Tyrosine recombination classification and mining for Recombinase-in-Trio elements

BINF6999 Thesis Project

Documentation in progress!

---

Contents

1. Project Objectives
2. Part 1: Classifier for tyrosine recombinases
   • Overview
   • Workflow
3. Part 2: Pipeline for identification of RIT elements
   • Overview
   • Workflow

---

Project Objectives

**1 **. Create an amino acid sequenece classifier to assign site-specific tyrosine recombinases (aka. phage integrase) proteins to the 20 subfamilies described by Smyshlaev et al. (2021), or 'Other' if the protein is not a tyrosine recombinase.

2. Create a pipeline for the identification of Recombinase-in-Trio (RIT) mobile elements from GenBank files, using the tyrosine recombinase classifier to identify RitA, RitB, and RitC proteins.

3. Classify tyrosine recombinases in the ICEberg2.0 database. to sequenced genomes and mobile element databases.

---

Part 1: Classifier for tyrosine recombinases (YRs)

Overview

• Datasets

  • SMART tyrosine recombinases reference data & other non-integrase sequences as negative examples.
  • These are combined, downsampled, and split to create a training/validation dataset for classifier development and a holdout dataset for final testing.

• Modelling Pipeline

  • For a given data split (eg. training/validation samples):
  • Create 20 alignments of YR catalytic domains, one for each subfamily.
  • Create 20 HMMs from the alignments, one for each subfamily
  • Score train and test seqs against HMMs using HMMER's hmmsearch
  • Normalize scores (using training data only to create normalizer)
  • Use SMOTE upsampling to create synthetic examples, bringing minority classes up to 25% of the largest class.
  • Fit classifiers to the prepared training sample.
  • Apply classifiers to validation sample.
  • Evaluate predictions and return performance metrics.

• 3-fold Nested-CV (repeated 3 times)

  • Apply the above pipeline, with a variety of models & hyperparameter settings, in nested CV.
  • Get an estimate of performance for the model-fitting procedure (not actually selecting the model hyperparameters at this stage).

• 3-fold CV (rep 3 times)

  • For the selection of models with best mean MCC across.
  • Best models are kept for the final holdout test.

• Holdout test

  • Fit 'best' models from CV using pipeline, with all training data.
  • Get point estimates of performance from prediction of holdout data.

• Fit final models

  • Align all catalytic domains for each subfamily, build 20 HMMs, score all sequences by hmmsearch, normalize scores, upsampling by SMOTE, fit classifiers.

• Produce visualizations

  • To show evaluation of classifiers in nested CV, CV, and final test.

---

Workflow

1. Dataset acquisition:

SMART database

• The script 1a_tidy_smart_data.R:
• Organizes data from the set of fasta files downloaded from the SMART database.
• Creates a dataframe with the names and sequences for tyrosine recombinase proteins and their catalytic domains.
• The total number of sequences is ~120k, but with large class imbalance, e.g., Xer has ~34k members and Int_Des has only 62.
• Problematically, some sequences are annotated with >1 subfamily, for the same portion of sequence (ambiguous classifications). These double-labelled sequences are removed, leaving 114,848 sequences.
• The dataset for modelling was downsampled to a max of 10k sequences per subfamily, with the removed sequences set aside for further classifier testing.
• The cleaned data was saved as smart_df.rds

Non-integrases

• The script 1b_get_refseq_non_integrases.R sent keyword searches to Entrez API to get a variety of non-YR protein sequences to serve as negative examples.

• I also downloaded:

  • Several Pfam families of transposases (domain sequences only)
  • Uniprot proteomes of yeast, Arabidopsis, and human proteomes.

• Any large groups were down-sampled to 2000 sequences.

• A total of 39,691 sequences were retained after downsampling.

• The script 1c_tidy_non_integrase.R cleans up data gathered by the steps above to create the dataset nonint_df.rds

2. Searching for best models

The script 2a_data_splitting.R:

• Combines the SMART integrases dataset and the non-integrase dataset.
• Splits data 75:25 for training/validation (building & tuning classifiers) and testing
• Creates 2 files in ./data/:
  • classif_train_set.rds (training 75 %)
  • classif_test_set.rds (testing 25 %)

The script 2b_nested_CV.R:

• Performs 3 x 3-fold nested CV
• Uses functions from the script 00_functions.R for the model fitting steps
• Uses the set of model specifications created in the script 00_get_model_specs.R (models: unfitted_parsnip_model_set.rds)
• Results are saved to results/3x3-fold_07-08_nest_cv_summary.rds

The script 2c_regular_CV.R

3. Final validation and model fitting

• 3a_final_test_set.R - fits best models from CV using the full training data, and evaluates predictions of the test set
• 3b_fit_final_model.R - fits the best model of each type (Elastic Net, Random Forest, k-Nearest Neighbors) using the full, down-sampled dataset

Visualizations

4_Classifier_results_plots.R generates the figures for the manuscript and presentation

---

Part 2: A pipeline to identify new RIT elements

Overview

• Obtain data from NCBI by linking CDD ->> proteins ->> nucleotides -> taxonomy
• Translate nucleotide sequences -> proteins with locations
• Classify protein sequences
• Figure out which nucleotides have RIT arrangement

Workflow

---