aditirk1 / DBSCAN

Implementation of an ML algorithm, DBSCAN, on biological data.

DBSCAN

Implementation of an ML algorithm, DBSCAN, on metabolome abundance data in treatment naive GBM patients.

Description of the data: Proteogenomic and metabolomic characterization of human glioblastoma. Whole genome or whole exome sequencing of 99 samples. Generated by CPTAC.

As seen in "Proteogenomic and metabolomic characterization of human glioblastoma. Cancer cell" by Wang LB, Karpova A, Gritsenko MA, et al. (doi:https://doi.org/10.1016/j.ccell.2021.01.006), Metabolite identifications and data processing were conducted as previously detailed (Webb-Robertson et al., 2014). GC-MS raw data files were processed using Metabolite Detector software v2.0.6 beta (Hiller et al., 2009). Retention indices (RI) of detected metabolites were calculated based on the analysis of the FAMEs mixture, followed by their chromatographic alignment across all analyses after deconvolution. Metabolites were identified by matching experimental spectra to an augmented version of the Agilent Fiehn Metabolomics Retention Time Locked (RTL) Library (Kind et al., 2009), containing spectra and validated retention indices. All metabolite identifications were manually validated. The NIST 08 GC-MS library was also used to cross validate the spectral matching scores obtained using the Agilent library and to provide identifications for metabolites that were initially unidentified. The three most abundant fragment ions in the spectra of each identified metabolite were automatically determined by Metabolite Detector, and their summed abundances were integrated across the GC elution profile. A matrix of identified metabolites, unidentified metabolite features, and their corresponding abundances for each sample in the batch were exported for statistics.

Executive Summary

The project aimed to perform metabolomic characterisation and analysis of Glioblastoma. Glioblastoma is a highly aggressive brain tumor with complex metabolic dysregulation, making it crucial to understand the underlying metabolomic patterns to gain prospective therapeutic insights.

High dimensional metabolic data was preprocessed, and clustered using the DBSCAN algorithm to describe metabolic abnormalities. Through principal component analysis (PCA) and differential expression analysis, we aimed to uncover subtypes and compare the data obtained from 99 GBM patients' whole genome generated by CPTAC.

Cancer cells often exhibit altered metabolic pathways to support rapid growth and proliferation. The analysis was performed on various attributes:

• Comparison of lactic acid, pyruvate, glucose levels
• Comparison of homocysteine and creatinine levels
• Comparison of galactitol and glucose levels
• Clustering based on age and BMI

The project utilized Python libraries such as pandas, NumPy, and matplotlib to preprocess and visualize metabolomic data. The knee point concept was used to determine the optimal value of the epsilon parameter.

Visualizations such as heatmaps and pathway maps elucidated disrupted metabolic pathways in GBM, providing critical insights into potential targets and biomarkers. This project highlights Python as a powerful tool for bioinformatics and metabolomics research, enabling comprehensive exploration and comparison of metabolite expression patterns in Glioblastoma.

Introduction

Density-Based Spatial Clustering of Applications with Noise (DBSCAN) is a machine learning algorithm for density-based clustering, first introduced in 1996 by Martin Ester and Hans Peter. It is a widely used unsupervised learning method for model construction and machine learning algorithms in various fields encompassing biological data analysis. DBSCAN is useful in identifying high-density clusters within data, separating them from low-density regions, and is suitable for arbitrarily shaped clusters. DBSCAN is found to be more advantageous than k-means clustering as it does not require the number of clusters to be specified beforehand, and it efficiently handles and identifies noise points. This makes it well-suited for very large datasets. DBSCAN's versatility has led to its application across various domains, including image processing, geospatial data analysis, data mining, astronomy, and anomaly detection. Glioblastoma (GBM) is an exceptionally aggressive and malignant type of brain tumor. It is a grade 4 glioma brain tumor arising from brain cells called glial cells. A brain tumor's grade refers to how likely the tumor is to grow and spread. Grade 4 is the most aggressive and serious type of tumor. Symptoms of GBM include headaches, seizures, changes in behavior, and motor weakness. Understanding the metabolic alterations in treatment-naive GBM is crucial for developing targeted therapies and increased progression-free survival.

1.1 About the Domain

DBSCAN is an unsupervised machine learning algorithm that uses two parameters: epsilon (ε), the maximum radius of the neighborhood around a point, and MinPts, the minimum number of points required to form a dense region or cluster. Its applications are far-reaching.

1. Geospatial Data Analysis: Identifying densely populated areas, urban growth patterns, and hotspots for infrastructure development; detecting clusters of pollution, wildlife populations, or natural resources.
2. Image Processing: Segmenting images based on pixel density to identify objects or regions of interest; analyzing clusters in MRI or CT scans to detect anomalies such as tumors or lesions.
3. Market Research: Grouping customers based on purchasing behavior, preferences, or demographic information for targeted marketing; identifying trends and patterns in sales data to optimize product placement and inventory management.

1.2 Scope and Objectives

The Problem

GBM is an aggressive brain tumor with a high degree of metabolic dysregulation. Traditional clustering algorithms, such as k-means, are not optimal for this type of data due to their requirement for predefined cluster numbers and sensitivity to noise. There is a need for advanced clustering methods that can handle the complexity and high dimensionality of metabolomic data from GBM patients.

Objectives

The primary objective of this study is to utilize DBSCAN for clustering based on the density of data points, identifying noise and outliers, and filtering out abnormalities in the dataset. This approach helps in understanding and analyzing the patterns and structures present in the data.

• Implement DBSCAN with optimized parameters using the knee point method for clustering.
• Analyze GBM metabolomic profiles using PCA and identify subclasses based on visualized patterns.
• Conduct analyses on various attributes, including:
  • Clustering based on the abundance of lactic acid, pyruvate, and glucose.
  • Clustering based on the abundance of homocysteine and creatinine.
  • Clustering based on the abundance of galactitol and glucose.
  • Clustering based on age and BMI.

Attributes

GBM cells, like most cancer cells, exhibit the Warburg effect, producing energy through glycolysis followed by lactic acid fermentation, even in the presence of oxygen. This metabolic shift can distinguish cancer patients from healthy individuals by analyzing their metabolic profiles. The selected attributes are vital in energy-producing metabolic pathways.

• Lactic Acid and Pyruvate Levels: Comparing these levels provides an understating about the shift towards glycolysis.
• Homocysteine Levels: As an intermediate in methionine metabolism, homocysteine is associated with oxidative stress.
• Creatinine Levels: Perturbations in creatinine levels offer insights into overall metabolic health.
• Galactitol Levels: Accumulation of galactitol indicates dysregulation in galactose metabolism.
• Age and BMI: Age can influence GBM progression, and clustering based on age can help identify age-specific metabolic patterns. BMI may also correlate with metabolic changes in GBM.

Detailed Design and Architecture

2.1 Proposed System Architecture

The proposed system architecture is designed to detect unique metabolic subgroups in Glioblastoma Multiforme (GBM) patients (treatment-naive) by grouping metabolite abundance profiles using the DBSCAN algorithm. The system is intended to take high-dimensional metabolomic data, preprocess it, and perform clustering to identify underlying metabolic abnormalities and possible subgroups in GBM. The system architecture's main components include data collection, preprocessing, clustering algorithm application, visualization and analysis.

2.2 Design Architecture

Python is a multi-paradigm language that supports various programming approaches, including object-oriented, procedural, and functional styles. It is an interpreted language, meaning that the code is executed directly without being compiled into machine code first. This interpreted nature makes Python highly interactive, allowing developers to test and experiment with code in real-time. In this code, the architecture follows a modular approach. The Data Input Module encapsulates functionality to handle input data files, ensuring compatibility across various formats and detecting potential errors or inconsistencies during file validation. It transforms input data into a structured format, typically a pandas DataFrame, to facilitate seamless processing. The Exploratory Data Analysis Module provides tools for computing descriptive statistics, generating correlation matrices, and visualizing data through heatmaps and scatter plots. This allows in uncovering underlying patterns and potential clusters within the metabolomic data. The Data Normalization and Scaling Module is used in preprocessing by standardizing feature scales using techniques like Min-Max scaling or standardization, ensuring uniformity across all features. The Dimensionality Reduction Module employs algorithms such as Principal Component Analysis (PCA) to reduce the dimensionality of high-dimensional data while preserving variance, thus enhancing visualization and facilitating effective clustering. The Clustering Module implements the DBSCAN algorithm for identifying clusters based on optimized parameters, determined using techniques like the KneeLocator algorithm for epsilon selection. Lastly, the Visualization enables the visualization of clustering results, creating scatter plots to highlight different clusters and distinguish between noise and cluster points, thus providing a visual aid for identifying subgroups.

2.3 Methodology

1. Data Acquisition and Preprocessing

The first step was to load the metabolomic data from a CSV file using pandas for efficient data handling and manipulation. Data integrity and completeness must be maintained by handling missing values through mean imputation. Finally, the dataset was transposed to align metabolite features as columns and patient samples as rows, optimizing the data structure for subsequent analysis.

2. Merging necessary files

Clinical patient data consisting of various categorical and numeric data (which were the features used in the analysis) was merged with metabolome abundance data based on the column “Patient_ID”. This ensured that both data were retained, with missing data being depicted as “NaN”. This step was performed for other analysis apart from the PCA derived data.

3. Exploratory Data Analysis (EDA)

Following the loading of required data, descriptive statistics using pandas to summarize the distribution of metabolite abundances across samples, including metrics such as mean, standard deviation, and quartiles, was performed. Utilized seaborn and matplotlib for generating correlation matrices and visualizing interrelationships among metabolites using heatmaps and scatter plots. These visualizations can determine potential metabolic patterns and outliers.

4. Data Normalization and Scaling

Normalization of metabolite abundances using MinMaxScaler from scikit-learn was performed to standardize data across a uniform range, preparing it for downstream analysis. This step ensures that all features contribute equally to algorithms and prevents biases due to varying scales.

5. Dimensionality Reduction

Implementation of Principal Component Analysis (PCA) from scikit-learn was done to reduce the dimensionality of the dataset while retaining essential variance. For the rest of the analysis, PCA was not performed as there were two dimensions only, when comparing two features in the data. Project the data into a lower-dimensional space using PCA, facilitating efficient visualization and clustering. This simplifies the interpretation of complex metabolomic profiles.

6. Clustering Analysis

Utilized the Density-Based Spatial Clustering of Applications with Noise (DBSCAN) algorithm from scikit-learn for clustering analysis. Optimized parameters such as epsilon and minimum samples using techniques like KneeLocator for enhanced clustering accuracy. Distinct metabolic subgroups within the data based on clustering results were identified, distinguishing noise from cluster points for comprehensive analysis.

7. Visualization and Interpretation

Visualized clustering outcomes using matplotlib and seaborn to create scatter plots and color-coded clusters. This visualization approach helps in effectively interpreting metabolic subgroups and highlighting important differences between clusters.

Implementation

3.1 About the code

DBSCAN was implemented using Python3 on the Jupyter Notebook. The code structure is as follows. Load Data file and Read the data Data was downloaded from cBioPortal for Cancer Genomics. The file consisted of various data including protein expression, mRNA expression, methylation, copy number variations and so on. We stuck to metabolome abundance data. The file was converted from .txt to .csv and loaded into python using the pandas library’s read_csv function as a dataframe.

import pandas as pd
data = pd.read_csv('metabolome.csv', index_col=0)

Dataset Structure The data consists of metabolite abundance data of 75 patients affected with glioblastoma multiforme (GBM) before any treatment regimes. There are a total of 134 metabolites, with some of them being labeled as “Unknown” by the authors. Columns are patient IDs and rows are metabolite names. The dataset was transposed using python by the code data.T in order to perform further deductions and conversions (scaling and dimensional reduction).

Data Visualization

• Utilized seaborn (import seaborn as sns) and matplotlib (import matplotlib.pyplot as plt) for data visualization.
• Created a heatmap using sns.heatmap() to visualize correlations between expression of metabolites.

Data Scaling and Dimensionality Reduction

• Employed StandardScaler() and MinMaxScaler() from scikit-learn for data scaling.
• Performed Principal Component Analysis (PCA) using PCA() from scikit-learn.
• Created a new DataFrame df with dimensionality-reduced data, with labels ‘X’ and ‘Y’.

Optimal Epsilon Determination

• Iterated over a range of k values using range().
• Fitted NearestNeighbors() from scikit-learn to compute distances and indices.
• Calculated mean distances and sorted them using NumPy operations.
• Employed KneeLocator() from the kneed library to find the optimal epsilon value.
• Plots the k-distance graph using plt.plot(), plt.axvline(), plt.xlabel(), plt.ylabel(), plt.title(), plt.legend().
• 2-dimensional data requires the use of DBSCAN’s default value of minpts = 4 (Ester et al., 1996)
• If the data has more than two dimension, minpts = dimensions*2, or minpts >= dimensions+1

Clustering with DBSCAN

• Applies DBSCAN clustering using DBSCAN() from scikit-learn.
• Stores cluster assignments in the DataFrame using .labels_.
• Separated clusters and noise points based on cluster labels.
• Visualized clusters using sns.scatterplot() and plt.scatter().

Further Data Processing and Clustering

• Merged data from ‘clincial.csv’ and ‘metabolites.csv’ CSV files using pd.merge().
• Metabolites.csv is different from metabolome.csv in that the index_col name is Patient_ID in order to merge the two files on the column Patient_ID.
• Selected a subset of features for clustering analysis.
• Repeated the process of optimal epsilon determination and DBSCAN clustering on the subset.
• Created scatter plots using plt.subplots() and sns.scatterplot() to visualize clustering results for different feature combinations relevant to GBM.
  • Lactic acid, pyruvic acid, glucose
  • Age, BMI
  • Galactitol, glucose
  • Homocysteine, creatinine

3.2 Functions and Tools used

1. Numpy: A fundamental package for scientific computing with Python. It provides support for arrays, matrices, and many mathematical functions.
2. Pandas: A data manipulation and analysis library offering data structures like DataFrame.
3. DBSCAN: A clustering algorithm that groups points closely packed together while marking outliers as noise.
4. StandardScaler: Standardizes features by removing the mean and scaling to unit variance.
5. PCA (Principal Component Analysis): A dimensionality reduction technique that reduces the number of variables in a dataset, retaining maximum information possible.
6. Matplotlib: A library intended to create interactive visualizations by creating plots and graphs.
7. NearestNeighbors: It is a sci-kit module that performs unsupervised nearest neighbor learning.
8. KneeLocator (from kneed library): A method to automatically identify the "knee" or elbow point in a plotted curve, used to determine the optimal epsilon value for DBSCAN.
9. Seaborn: A data visualization library based on matplotlib, providing a high-level interface for creating attractive and informative statistical graphics.
10. MinMaxScaler: A scikit-learn preprocessor that scales features to a given range, in this case, scaling the data to the range [0, 1].

Results

We compared levels of key metabolites such as lactic acid, pyruvate, glucose, homocysteine, creatinine, and galactitol, which resulted in notable differences that highlight the unique metabolic reprogramming in Glioblastoma Multiforme. To add to the interpretation, clustering based on demographic factors like age and BMI provided further context, showing how these variables might influence or correlate with metabolic changes in the tumor environment.

data.head()

After transposing the data,

data.info()

data.corr()

Elbow plots based on k nearest neighbor were highly resourceful in identifying the best epsilon value for the dimensionally reduced or only scaled data. The minpts (minimum number of samples to consider as a cluster in DBSCAN) value was determined by the appropriate k value (as k = minpts). In this case, optimum eps value was determined to be 0.2767554389483952 at k = 4.

The results of the clustering showed 4 major clusters that could indicate different metabolic subgroups within the GBM patient cohort. These subgroups may reflect variations in metabolite abundances and metabolic pathways across different patients. They may have clinical implications, such as guiding personalized treatment strategies based on metabolic profiles. The obtained subgroups need further validation with rigorous testing and additional data analysis.

Next, the three metabolites that are most implicated in GBM due to the Warburg effect - an increase in the rate of glucose uptake and preferential production of lactate, even in the presence of oxygen – pyruvate, lactate, and glucose, where considered for clustering.

The results were pretty significant for 2 plots out of 3. There were two subgroups for glucose and lactic acid abundance, where there is an indication that one group has a higher lactic acid abundance compared to the other, for the same glucose levels. Another interesting finding was pyruvate vs. lactic acid – again, two subgroups that were clearly separated and clustered well using DBSCAN. The figure is suggestive of the trend that for similar lactic acid levels, there is one sub group expressing pyruvate at a higher level than the other. The implications and potential outcomes of the variation between these subgroups must be investigated further.

Age and BMI of the patient cohort were also clustered and visualized to see the spread of population in the given cohort. 3 clusters were obtained, and by visual interpretation, the cohort seems to consist of patients aged 60-70 primarily, with BMI varying between ~18 to 33, whereas the second cluster of middle aged people between 40-53 have BMI in the higher range ~24-33. The final cluster is a small sample of ages 50-54 with BMI well within 25 and above 20.

Creatinine is a breakdown product of creatine phosphate from muscle and protein metabolism. It is commonly used as a marker of kidney function. Its levels can vary in different physiological states and diseases, including cancers. Glioblastoma (GBM) is a malignancy dominated by the infiltration of tumor-associated myeloid cells (TAMCs). Examination of TAMC metabolic phenotypes in mouse models and patients with GBM identified the de novo creatine metabolic pathway as a hallmark of TAMCs. Homocysteine is an amino acid that is involved in methionine metabolism and elevated levels are associated with cardiovascular disease and potentially with certain cancers. Homocysteine has a detrimental influence on human neurons as seen in recent studies. The proneural-like subtype of GBM shows significantly increased levels of creatinine and homocysteine compared to other subtypes.

Three clusters were identified, but the significance of clustering can only be determined after investigating the possible overlapping biological pathways and molecular basis for the upregulation of the metabolites, as well as whether the correlation holds valid by performing laboratory experiments.

The final plot was to cluster expression of galactitol and glucose. Galactitol is a glucose epimer produced from aldose reductase, the first enzyme of the polyol pathway [9]. Physiologically, this pathway converts excess glucose into fructose through the sequential activity of two enzymes: aldose reductase and sorbitol dehydrogenase. To check if there is any evidence of clustering, a scatter plot was plotted to visualize raw data:

Since the plot suggested some form of clustering and variance in expression, the same procedure was followed as above to generate clusters using DBSCAN for these two metabolites. An elbow plot was plotted once again to check for the optimal k and epsilon value. The value was found to be k = 4 and eps = 0.8010010322532317. Clustering results were distinct, with a total of 7 clusters! This possibly suggests that for different subgroups of GBM, the expression of glucose varies from ~ 20 to 26.

For the various levels of glucose expression, galactitol shows a markedly different pattern of expression in each subgroup, suggesting the possibility that the glycolysis pathways may be diverted in some patients to the polyol pathway to make up for excess glucose.

Conclusion and Future Enhancement

The clustering results obtained from the DBSCAN analysis on the metabolomic data represent only the first step of many in finding potential metabolic subtypes and patterns in glioblastoma multiforme (GBM) patients. These initial clusters require much more in-depth analysis and interpretation to fully understand their biological significance and clinical implications. A key step is to look at the exact metabolic pathways and activities connected with each cluster. This can be accomplished by doing pathway enrichment analysis, which is basically mapping the metabolites in each cluster to known metabolic pathways and detecting those that are highly enriched or dysregulated. Understanding the underlying metabolic pathways provides an idea about fundamental biological mechanisms that drive the observed metabolic subtypes, as well as their possible role in tumor growth, therapy response, and patient outcomes.

It is important to note that DBSCAN is just the initial step in a multi-faceted analysis pipeline. The obtained clusters can serve as a starting point for further exploration and validation using additional statistical methods such as the t-test, ANOVA, machine learning techniques, or in vitro and in vivo experiments. These subsequent analyses can help refine the clustering results, identify potential drivers of GBM, and ultimately translate the findings into clinical applications, such as developing novel diagnostics, prognostic markers, or therapeutic strategies tailored to specific metabolic subtypes of GBM.

One challenge that may arise in the analysis of metabolomic data is the curse of dimensionality. As the number of features (metabolites) increases, the computational complexity and computational resources required for clustering algorithms like DBSCAN can increase exponentially. High-dimensional data can lead to sparse data distributions, making it difficult to identify meaningful clusters and patterns. To address this challenge, dimensionality reduction techniques like Principal Component Analysis (PCA), UMAP or t-SNE can be employed to project the high-dimensional data into a lower-dimensional space while preserving the most relevant information. Although PCA was used to reduce the dimensionality of the data, analysis of the clusters and meaningful interpretations can only be performed with thorough domain knowledge.

The future of DBSCAN and its applications in metabolomics and other omics fields holds great promise. Continuous improvements in user-friendliness, computational efficiency, and versatility will enhance the ability to analyze and interpret complex biological data. Enhancing parallel processing methods for multi-core and GPU-based systems and integrating DBSCAN into distributed computing frameworks such as Apache Spark are two ways to boost scalability. This algorithm has huge potential for improvement in the field of biological data science, and is widely being used to this day in high-dimensional data analysis in biology.

References

1. Rashidi A, Billingham LK, Zolp A, et al. Myeloid cell-derived creatine in the hypoxic niche promotes glioblastoma growth. Cell metabolism. 2024;36(1):62-77.e8. doi:https://doi.org/10.1016/j.cmet.2023.11.013
2. Agapito G, Milano M, Cannataro M. A Python Clustering Analysis Protocol of Genes Expression Data Sets. Genes. 2022;13(10):1839-1839. doi:https://doi.org/10.3390/genes13101839
3. Wang LB, Karpova A, Gritsenko MA, et al. Proteogenomic and metabolomic characterization of human glioblastoma. Cancer cell. 2021;39(4):509-528.e20. doi:https://doi.org/10.1016/j.ccell.2021.01.006
4. https://www.kdnuggets.com/2022/08/implementing-dbscan-python.html (code snippet)
5. https://www.cbioportal.org/
6. https://www.reneshbedre.com/blog/dbscan-python (code snippet)
7. https://matplotlib.org/stable/users/explain/colors/colormaps.html
8. Mullin T. DBSCAN Parameter Estimation Using Python - Tara Mullin - Medium. Medium. Published July 10, 2020. https://medium.com/@tarammullin/dbscan�parameter-estimation-ff8330e3a3bd