## A Denoised Multi-omics Integration Framework for Cancer Subtype Classification and Survival Prediction

---

### What we do?

- We developed a new feature selection method, Feature Selection with Distribution (FSD), for multi-omics data denosing and feature selection.

- We developed a biologically informed deep learning algorithm for multi-omics integration to predict cancer subtypes and patient survival. 

- Commonly used feature selection methods, ANOVA, RFE, LASSO, PCA, were incorporated for comparison.

- Several machine learning and deep learning algorithms, including Random Forest, XGboost, SVM, DNN, MOGONET<sup>1</sup>, Moanna<sup>2</sup>, were integrated for multi-omics integration for cpmparison. MOGONET used graph convolutional networks for multi-omics integration, and Moanna is a Autoencoder-based neural network.

---

<div align=center>

<img src="https://github.com/BioAI-kits/AttentionMOI/blob/master/img/Figure1.jpg" />

</div>

**Introduction of project**. The availability of high-throughput sequencing data create opportunities to comprehensively understand human diseases as well as challenges to train machine learning models using such high dimensions of data. Here, we propose a denoised multi-omics integration framework for cancer subtype classification and survival prediction. Firstly, a distribution based feature denosing algorithm, Feature Selection with Distribution (FSD), were designed to reduce dimensions of omics features. Secondly, we introduced a a multi-omics integration framework, Attention Multi-Omics Integration (AttentionMOI), which is inspired by the central dogma of biology. We demonstrated that FSD improved model performance either using single omics data or multi-omics data in 13 TCGA cancers for survival prediction and kidney cancer subtype identification. And our integration framework outperformed traditional artificial intellegnce models current multi-omics integration algorithms under high dimensions of features. Furthermore, FSD identisied features were related to cancer prognosis and could be considered as biomarkers. 

---

### Install

You can install programs and dependencies via pip. We recommend using conda to build a virtual environment with python version 3.9 or higher.

(optional) Create a virtual environment

```bash

conda create -n env_moi python=3.9

conda activate env_moi  # Activate the environment

```

Install

```bash

pip install AttentionMOI

```

### Parameters

After your installation is complete, your computer terminal will contain a `moi` command. This is the only interface to our program. You will use this command to build an omics model.

First, you can execute the following command line to get detailed help information.

```

moi -h

```

Then, we also introduce these parameters in the following documents: 

**1. Input**

The input file format is described below, or you can refer to the reference data we provide (https://github.com/BioAI-kits/AttentionMOI/tree/master/AttentionMOI/example).

f | omic_file

> REQUIRED: File path for omics files (should be matrix)

**NOTE:The file must be in csv format, such as rna.csv. Of course, it can be compressed with gz, such as rna.csv.gz.**. Example: The first line is the header, patient_id and gene (features) names.

>  patient_id,A1BG,A1CF,A2BP1,A2LD1,....

>

>  TCGA.KL.8323,3.3491,0.0,0.0,5.8939,....

>

>  TCGA.KL.8324,2.922,0.5557,0.5557,6.4226,....

n | omic_name

> REQUIRED: Omic names for omics files, should be the same order as the omics file

l | label_file

> REQUIRED: File path for label file

**NOTE:The file must be in csv format, such as label.csv. Of course, it can be compressed with gz, such as label.csv.gz.**. Example: The first line is the header, patient_id and label represent the sample name and sample classification label respectively. 

> patient_id,label

>

> TCGA.KL.8328,0

>

> TCGA.KL.8339,0

>

> TCGA.KM.8439,1

>

> TCGA.KM.8441,1

>

> TCGA.KM.8442,1

**2. Output**

o | outdir

> OPTIONAL: Setting output file path, default=./output

**3. Feature selection**

method

> OPTIONAL: Method of feature selection, choosing from ANOVA, RFE, LASSO, PCA, default is no feature selection

percentile

> OPTIONAL: Percent of features to keep for ANOVA (integer between 1-100), only used when using ANOVA, default=30

num_pc

> OPTIONAL: Number of PCs to keep for PCA (integer), only used when using PCA, default=50

FSD

> OPTIONAL: Whether to use FSD to mitigate noise of omics. Default is not using FSD, and set --FSD to use FSD

i | iteration

> OPTIONAL: The number of FSD iterations (integer), default=10

s | seed

> OPTIONAL: Random seed for FSD (integer), default=0

threshold

> OPTIONAL: FSD threshold to select features (float), default=0.8 (select features that are selected in 80 percent FSD iterations)

**4. Building Model**

m | model 

> OPTIONAL: Model names, choosing from DNN, Net (Net for AttentionMOI), RF, XGboost, svm, mogonet, moanna, default=DNN.

t | test_size

> OPTIONAL: Testing dataset proportion when split train test dataset (float), default=0.3 (30 percent data for testing)

b | batch

> OPTIONAL: Mini-batch number for model training (integer), default=32

e | epoch

> OPTIONAL: Epoch number for model training (integer), default=300

r | lr

> OPTIONAL: Learning rate for model training(float), default=0.0001

w | weight_decay

> OPTIONAL: weight_decay parameter for model training (float), default=0.0001

---

### Example

Example (Data can be downloaded from https://github.com/BioAI-kits/AttentionMOI ): 

```

moi -f GBM_exp.csv.gz -f GBM_met.csv.gz -f GBM_logRatio.csv.gz -n rna -n met -n cnv -l GBM_label.csv --FSD -m Net -o GBM_Result

```

---

### Ref.

1. MOGONET integrates multi-omics data using graph convolutional networks allowing patient classification and biomarker identification

2. Moanna: Multi-Omics Autoencoder-Based Neural Network Algorithm for Predicting Breast Cancer Subtypes 

---

All rights reserved.