PheSeq, A Bayesian Deep Learning Model to Enhance and Interpret the Gene Disease Association Studies

In the midst of numerous genotype-phenotype association studies, the derived association findings frequently exhibit a lack of robustness and meaningful interpretations. In response to these challenges, we have introduced PheSeq, a Bayesian deep learning model. This model enhances and elucidates association studies by integrating and comprehending phenotype descriptions.

The repository comprehensively encompasses algorithms and visualization code for PheSeq designed for three exemplary diseases: Alzheimer's Disease, Breast Cancer, and Lung Cancer. Additionally, it provides a flexible data interface, alternative gene network embeddings, and data for 32 Pan-Cancer types. This allows users to seamlessly extend PheSeq to other data modalities and diverse diseases.

PheSeq has undergone testing with Python 3.7 on Ubuntu 21.04 and relies on the following primary dependencies for execution on both CPU and NVIDIA GeForce RTX 3090Ti GPU:

torch==1.7.1
sympy==1.8.0
scipy==1.1.2
transformers==4.10.2
numpy==1.19.5
spacy==2.3.5
scikit-learn==0.20.0

Other dependency packages can be found in the requeirement.txt, and batch installed with the following commend line.

pip3 install -r requirements.txt

To streamline the implementation of PheSeq for additional disease cases, pre-processed text and embedding data for 32 types of Pan-Cancers in the TCGA database are provided at http:https://lit-evi.hzau.edu.cn/Bayes/more-diseases. This dataset comprises extensively annotated sentence evidence and pre-computed embeddings for each gene.

For these 32 cancers, users can directly download the pre-computed embedding data and seamlessly integrate it with the sequence analysis data they have already developed, utilizing the PheSeq model.

The HeterogeneousData folder contains both Embedding data and p-value data for three diseases.

The text data in HeterogeneousData/EmbeddingData/TextData are downloaded from PubTator (https://www.ncbi.nlm.nih.gov/research/pubtator/). In the case when ones would like to collect all literature data related to an interest disease, please search the disease name in PubTator database and download all the Json/PubTator/BioC files.

The graph embedding HeterogeneousData/EmbeddingData/GraphData are downloaded from BioNEV (https://github.com/xiangyue9607/BioNEV).

We additionally provide 5 alternative gene network embeddings, namely node2vec_PPI, Mashup_PPI, BioPlex 3.0, HuRI, Drug-Target Network. The corresponding edge files and pre-trained embeddings can be found in HeterogeneousData/EmbeddingData/GraphData. More information about alternative gene network embeddings can be found in the corresponding folders.

The download link of p-value data is recorded in HeterogeneousData/P-ValueData/README.md. The GWAS Summary data for AD are collected from GWAS Catalog (https://www.ebi.ac.uk/gwas/), and both transcriptome data for BC and methylation data for LC are collected from TCGA (https://www.cancer.gov/about-nci/organization/ccg/research/structural-genomics/tcga). For the disease under consideration, GWAS Summary data need be collected from resource like GWAS Catalog. Please be sure to include both gene site and p-value in the file.

More information about collection and pre-processing of heterogeneous data can be found in "HeterogeneousData/README.md"

If you want to use the PheSeq model for your own heterogeneous data -- a set of embedding data and a set of p-value data, then the two files need to be constructed.
1. Summary data

Summary data, which including the sentence descriptions and p-value for each gene.
Example files can be found in HeterogeneousData/EmbeddingData/TextData.
The Summary file format as flowers (Tab separated):

GENE_LINE: $GENE_Symbol $Entrez_ID $p-valie
$PMID_1 Sentence_1 {$Tag_1, Tag_2}
$PMID_2 Sentence_2 {$Tag_1, Tag_2}

2. Embedding data

Embedding data, it can be derived from different representation learning methods, such as Graph embedding, text embedding.
Embedding files for mat as flowers (Tab separated):

$Entrez_ID_1 $embedding_vec_1
$Entrez_ID_2 $embedding_vec_2

1. Static-PheSeq

python src/run_model.py -ef [embedding_file] -sf [summary_file] --sl -lp [log_save_path] -lf [log_prefix] -mh [multi_hidden] --rs [random_seed] -ed [embedding_size] -lr [learning_rate] -hd [hidden_dim] -tt [train_time] -bs [batch_size] -pt [p_value_threshold]

[embedding_file]: The embedding file.
[summary_file]: The summary file, file format is described in "HeterogeneousData/EmbeddingData/README.md".
[log_save_path]: log_save_path.
[log_prefix]: "predict", prefix of the log file name.
[multi_hidden]: False, Use more hidden layers.
[random_seed]: 126, random size.
[embedding_size]: 128, embedding size.
[learning_rate]: 5e-3, learning rate. [hidden_dim]: 50, the dimension of hidden layer. [train_time]: 100, training time. [batch_size]: 128, batch size.
[p_value_threshold]: 5e-3, The threshold of p-value.

2. Dynamic-PheSeq

python src/Dynamic PheSeq.py -mf [summary file] -ef [embedding file] -ds [data size] -ed [embedding size] -mh [multi hidden] -hs [hidden dim] -zd [z dim] -bs [batch size] -tt [train time] -lr [learning rate] -mr [mu learning rate] -sr [sigma learning rate] -rs [random seed] -ga [gradient accumulation] -sl [save log] -lp [log save path] -lf [log save prefix] -uf [use fake data]

[summary file]: The summary file, file format is described in ”HeterogeneousData /Embed-dingData /README.md”.
[embedding file]: The embedding file.
[data size]: The data size (genotype-phenotype association count) of the input data.
[embedding size]: The embedding dimension of the embedding data.
[multi hidden]: Use more hidden layers.
[hidden dim]: The hidden layer dimension of the deep learning module.
[z dim]: The dimension of the latent variable Z.
\v[log save path]: The folder used to save the log file.
[log prefix]: The name prefix of the log file.
[batch size]: Data count of each mini-batch.
[train time]: The training epoch.
[batch size]: Batch size.
[learning rate]: The learning rate.
[mu learning rate]: The learning rate for μ-related part in the deep learning module.
[sigma learning rate]: The learning rate for σ-related part in the deep learning module.
[random seed]: Random size.
[gradient accumulation]: Using gradient accumulation strategy in model training or not.
[save log]: Save the log file or not.
[log save path]: The folder for saving the log file.
[log save prefix]: The name prefix for the saved log file.
[use fake data]: True, using test file or not.

python src_for_result/Pathological_evidence_network_visualization.py --report_file [report_file] -evidence_file [evidence_file] --save_file [save_file] --topn [topn] --tax_id [tax_id]  --sequence_analysis_gene_file [sequence_analysis_gene_file] --add_STRING [add_STRING] --filter_keyword [filter_keyword]

[report_file]: report file generated by Generate_reports.py.
[evidence_file]: evidence file. It can be downloaded from http:https://lit-evi.hzau.edu.cn/Bayes/more-diseases.
[save_file]: saved pathological network visualization result, the suffix must be ".html".
[topn]: Top-n genes for visualization.
[tax_id]: tax id for species of interest, i.e. 9606 for Homo Sapiens.
[sequence_analysis_gene_file]: Significant gene file from sequence analysis, containing one column, each Entrez id per row.
[add_STRING]: add edges of gene-gene interaction from STRING database, default: False.
[filter_keyword]: keyword used to filter evidence example in pathological network visualization.