8.1 Why Learning Begins with Unlabeled Data

Chapter 8 — Clustering: Discovering Structure in Biological Data

8.1 Why Learning Begins with Unlabeled Data

Learning Goals

After completing this section, you should be able to:

• Explain why modern biology increasingly relies on data-driven discovery.
• Distinguish between supervised and unsupervised learning.
• Describe the biological motivation behind clustering.
• Explain why similarity can reveal functional relationships between biological objects.

Modern biology has entered an era of unprecedented data generation. Advances in next-generation sequencing, mass spectrometry, high-content microscopy, and single-cell technologies allow researchers to measure thousands—or even millions—of biological features simultaneously. Entire transcriptomes, proteomes, metabolomes, or populations of individual cells can now be characterized in a single experiment.

This remarkable technological progress has fundamentally changed the nature of biological research. For many decades, the primary challenge was obtaining sufficient experimental data. Today, the challenge is increasingly to interpret the enormous amounts of data that modern technologies produce. Biological discovery has therefore become a problem of learning from data.

Unlike many classical scientific problems, however, these datasets rarely come with predefined answers. When measuring the expression of thousands of genes, we usually do not know which genes participate in the same biological pathway. When profiling individual cells, we often do not know how many cell types are present. Similarly, in proteomics experiments, we typically cannot identify functional protein complexes directly from the measurements.

Instead, the data themselves must reveal their internal organization.

This is the domain of unsupervised learning.

In supervised learning, which will be introduced later in this part of the book, algorithms learn from examples with known answers. Each observation is associated with a label—for example, a disease diagnosis, a protein localization, or an enzyme class—and the goal is to predict these labels for new observations.

Unsupervised learning addresses a fundamentally different problem. Here, no labels are available. Instead of predicting known outcomes, the objective is to discover hidden patterns, structures, or relationships directly from the data. Rather than confirming existing knowledge, unsupervised learning generates new hypotheses by revealing previously unknown organization within biological systems.

Among all unsupervised learning methods, clustering is arguably the most fundamental. Its purpose is to identify groups of biological objects that are more similar to each other than to the rest of the dataset.

At first sight, this objective appears almost trivial. If two genes exhibit nearly identical expression profiles, it seems reasonable to assume that they perform similar biological functions. Yet this immediately raises two important questions.

First, what does it actually mean for two biological objects to be similar?

Second, how can similarity be transformed into biological knowledge?

These questions lie at the heart of clustering and, more generally, of data-driven biology. Throughout this chapter, we will see that the answers are far from obvious. Similarity is not an intrinsic property of biological objects but a consequence of how we choose to represent them mathematically. Likewise, clusters are not objective entities waiting to be discovered; they emerge from specific mathematical assumptions about biological organization.

Understanding these assumptions is essential for interpreting clustering results correctly. Rather than treating clustering as a black-box algorithm, we will approach it as a mathematical model for discovering functional organization in biological systems.

Conceptual Summary

• Modern biology produces large, high-dimensional datasets that require computational analysis.
• Many biological datasets are unlabeled, making unsupervised learning essential.
• Clustering aims to discover groups of similar biological objects without prior knowledge.
• The biological interpretation of clusters depends on how similarity is defined.
• Clustering generates hypotheses about biological organization rather than proving biological relationships.

Self-Check Questions

1. Why has modern biology become increasingly dependent on data-driven analysis?
2. What distinguishes supervised from unsupervised learning?
3. Why is clustering considered an exploratory method?
4. Why should clustering results be interpreted as hypotheses rather than definitive biological conclusions?