Related Experiment Video
Updated: Apr 14, 2026
09:23
Assessing Two-dimensional Crystallization Trials of Small Membrane Proteins for Structural Biology Studies by Electron Crystallography
Published on: October 29, 2010
11.7K
Evaluation of Semi-supervised Learning for Classification of Protein Crystallization Imagery
Madhav Sigdel1, İmren Dinç1, Semih Dinç1
1DataMedia Research Lab, Department of Computer Science, University of Alabama in Huntsville, Huntsville, Alabama 35899, United States.
Proceedings of IEEE Southeastcon. IEEE Southeastcon
|April 28, 2015
Summary
Semi-supervised learning methods, self-training and Yet Another Two Stage Idea (YATSI), improve protein crystallization image classification accuracy with Naïve Bayesian and SMO classifiers. Random forest excelled with supervised learning.
Area of Science:
- Computational Biology
- Machine Learning
- Structural Biology
Background:
- Accurate classification of protein crystallization images is crucial for drug discovery and materials science.
- Limited labeled data presents a significant challenge for training effective classification models.
- Semi-supervised learning offers a promising approach to leverage unlabeled data.
Purpose of the Study:
- To evaluate the performance of two wrapper-based semi-supervised learning methods for protein crystallization image classification.
- To compare the effectiveness of self-training and Yet Another Two Stage Idea (YATSI) with various base classifiers.
- To assess the impact of semi-supervised learning against traditional supervised learning in a low-data regime.
Main Methods:
- Implemented self-training using Naïve Bayesian (NB) and Sequential Minimum Optimization (SMO) classifiers, utilizing high-confidence predictions for iterative training.
- Applied Yet Another Two Stage Idea (YATSI) semi-supervised learning with NB, SMO, Multilayer Perceptron (MLP), J48, and Random Forest (RF) classifiers.
- Conducted experiments on a dataset of 2250 protein crystallization images, varying training and testing data proportions.
Main Results:
- NB and SMO classifiers demonstrated improved accuracy when using both self-training and YATSI semi-supervised approaches compared to supervised learning.
- MLP, J48, and RF classifiers performed better under standard supervised learning conditions.
- The Random Forest classifier achieved the highest accuracy overall when employing supervised learning.
Conclusions:
- Semi-supervised learning, particularly with self-training and YATSI, can enhance classification performance for specific algorithms (NB, SMO) in protein crystallization image analysis with limited labeled data.
- The choice of classifier significantly influences the effectiveness of semi-supervised versus supervised learning strategies.
- For this dataset, Random Forest with supervised learning provided the best accuracy, highlighting the trade-offs between methods.
