jiantaoshi/tramhap

TRAmHap is a novel deep-learning framework, which predicts transcriptional activity using characteristics of DNA methylation haplotypes in proximal promoters and enhancers as far as 25 kb away from TSS.

TRAmHap

TRAmHap is a novel deep-learning framework, which predicts transcriptional activity using characteristics of DNA methylation haplotypes in proximal promoters and enhancers as far as 25 kb away from TSS.

Prerequisites

Python 3.8.8 or greater with the following packages:

1. pytorch v1.9.1+cu102
2. numpy v1.21.5
3. pandas v1.3.5
4. matplotlib v3.5.1

Workflow

Data preprocessing

Datasets

To demonstrate the usage of TRAmHap, we have provided a mouse heart dataset, which can be downloaded from Test DataSets. It contains the following heart tissue samples:

Tissue	Name	Description	GEO Accession
heart	ENCBS366XOW	Heart from 69 embryonic 10.5 day mice	GSM2701082
heart	ENCBS490ZWZ	Heart from embryonic 13.5 day mice	GSM2191246
heart	ENCBS004ZLN	Heart, Mus musculus	GSM2191412
heart	ENCBS909LOF	Heart from 62 embryonic 15.5 day mice	GSM2191436
heart	ENCBS955JPC	Heart pooled from embyonic 16.5 day mice	GSM2191004
heart	ENCBS557AAA	Heart from pooled postnatal 0 day mice	GSM2191652
heart	ENCBS225RIF	Heart from 109 embyronic 11.5 day mice	GSM2192039

Window definition file

Since TRAmHap uses window-based summary statistics for prediction, these intervals should be provided in advance. Specifically, two BED files are needed, one for 5 kb region, and another one for 50 kb region. The BED files have the following columns:

Column	Description
Chr	chromosome name
Start	start position
End	end position
gene_win	region name, in the format of geneID_windowNumber

DNA methylation metrics

We used the DNA methylation metrics in TRAmHap, i.e., average methylation, MCR, and MHL. Average and MCR were calculated using mHapTools, and MHL was calculated with an in-house script. Recently, we have developed a new software package mHapTk, which can be used to calculate these three DNA methylation metrics as well as several others.

Finally, all data were loaded into numpy matrix in Python and saved as a .npy file. The matrix has the following columns:

Column	Description
gene	geneID
mM5	mean methylation for 5 kb regions
MHL5	MHL for 5 kb regions
MCR5	MCR for 5 kb regions
mM50	mean methylation for 50 kb regions
MHL50	MHL for 50 kb regions
MCR50	MCR for 50 kb regions

Response variables

RNAseq data were processed using existing pipelines. Gene expression was loaded into numpy and saved as a .npy file.

Column	Description
gene	GeneID
TPM(log2)	Gene expression, log2(TPM+1)

Inputs

TRAmhap has two types of inputs, training input & predicting input.

Training Input

For model training, all the training data should be preprocessed as .npy files. The training command has the following options:

    -S : (str) Naming the output model.
    -I : (str) Data folder with processed data.
    -M : (str) Model folder for saving model parameters.
    -E : (int) Number of epochs for training (default=90).
    -K : (str) Regions to be used. [5, 50, all]
    -T : (str) Statistics to be used. [0,1,2] represent mean methylation, MHL, and MCR, respectively.

Predicting input

For making predictions, the data should be preprocessed as .npy files. The prediction command has the following options:

    -S : (str) Naming the output results.
    -I : (str) Data path to be predict.
    -O : (str) Output Folder for saving result.
    -M : (str) Model parameters to be used.
    -K : (str) Regions to be used. [5, 50, all]
    -T : (str) Statistics to be used. [0,1,2] represent mean methylation, MHL, and MCR, respectively.

Example of usage

We use ENCBS004ZLN sample as the test data set of the model, and we pooled all other heart samples, which are then split into training and validating sets with the ratio of 8:2.

Run prediction

#Training
python3 ~/TRAmHap/RunModel/run_TRAmHap_Train.py -S ENCBS004ZLN \
-I train_data \
-M ~/TRAmHap/TestDatasets/trainning/demo_model \
-T 0,2  # mm & MCR

#Predicting
python3 ~/TRAmHap/RunModel/run_TRAmHap_Predict.py \
-I ~/TRAmHap/TestDatasets/demo_testdata/data/x_ENCBS004ZLN.npy \
-O ~/TRAmHap/TestDatasets/test_file_fig \
-M ~/TRAmHap/TestDatasets/trainning/demo_model/model_ENCBS004ZLN.pkl \
-T 0,2

The predicting output file is .csv file with 4 columns, including index, P_log2(TPM+1).

(The above scatterplot is from R-ggplot)