alanadeng / hiv_subtypes Goto Github PK

• HIV-1 is classified into 4 groups: M, N, O, and P, and of these groups, group M is the most widespread and clinically relevant.
• Group M is subdivided into 9 distinct subtypes: A, B, C, D, F, G, H, J, K.
• 2 or more subtypes can combine to form a hybrid form called a circulating recombinant form (CRFs).
• Rates of disease progression vary among HIV-1 subtypes, with some subtypes showing increased drug resistance and virulence.
• Understanding the distribution of HIV-1 subtypes is crucial for the development of vaccines and the clinical management of HIV-1 infections. For infected individuals, classifying HIV-1 subtype is a crucial step in clinical infection management.
• HIV-1 subtyping = classifying HIV-1 into subtypes.

• Most HIV-1 subtype classification methods rely on aligning an input sequence against a set of pre-defined subtype reference sequences. These alignment-based methods are computationally expensive, especially for long sequences and rely on ad hoc parameter settings, which can limit reproducibility.
• To address the limitations of alignment-based methods, alignment-free methods have been developed.
• Few studies (if any) have compared DNA vectorization methods for HIV-1 subtyping.
• Few studies (if any) have applied multi-task learning to HIV-1 subtype classification.

• To characterize the effect of DNA sequence vectorization methods on HIV-1 subtyping.
• Apply multi-task learning to HIV-1 subtype classification.
• Develop an improved HIV-1 subtype classification method.

1. Explore different ways to vectorize DNA sequences:

   • Word2Vec
   • K-mers (of varying length)
   • Subsequence natural vector: includes number and distrubution of nucleotides, position, and variance
   • Natural vector: number and distribtion of nucleotides
   • Other language model encodings

2. Feature selection:

   • LASSO
   • PCA
   • Pre-trained CNN (as Kyle suggested)

3. Machine Learning & Deep Learning

   • Classical Machine Learning
     • SVM (one-vs-rest)
     • Multi-class logistic regression
     • XGBoost
     • LASSO
     • Naive Bayes?
     • KNN clustering
   • Deep Learning
     • 1D-CNN
   • Multi-task learning (time-permitting)

4. Evaluation Metrics

   • Confusion Matrix
   • Accuracy, precision (macro-precision), recall, F1-score (macro F1-score)
   • AUROC, AUPRC
   • Cohen's Kappa

At our meeting, we had discussed evaluating our models with and without applying feature selection.

• Processing data - Kaitlyn
• Vectorize DNA Sequences** - Gen and Kaitlyn
  • Kaitlyn: k-mers, natural (and sub-sequence) vectors
  • Gen: Word2Vec, other natural language encoding methods
• Feature Selection: Alana
• Machine Learning and Deep Learning: Kyle - Done!
• Multi-task Learning - ?

• Background and introduction: Kaitlyn
• Methods
  • Data-processing: Kaitlyn
  • Sequence vectorization: Kaitlyn and Gen
  • Feature Selection: Alana
  • ML/DL: Kyle
• Results and Discussion: Kyle
• Figures: Kyle

• hiv-db.tar.xz: contains 20,386 unprocessed sequences from 289 subtypes LANL HIV Sequence Database
• After processing there are 15,018 sequences and 28 subtypes.
• Each sequence is approx. 10,000 nucleotides.
• hiv.txt: the sequences (atgcgctagatcga)
  • Each line corresponds to one sequence
• labels.txt: the subtype label
  • The first line in labels.txt corresponds to the first sequence (line) in hiv.txt.

• pre-process.py: reads hiv-db.tar.xz
  • Removes sequences with unknown characters (eg: N) and subtypes with too few examples
  • Creates hiv.txt and labels.txt

hiv.txt can be used to vectorize the sequences.

• classify.py: take X and y as inputs, where X is a 2D feature vector and Y is a vector

  • Before runing the codes, please first set configs, including output directory, possible hyper-parameters (optional)
  • Output: Confusion metrices and numerical results, which will be saved in your output directory.

• feature_selection.ipynb: script for dimensionality reduction (PCA, LR-Lasso, 1D-CNN). Specific instructions can be found in the ipython notebook.