Skip to content

Machine Learning Task Classification in Genomics

Understanding how biological problems map to different machine learning (ML) task types is essential for choosing the right models, loss functions, and evaluation metrics. In genomics, four major ML task categories cover most practical applications.

Four Main Task Types

Task Type Description Biological Example Output Format
Sequence Classification Assign one or more labels to an entire sequence Is this DNA sequence a promoter? Single label or label set
Sequence Regression Predict continuous numerical values for a sequence Predicting protein stability score (0.0–1.0) Continuous values
Token Classification Assign a label to each nucleotide/amino acid in the sequence Identifying transcription factor binding sites One label per position
Sequence-to-Sequence Transform an input sequence into a different output sequence RNA sequence → RNA secondary structure Output sequence

These categories mirror classical ML tasks but are adapted to the structured, position-dependent nature of biological sequences.

Examples of Common Genomic ML Tasks

Sequence Classification Tasks

  • Translation efficiency: High vs Low
  • Transcription factor binding: Binding vs Non-binding
  • Promoter identification: Promoter vs Background
  • Subcellular localization: Nucleus vs Cytoplasm vs Mitochondria

Sequence Regression Tasks

  • Gene expression levels (e.g., FPKM prediction)
  • Protein stability (ΔΔG value prediction)
  • Binding affinity (Kd prediction)
  • Enzyme activity (kcat prediction)

Token Classification Tasks

  • Transcription factor binding site probability per nucleotide
  • Splice site boundary prediction
  • RNA modification sites (e.g., m6A position identification)
  • Protein secondary structure (α-helix, β-sheet, random coil)

Sequence-to-Sequence Tasks

  • RNA secondary structure prediction: Sequence → Structure notation
  • Protein design: Function description → Amino acid sequence
  • Sequence optimization: Wild-type → Optimized sequence
  • Reverse translation: Protein sequence → DNA codon sequence

How to Select the Correct Task Type

1. Ask: What do I want to predict?

Properties of an entire sequence
→ Sequence classification or regression
Examples: promoter vs non-promoter, expression level prediction

Properties of specific positions
→ Token classification
Examples: splice sites, binding motifs

Entirely new sequence as output
→ Sequence-to-sequence
Examples: RNA structure prediction, protein design

2. Consider the Output Type

  • Category labels (Yes/No, High/Low, etc.)
    → Classification

  • Continuous numerical values (0.1, 2.5, etc.)
    → Regression

  • Multiple labels for the entire sequence
    → Multi-label classification

  • One label per nucleotide/residue
    → Token classification

Summary

By classifying biological problems into one of these four ML task types, you can systematically decide:

  • Which model architecture is suitable
  • What loss function to use
  • Which evaluation metrics matter
  • How to structure your dataset

This task-aware approach ensures more efficient experimentation and clearer scientific interpretation.