EXPERT-Disease-GGMP

Recently, machine learning based on microbial features has drawn increasing interests in diagnosis of diverse diseases such as IBD, Type-2 Diabate, etc. However, current methods were unable to mitigate the regional effects, which made it difficult to perform microbial-based cross-regional diagnosis of diseases.

In this study, we proposed a machine learning framework EXPERT-Disease which integrated the neural network and transfer learning for microbial-based cross-regional diagnosis of diseases.We selected Guangdong Gut Microbiome Project (GGMP, EBI ID: PRJEB18535) samples and categorized them into 14 cities according to the sample sources. EXPERT-Disease trains from scratch to build a disease neural network model (DNN) based on the training set of each city, and then uesd training set from other cities for transfer learning to obtain the transferred neural network model (TDNN). Finally, the TDNN can be used for disease diagnosis in other cities. To get the the disease model of each city, please download DNN model for each city.

Get and use

To learn how to install the model and how to use it, click here

The transfer learning workflow of microbial-based cross-regional diagnosis of diseases: (1) Fundamental model establishment: ab initio train the disease neural network (DNN) model on the training set of city A. (2) Transfer model establishment: implement transfer learning to generate a transfer DNN for another city B by using the training set of city B. (3)Classification test: test the transfer DNN on the testing set of city B.

Example

Here we choose Shenzhen(City A) and Guangzhou(City B) as the source city and target city, and we obtained genus-level species abundance tables for these two cities, individuals with T2D are the positive samples and controls are the negitive samples.

Genus-level species abundance tables(reference format): city1_train.tsv city1_test.tsv city2_train.tsv city2_test.tsv
Biome profiles(reference format): biome.tsv
Mapper profiles(reference format): city1_train_mapper.tsv city1_test_mapper.tsv city2_train_mapper.tsv city2_test_mapper.tsv

Ontology construct

• Construct a biome ontology representing stages of T2D. You'll see constructed ontology like a tree in the printed message.

expert construct -i biome.tsv -o ontology.pkl

Source mapping

• Map microbial community samples to the biome ontology to obtain hierarchical labels. You'll see counts of the samples on each biome ontology layer in the printed message.

expert map --to-otlg -i city1_train_mapper.csv -t ontology.pkl -o city1_train_labels.h5
expert map --to-otlg -i city1_test_mapper.csv -t ontology.pkl -o city1_test_labels.h5
expert map --to-otlg -i city2_train_mapper.csv -t ontology.pkl -o city2_train_labels.h5
expert map --to-otlg -i city2_test_mapper.csv -t ontology.pkl -o city2_test_labels.h5

Data convert

• Convert input abundance data to model-acceptable hdf file. The EXPERT model only accepts standardized abundance data. Here we standardize the abundance data using convert mode.

ls city1_train.tsv > inputlist; expert convert -i inputlist -o city1_trainCM.h5 --in-cm;
ls city1_test.tsv > inputlist; expert convert -i inputlist -o city1_testCM.h5 --in-cm;
ls city2_train.tsv > inputlist; expert convert -i inputlist -o city2_trainCM.h5 --in-cm;
ls city2_test.tsv > inputlist; expert convert -i inputlist -o city2_testCM.h5 --in-cm;

Ab initio training

• Train the disease neural network model from scratch. Here we will use ontology.pkl and hdf files.

expert train -i city1_trainCM.h5 -l city1_train_labels.h5 -t ontology.pkl -o city1_DNN

Transfer learning

• Transfer the knowledge of city B to the DNN model of city A for better performance in disease diagnosis on city B. You'll see running log and training process in the printed message.

expert transfer -i city2_trainCM.h5 -l city2_train_labels.h5 -t ontology.pkl -m  city1_DNN -o Transfer_DNN

Search

• Search the test set of city B against the transferred DNN model.

expert search -i city2_testCM.h5 -m Transfer_DNN -o Search_Transfer_DNN

Evaluation

• Evaluate the performance of the Transferred DNN model. You'll obtain a performance report.

expert evaluate -i Search_Transfer_DNN -l city2_test_labels.h5 -o Evaluation

The illustration of the R code

city.R

• Randomly selected samples of each city for each disease, into training subset(80%) and testing subset(20%). Generated the biome file, mapper file and the path file for each experiment.

city_gradient.R

• The gradient experiments of the disease diagnosis across cities, the proportion of the training subset was 20%, 40%, 60% and 80% respectively.

draw_figure_city_AUROC.R

• The evaluation results of the three models of 14 cities respectvely and the collective results.

draw_figure_city_heatmap.R

• The heatmap of the average AUROC of the three models of each city pair.

draw_figure_city_relavance.R

• The boxplots and the lines showed the correlation between the sample size for transfer learning and the accuracy of the three models.

story.R

• Using the "leave-one-out" method to discover the "region-spectific" microbes and the "region-shared" microbes.