This repository contains workflow, data and extra documentation required to train twelve models based on Biomarker Identification for Protein Sequences obtained through MALDI-TOF Mass Spectrometry as exposed in the manuscript: "An experimental study for Breast Cancer Prediction based on Protein Biomarkers and Ensemble Learning" published in IEEE Latin America Transactions. Also, contains the ready-to-use ML model accesible through Python (>=3.9) and related packages in order to describe relation between proteic composition and breast cancer. The implementation allows users to get breast cancer predictions by applying a Bagging-SVM model over Proteic Sequences.
- Input: A string from protein samples
- Output: Probability and prediction for positive (1) or negative (0) findings.
The project hierarchy and folders description is as follows
- Data Warehousing: Clean, scale, balance and feature subset selection. Here is the data warehousing process. It includes: outliers removal, minmax scaling and invariant columns removal.
Data from Soto, (2020) is recopiled and the process for data warehousing is replicated as a solo script. Relational data format is setted up. A minimax scaler is adapted for every column and then exported as .pkl file. Duplicated entries are removed, and then invariant columns, reducing features from 8742 to 8709. A PCA approach is applied and concludes more than 97% of the variance in class response is explained with 300 columns. Feature Subset Selection is applied, and unselected features are exported. All data files are compressed with a .gz using pigz. You can use unpigz file.pkl.gz or unpigz file.csv.gz to explore them with a gui or simply load them with SQL, Python Pandas or readr package.
-
Model Training: Hyperparameter setting, Training and validation of metrics. The refined model is evaluated with three partitions (70:20:10). The four best classifiers from muntisa paper are selected, trained and exported: Decision Tree, Support Vector Machine with Radial kernel basis function, Logistic Regression and multilayer perceptron. Cross-validation and stratified cross-validation processes are executed to estimate accuracy per model. Then, accuracy, recall, precision, F1 and roc-auc scores are calculated from the confusion matrix and get exported. The ensemble process includes twelve models for which Bagging SVMs have the best properties for prediction.
a) Ensemble metrics
b) Models
c) Supplemental information
-
Evaluation of the selected model: Experimental setting for the best model.
- pigz (> 2.0)
- Python (> 3.9)
- NumPy (> 1.14.3)
- SciPy (> 1.0.0)
- Scikit-learn (> 1.0.0)
- pandas (> 0.22.0)
- joblib (> 1.1)
- r-base (> 4.0)
- Rcpi (>1.2)
- Optional: Run through Ubuntu terminal.
This repo contains two implementations. You can set a list of proteins as the input.fasta file, clone the whole repo and run the bash file:
sh spencer.shOr try the fresh and smooth shiny implementation.
Biomarkers are calculated from input sequences, and then used for final predicition. They include, but are not limited by:
| Descriptor | Description |
|---|---|
| CIDH920105 | Normalized average hydrophobicity scales (Cid et al., 1992) |
| BHAR880101 | Average flexibility indices (Bhaskaran-Ponnuswamy, 1988) |
| CHAM820101 | Polarizability parameter (Charton-Charton, 1982) |
| CHAM820102 | Free energy of solution in water, kcal/mole (Charton-Charton, 1982) |
| CHOC760101 | Residue accessible surface area in tripeptide (Chothia, 1976) |
| BIGC670101 | Residue volume (Bigelow, 1967) |
| CHAM810101 | Steric parameter (Charton, 1981) |
| DAYM780201 | Relative mutability (Dayhoff et al., 1978b) |
This repository is licensed under the terms of the GNU AGPLv3 license.