C code for HDPGM parallel computing. Currently, the code can be run on Linux and MacOS.

HDPGM is short for the hiearachical and dynamic Poisson graphical model. The model was proposed by Xiangyu Luo and Yingying Wei to recover transcription factor (TF) networks from ChIP-seq count data. On the one hand, TF networks are heterogeneous across the genome. On the other hand, TF networks are dynamic with respect to different cell types. Under the mild assumption that the whole genome can be partitioned into several (unknown) sets and genomic locations in the same set share the same TF networks, HDPGM can be used to simultaneously construct the dynamic TF networks and cluster genomic locations without need to pre-specify the number of clusters.

The Markov chain Monte Carlo algorithm (MCMC) was used to draw posterior samples from HDPGM. For more technical details, please refer to our paper Nonparametric Bayesian Learning of Heterogeneous Dynamic Transcription Factor Networks (published in Annals of Applied Statistics: https://projecteuclid.org/euclid.aoas/1536652973).

The idea of the message passing interface (MPI) was employed to parallelize the MCMC algorithm. Among MPI functions, MPI_Rec and MPI_Send were used the most frequently in the C code.

Assume there are N genomic locations, D conditions, p TFs, and R replicates.

1. Download the source file HDPGM.c and the header file HDPGM.h in your working directory.
2. Prepare your data file (a N by D*p*R matrix with row names and column names removed) where rows correspond to N genomic locations. For the columns, column 1 corresponds to (condition 1, TF 1, replicate 1), which is coded as (1,1,1). Using the same notation, column 2 is (1,1,2), ... , column R is (1,1,R), column R+1 is (1,2,1), column p*R is (1,p,R), column p*R+1 is (2,1,1) and so on so forth until column D*p*R which is (D, p, R).
3. Name your data file as "Data.txt" and put it in your working directory.
4. In your working directory, make a new directory "PosteriorSamples". Under the folder "PosteriorSamples", make four new directories: "P_dp_t", "cla_t", "L_t", and "Lambda_t" which are used to collect posterior samples for cluster proportions, cluster indicators, edge indicators, and dependence intensity parameters.
5. In the source file HDPGM.c, set N, D, p, R, and the maximum cluster number M.
6. Open the terminal, change the current diretory to your working directory and input the following two commands (note that 10 threads are used here, and you could flexibly set it).

mpicc HDPGM.c -o HDPGM -lm -std=c11
mpirun -np 10 ./HDPGM

1. The number of threads must be greater than one.
2. If N cannot be divisible by the thread number L, the first L-1 threads have [N/L] genomic locations each and the last thread has N - (L-1)[N/L] genomic locations, where [x] is the largest integer less than or equal to x.
3. When the program is running, there will be a "log.txt" file showing the running progress.
4. If you find any bug or have any question, please contact us via email to xyluo1991 at gmail dot com.