This work started as class project for NAIL107 Machine Learning in Bioinformatics class at Charles University of Prague. Inspired by [1], given chemical formula of compound, we tried to predict metabolic pathways in which this compounds is involved. But unlike the approach in the book, we formulated the problem as multi-label classification, that is, every compound can be involved in multiple pathways.
Implementation is in python using sci-kit learn (https://scikit-learn.org/) and scikit-multilern (http://scikit.ml/) libraries. Data are from KEGG database (https://www.kegg.jp/).
Overall accuracy/precision/recall was calculated as average accuracy/precision/recall over all classes.
| Method | Overall accuracy | Overall precision | Overall recall |
|---|---|---|---|
| DecisionTreeClassifier (LabelPowerset/BinaryRelevance) | 89.71% | 99.60% | 89.94% |
| MLkNN | 86.26% | 93.05% | 91.00% |
| RandomForestClassifier | 89.75% | 99.99% | 89.75% |
| SVC (BinaryRelevance) | 89.31% | 98.15% | 90.48% |
| MLPClassifier | 89.76% | 99.83% | 89.83% |
| MLPClassifier 50-50 | 89.81% | 99.86% | 89.87% |
| MLPClassifier 50-100-50 (no convergence) | 89.77% | 99.44% | 90.07% |
| MLPClassifier 1000 | 89.74% | 99.60% | 89.97% |
| AdaBoostClassifier (BinaryRelevance) | 89.69% | 99.88% | 89.77% |
Measures per class (DecisionTreeClassifier):
| Class | Accuracy | Precision | Recall |
|---|---|---|---|
| Carbohydrate | 90.74% | 100.00% | 90.74% |
| Energy | 82.84% | 99.54% | 83.02% |
| Lipid | 98.20% | 100.00% | 98.20% |
| Nucleotide | 77.51% | 99.29% | 77.65% |
| Amino acid | 96.24% | 100.00% | 96.24% |
| Other amino acids | 83.17% | 99.93% | 83.21% |
| Glycan | 89.16% | 100.00% | 89.16% |
| Cofactors/vitamins | 97.22% | 100.00% | 97.22% |
| Terpenoids | 86.55% | 99.69% | 86.78% |
| Secondary metabolites | 90.03% | 100.00% | 90.03% |
| Xenobiotics | 96.30% | 100.00% | 96.30% |
Contrast that with what we belive to be current state of the art [2], which achieved accuracy over 95% for all classes using structure data and graph neural networks.
[1] Yang, Z. R. (2010). Machine learning approaches to bioinformatics (Vol. 4). World scientific.
[2] Mayank Baranwal, Abram Magner, Paolo Elvati, Jacob Saldinger, Angela Violi, Alfred O Hero, A deep learning architecture for metabolic pathway prediction, Bioinformatics, Volume 36, Issue 8, 15 April 2020, Pages 2547–2553, https://doi.org/10.1093/bioinformatics/btz954