BlanchetteLab / PhyloPGM

PhyloPGM

PhyloPGM

Requirements

pip install -r requirements.txt

TFBS Prediction Problem

The goal is to improve the prediction accuracy of a previously trained TFBS predictor on human using orthologous data. Due to vast amount of data size, we do not provide the prediction scores from our TFBS prediction experiments in this repo.

Data Preparation

Please note that the data used in our experiments for the TFBS prediction problem is huge and is not possible to provide the processed data in our github repo. Please follow the instructions below to prepare the data that we used in our TFBS prediction experiments.

1. Download data for "within-cell type" problem of ENCODE-DREAM in vivo Transcription Factor Binding Site Prediction Challenge https://www.synapse.org/#!Synapse:syn6131484/wiki/402026

2. There should be 13 ChIP-Seq data consisting of 12 TFs and 3 cell types.

3. In train set, sub-sample the number of non-binding sites as the same number of binding sites.

4. Keep 20% of train set as validation set.

5. The input examples in train, validation and test sets are 200 bps, kindly extend each example from both sides to make an input example of 1000 bps.

6. Train FactorNet on the training set and use valdiation set to avoid overfitting.

7. Now, we need to compute orthologous regions to make predictions on orthologous regions using the trained base models

8. The 13 ChIP-Seq data from ENCODE-DREAM competition is based on hg19 human reference, so first convert them to hg38 human reference using liftOver tool

9. Use mafsInRegion program (https://hgdownload.soe.ucsc.edu/admin/exe/linux.x86_64/mafsInRegion) to extract orthologous regions from a 100-way vertebrate whole-genome alignment available from the UCSC Genome browser (Kent et. al. 2002) and computationally predicted ancestral sequences produced by Ancestor1.0 (Diallo et. al. 2009)

10. Ignore orthologous regions whose length is less than 700 bps

11. Symmetrically trim or extend orthologous regions to make input example of 1000 bp.

Train FactorNet

We can use the script train_factornet.py to train a FactorNet (Quang and Xie, 2019) model. Below is an example to run the script

python FactorNet/train_factornet.py toy-data/chipseq-ortho-test.csv toy-data/chipseq-ortho-test.csv toy-data/chipseq-ortho-test.csv 2 10 toy-data/model-trial.pth

FactorNet Predictions

The trained FactorNet models can be used to predict on orthologous regions. We need to predict on the orthologous regions of the validation and test sets, which are later used in the PhyloPGM pipeline.

Below is an example of predicting on the orthologous regions of the validation set

python FactorNet/predict_factornet.py toy-data/chipseq-ortho-val.csv toy-data/model-human-FOXA1.pth toy-data/pred-chipseq-ortho-val.csv

Below is an example of predicting on the orthologous regions of the test set

python FactorNet/predict_factornet.py toy-data/chipseq-ortho-test.csv toy-data/model-human-FOXA1.pth toy-data/pred-chipseq-ortho-test.csv

Compute PhyloPGM Scores

We need to prepare input for the PhyloPGM computations.

Below is an example to prepare input for PhyloPGM from the predictions made on the orthologous regions of the validation set

python PhyloPGM/create_PhyloPGM_input.py toy-data/pred-chipseq-ortho-val.csv toy-data/df-pred-chipseq-ortho-val.csv PhyloPGM/tree.pkl 1000

Below is an example to prepare input for PhyloPGM from the predictions made on the orthologous regions of the test set

python PhyloPGM/create_PhyloPGM_input.py toy-data/pred-chipseq-ortho-test.csv toy-data/df-pred-chipseq-ortho-test.csv PhyloPGM/tree.pkl 1000

Now, we can run PhyloPGM script to compute the final PhyloPGM predictions. Below is an example of running PhyloPGM script

python PhyloPGM/run_PhyloPGM.py toy-data/df-pred-chipseq-ortho-val.csv toy-data/df-pred-chipseq-ortho-test.csv PhyloPGM/tree.pkl toy-data/df-pgm-100

The above command computes PhyloPGM scores for each human example in the toy-data/df-pred-chipseq-ortho-test.csv dataset.

Compute PhyloStackNN Scores

PhyloPGM combines the orthologous prediction scores to improve the prediction accuracy on human. Alternatively, We can stack a multilayer perceptron on the orthologous prediction scores to obtain a combined prediction. Below is an example of training such a model.

python PhyloStackNN/run_phylostacknn.py --validate-data toy-data/df-pred-chipseq-ortho-val.csv --test-data toy-data/df-pred-chipseq-ortho-test.csv --output-fname toy-data/df-chipseq-stack.csv

The above command trains a multilayer perceptron on the orthologous prediction scores in toy-data/df-pred-chipseq-ortho-val.csv dataset and the trained model combines the prediction scores in toy-data/df-pred-chipseq-ortho-test.csv dataset.

RNA-RBP Binding Prediction Problem

The goal is to improve the prediction accuracy of a previously trained RNA-RBP binding predictor on human using orthologous data. We provide the trained base models models.rnatracker.tar.gz and PhyloPGM prediction scores rnatracker-phylopgm-predictions.tar.gz in this repo.

Data Preparation

Please note that the data used in our RNA-RBP binding prediction experiments is huge. Please follow the instructions below to prepare the data

1. We use the CLIP-seq data of 31 RNA-RBP binding experiments curated by Strazar et. al. 2016. https://github.com/mstrazar/iONMF/tree/master_full/datasets/clip
2. In order to retrieve orthologous regions, we first convert sequences from hg19 human reference genome to hg38 human reference using liftOver.
3. Use mafsInRegion program (https://hgdownload.soe.ucsc.edu/admin/exe/linux.x86_64/mafsInRegion) to extract orthologous regions from a 100-way vertebrate whole-genome alignment available from the UCSC Genome browser (Kent et. al. 2002) and computationally predicted ancestral sequences produced by Ancestor1.0 (Diallo et. al. 2009)
4. Ignore orthologous regions whose length is less than 70 bps.
5. Symmetrically trim or extend orthologous regions to make input example of 101 bp.

Train RNATracker

We can train a RNATracker (Yan et al. 2019) model using the script train_rnatracker.py. Below is an example to train a RNATracker model

time python RNATracker/train_rnatracker.py toy-data/ortho-val-100 toy-data/toy_model.pth

RNATracker Predictions

We can predict on orthologous regions using a trained RNATracker model. Below are two examples that will be later used in PhyloPGM pipeline

python RNATracker/predict_rnatracker.py toy-data/ortho-val-100 toy-data/base_model.pth toy-data/pred-ortho-val-100

python RNATracker/predict_rnatracker.py toy-data/ortho-test-100 toy-data/base_model.pth toy-data/pred-ortho-test-100

Compute PhyloPGM Scores

We need to prepare input for the PhyloPGM computation using the predicted scores on the orthologous regions. Below are two examples,

python PhyloPGM/create_PhyloPGM_input.py toy-data/pred-ortho-val-100 toy-data/df-pred-ortho-val-100 PhyloPGM/tree.pkl 1000

python PhyloPGM/create_PhyloPGM_input.py toy-data/pred-ortho-test-100 toy-data/df-pred-ortho-test-100 PhyloPGM/tree.pkl 1000

Now, we can run PhyloPGM on the inputs to obtain the final PhyloPGM scores. Below is an example

python PhyloPGM/run_PhyloPGM.py toy-data/df-pred-ortho-val-100 toy-data/df-pred-ortho-test-100 PhyloPGM/tree.pkl toy-data/df-clipseq-pgm-100

The above command computes PhyloPGM scores for each human example in the toy-data/df-pred-ortho-test-100 dataset.

Compute PhyloStackNN Scores

Alternatively, we can combine the orthologous prediction scores using a multilayer perceptron. Below is an example,

python PhyloStackNN/run_phylostacknn.py --validate-data toy-data/df-pred-ortho-val-100 --test-data toy-data/df-pred-ortho-test-100 --output-fname toy-data/df-clipseq-stack.csv

The above command trains a multilayer perceptron on the orthologous prediction scores in toy-data/df-pred-ortho-val-100 dataset and the trained model combines the prediction scores in toy-data/df-pred-ortho-test-100 dataset.

Putative Functional Sites Analysis

We consider the TF/RBP binding sites as putativelu functional if they overlap with

1. the non-coding portion of the ClinVar database (Landrum et al., 2016), which human mutations associated to diseases
2. the non-coding human variants linked to phenotypic consequences through several publications (Biggs et al., 2020)
3. the list of deleterious non-coding variants identified through machine learning and other computational techniques (Wells et al., 2019)

The above three data sets are combined in nc-clinVar.bed.

References

1. Quang, D. and Xie, X. (2019). Factornet: a deep learning framework for predicting cell type specific transcription factor binding from nucleotide-resolution sequential data. Methods, 166, 40–47.
2. Yan, Z. C., Lecuyer, E., and Blanchette, M. (2019). Prediction of mrna subcellular localization using deep recurrent neural networks. Bioinformatics, 35(14), I333–I342.
3. Landrum, M. J., Lee, J. M., Benson, M., Brown, G., Chao, C., Chitipiralla, S., Gu, B., Hart, J., Hoffman, D., Hoover, J., et al. (2016). Clinvar: public archive of interpretations of clinically relevant variants. Nucleic acids research, 44(D1), D862–D868.
4. Biggs, H., Parthasarathy, P., Gavryushkina, A., and Gardner, P. P. (2020). ncvardb: a manually curated database for pathogenic non-coding variants and benign controls. Database, 2020.
5. Wells, A., Heckerman, D., Torkamani, A., Yin, L., Sebat, J., Ren, B., Telenti, A., and di Iulio, J. (2019). Ranking of non-coding pathogenic variants and putative essential regions of the human genome. Nature communications, 10(1), 1–9.
6. Kent, W. J., Sugnet, C. W., Furey, T. S., Roskin, K. M., Pringle, T. H., Zahler, A. M., and Haussler, D. (2002). The human genome browser at ucsc. Genome research, 12(6), 996–1006.
7. Diallo, A. B., Makarenkov, V., and Blanchette, M. (2009). Ancestors 1.0: a web server for ancestral sequence reconstruction. Bioinformatics, 26(1), 130–131.
8. Strazar, M., Zitnik, M., Zupan, B., Ule, J., and Curk, T. (2016). Orthogonal matrix factorization enables integrative analysis of multiple RNA binding proteins. Bioinformatics, 32(10), 1527–1535.