dictionnary and stores it as a neighbors_dataset_k_m.p file. Some of the auxiliary dict can be downloaded from: https://drive.google.com/drive/folders/1EwTum27-v8h0Ir6PBHpZaCqHmAFnYKUH?usp=sharing