8.2 From Biological Measurements to Feature Spaces

8.2 From Biological Measurements to Feature Spaces

Learning Goals

After completing this section, you should be able to:

• Explain why biological objects must be represented mathematically before they can be analyzed computationally.
• Describe how experimental measurements are transformed into feature vectors.
• Interpret a gene expression matrix as a mathematical representation of biological behavior.
• Explain why the choice of representation determines what biological similarity means.

Before biological data can be analyzed computationally, they must first be represented in a mathematical form. This seemingly simple step is one of the most important modeling decisions in bioinformatics because every computational method operates on mathematical objects rather than on biological concepts themselves.

A gene, a protein, or an individual cell has no inherent mathematical representation. Instead, we choose a set of measurable properties—called features—that describe the biological object. Once these features have been selected, each biological object can be represented as a numerical vector, allowing algorithms to compare, classify, or cluster the data.

This idea has already appeared repeatedly throughout this book. In sequence analysis, biological sequences were represented as strings of nucleotides or amino acids. In phylogenetic reconstruction, similarity between sequences served as evidence of common evolutionary history. Clustering now extends this principle to much more general biological measurements. Rather than comparing sequences, we compare patterns of biological behavior.

Features Describe Biological Properties

The choice of features depends entirely on the biological question being addressed.

For gene expression experiments, the features are typically the measured expression levels under different experimental conditions or at different time points. Each measurement captures one aspect of how a gene responds to its environment.

Other applications use entirely different features. A protein might be represented by its amino acid composition, predicted secondary structure, hydrophobicity, or physicochemical properties. A single cell may be described by the expression levels of thousands of genes, while an organism could be characterized by physiological measurements or phenotypic traits.

The important point is that features are not discovered—they are chosen.

They represent our mathematical description of biological reality and therefore determine which aspects of biology become visible to the algorithm.

This illustrates a recurring theme throughout this book:

Every computational analysis begins with a modeling decision.

Before any algorithm is applied, we have already decided which biological information should be retained and which information should be ignored.

Expression Profiles as Feature Vectors

Consider a typical gene expression experiment.

Suppose we measure the expression level of a gene under six different environmental conditions. The measurements might correspond to different time points during development, different stress conditions, or different nutrient availabilities.

The resulting expression profile can be written as

$$\mathbf{x} = (x_1,x_2,\ldots,x_6),$$

where each component represents the measured expression level under one experimental condition.

Rather than thinking of the gene simply as a biological entity, we now interpret it as a point in a six-dimensional mathematical space.

This change in perspective is profound.

The biological question

How similarly do two genes behave?

has become the mathematical question

How similar are two vectors?

From this point onward, clustering is no longer concerned with genes directly. It operates entirely on their mathematical representations.

The Expression Matrix

Real biological experiments rarely measure only a few genes. Modern transcriptomic experiments routinely quantify the expression of tens of thousands of genes simultaneously.

The complete dataset is conveniently represented as an expression matrix.

In this matrix,

• each row corresponds to one gene,
• each column corresponds to one experimental condition,
• and each entry contains the measured expression value.

Mathematically, the expression matrix can be written as

$$X= \begin{pmatrix} x_{11} & x_{12} & \cdots & x_{1m}\\ x_{21} & x_{22} & \cdots & x_{2m}\\ \vdots & \vdots & \ddots & \vdots\\ x_{n1} & x_{n2} & \cdots & x_{nm} \end{pmatrix},$$

where (n) denotes the number of genes and (m) the number of experimental conditions.

Each row therefore represents one feature vector describing the behavior of a single gene.

This matrix forms the starting point for virtually all machine learning methods in molecular biology. Clustering, principal component analysis, classification algorithms, and neural networks all operate on this mathematical representation rather than on the original biological measurements.

Representation Determines Similarity

Once biological objects have been converted into feature vectors, the next challenge is to compare them.

At first glance, this appears straightforward. Two genes with nearly identical expression profiles should be considered similar, whereas genes with very different responses should not.

However, this raises a subtle but fundamental question:

What exactly does “similar” mean?

Should two genes be considered similar if their expression levels differ only by small numerical amounts?

Or should they be considered similar whenever they increase and decrease together, even if one gene is consistently expressed at much higher levels than the other?

These different interpretations lead to different mathematical definitions of similarity and, ultimately, to different clustering results.

In other words, similarity is not an intrinsic property of the data. It depends on the mathematical measure that we choose.

Choosing an appropriate similarity measure is therefore just as important as choosing the biological features themselves. Together, these two modeling decisions determine which biological patterns become visible and which remain hidden.

The next section explores this question in detail by introducing mathematical distance measures that quantify similarity between biological feature vectors.

Conceptual Summary

• Biological objects must be converted into mathematical representations before computational analysis is possible.
• Features define which biological properties are represented numerically.
• Gene expression profiles can be represented as feature vectors.
• An expression matrix provides a compact mathematical description of an entire experiment.
• The mathematical representation determines what biological similarity means and therefore influences every downstream analysis.

Self-Check Questions

1. Why must biological objects be represented as feature vectors before computational analysis?
2. What is the difference between a biological object and its mathematical representation?
3. How is a gene expression experiment represented as an expression matrix?
4. Why is the choice of features already a modeling decision?
5. Why does the mathematical representation influence the outcome of clustering?