1. @ARTICLE{SG-LSTM-FRAME,
2. author = {Weidun Xie(shelton xie),Jiawei Luo,Chu Pan,Ying Liu},
3. title = {SG-LSTM-FRAME: A computational frame using sequence and geometrical information via LSTM to predict miRNA-gene associations},
4. year = {2019},
5. journal = {},}

• Hardware: Core: Intel i7-7700HQ ; Graphic Card：GTX1060 6G; RAM:32 GB ddr4 2666
• CUDA Version 10.0.130
• CuDNN version 7.4.2
• OS: Win 10 64bit
• Python version: IDLE (Python 3.7 64-bit)#yep I wrote directly in IDLE :)
• (In my another computer, I don't have CUDA for GPU and tensorflow-gpu, but I tested that it could still run in that computer)
• Hardware: Core: Intel i5-2420M,RAM:16 GB ddr3 1600
• OS: Win 10 64bit
• Python version: IDLE (Python 3.7 64-bit)

• PyQt5 5.10.1(for UI)
• Pandas 0.24.1(for file read)
• Sklearn 0.0(for comparison)
• tensorflow-estimator 1.13.0(for LSTM)
• tensorflow-gpu 1.13.1(for LSTM)
• numpy 1.16.2(for npy read and other functions)
• matplotlib 3.0.3(for roc curve plot)

This is the code of SG_LSTM_core, the only difference between SG_LSTM_core and SG_LSTM_WHOLE is the difference between dataset and the way we store the unknown pairs' score.

• print('In this test, The dataset is randomly separated into train and test set, you don’t have to generate other file')
• what you need is the dataset downloaded or generated use SG_LSTM_Dataset UI
• then change RNN(LSTM) - fast validate-train.py your_dataset_file_name='AXB_383_gene_default.csv' into
• your_dataset_file_name='{}.csv' .format(your file's name!!!)
• open the RNN(LSTM) - fast validate-train.py in IDLE and Press F5
• It will automatically begin to train ,the model will be saved in the same folder

AXB_383_gene_default.csv 159.1MB link：https://pan.baidu.com/s/1Yy8RNfinV2B0G9l_Gb8-Ag password：11v1

• In order to fast test: first put RNN(LSTM)-fast-validate-test.py and AXB_383_gene_default.csv and models in the same folder.
• Then change:
• your_dataset_file_name='AXB_383_gene_default.csv' into your_dataset_file_name='your file generated.csv' if you decide to generate your own dataset. otherwise ,skip this step
• Then change:
• your_file_path="C://Users/shelton/Desktop/fast validate/models" into your file root , otherwise it cannot find the model.
• After the model successfully loaded, test will automatically run.
• it will produced the y_score and y_label for your test set, named LSTM_y_score and LSTM_y_label ,and also after running the RNN(LSTM)-fast-validate-test.py,a ROC -curve would appear to indicate the performance of the model.

• 0 SG-LSTM(for test the model).py can help to build the model
• 0 SG-LSTM(for roc curve generation).py can use the model to do the test job, (generation of ROC curve) also the scores!
• 1.put SG-LSTM(for test the model).py and X_train5.csv and X_test5.csv in the same folder
• 2.run SG-LSTM(for test the model).py (after a while of running, a folder named model5 would be generated)
• 3.put SG-LSTM(for roc curve generation).py and X_train5.csv and X_test5.csv in the same folder
• 4.You have to change the file root so that the roc curve could be generated successfully:
• From module_file = tf.train.latest_checkpoint("I://DNN模型重新测试/LSTM测试/测试24 SG相乘128 五折交叉验证/LSTM for test/models{}".format(gg))
• To module_file = tf.train.latest_checkpoint("C://Users/shelton/Desktop/SG-LSTM_master/models{}".format(gg))
• here 'gg' means the iteration of fold. gg=5 means the last fold of SG-LSTM-core
• please change it into your file root, otherwise it can not load the model.

• X_train5.csv links：https://pan.baidu.com/s/1ZvlfgQ8IjEo61extIU6WvQ password：5o11
• X_test5.csv links：https://pan.baidu.com/s/1X7OmbCQq8OeudzWmkZydqQ password：ld7z

• 5_fold_cross_divided.py and 'AXB_383_gene_default.csv'

• 5_fold_cross_divided.py can divide 'AXB_383_gene_default.csv'(or other dataset) into five fold of train and test parts.

• all_unkown_sample_generation.py can generate all negative samples for prediction. Here in SG-LSTM-core, we generate all unknown samples by generating file. But in the SG-LSTM-WHOLE, because there are more than 14,000,000 pairs, we generate the score by producing a matrix to record the scores.

• S_gene2vec_320_34567.csv S_mrna2vec_384_34567.csv is the embedding generated from sequence information of 320 gene and 384miRNA

• G_gene2vec_328_gene.csv G_mirna2vec_384_mirna.csv is the embedding generated from geometric information.

• S_gene2vec_320_34567.csv+G_gene2vec_328_gene.csv can generate SG128_317_gene.csv.

• S_mrna2vec_384_34567 +G_mirna2vec_384_mirna.csv can generate SG128_383_miRNA.csv

• SG128_317_gene.csv is the embedding merged from Geometric information embedding and Sequence information embedding of gene.

• SG128_383_miRNA.csv is the embedding merged from Geometric information embedding and Sequence information embedding of miRNA.

• Using SG128_317_gene.csv and SG128_383_miRNA.csv and a file from mitarbase, you can generate positive samples; After calculate all the pairs' Euclidean distance and cosine-similarity, you can earn a average distance. Using the positive samples and distance, you can generate the dataset, like 'AXB_383_gene_default.csv'

• 'AXB_383_gene_default.csv' is the training Dataset of SG-LSTM-core.

I designed an UI for dataset construction. It contains all the content about how to use SG128_317_gene.csv and SG128_383_miRNA.csv to generate file like 'AXB_383_gene_default.csv'

• SG_LSTM_Dataset UI https://github.com/Xshelton/SG_LSTM_DatasetUI

#If your only want the results of my research:

• SG-LSTM_core_result: https://github.com/Xshelton/SG_LSTM_core_result

In the WHOLE part,more than 14,000,000 pairs of miRNA and Gene were predicted.

• SG-LSTM_WHOLE_result https://github.com/Xshelton/SG-LSTM_WHOLE_result

3186 pairs of the associations were predicted by all of three mentioned methods.