Questions: Machine Learning in Genomics

Question 1

Why are convolutional neural networks (CNNs) particularly well-suited for learning patterns in DNA sequences?

• A. Because DNA sequences are always the same length
• B. Because CNNs can detect local sequence patterns (motifs) regardless of their position in the input sequence
• C. Because CNNs require less training data than other methods
• D. Because DNA has exactly four bases, which matches the typical CNN architecture

Answer: (B)

Transcription factor binding sites and other functional sequence elements are short motifs (typically 6-20 bases) that can occur at various positions within a regulatory region. CNNs use learned convolutional filters that slide along the input sequence, detecting these motifs wherever they appear — a property called translation invariance. This mirrors how biology works: a TATA box functions regardless of whether it is at position 50 or position 200 in the input. The learned filters often correspond directly to known transcription factor binding motifs, providing interpretability.

Question 2

A deep learning model for predicting gene expression from DNA sequence achieves 99% accuracy on the training data. This means the model has successfully learned the true biological relationship.

• T. True
• F. False

Answer: False

99% training accuracy with no evaluation on held-out data is a classic warning sign of overfitting — the model may have memorized the training examples (including noise) rather than learning generalizable patterns. Performance must be evaluated on independent test data that the model never saw during training. Furthermore, even good test performance can be misleading if the data split was not done properly (e.g., if homologous genes or linked genomic regions appear in both training and test sets, creating data leakage).

Question 3

Explain why interpretability is particularly important for machine learning models applied to genomic data, compared to many other ML applications.

Model answer: In genomics, the goal is usually not just prediction but biological understanding — we want to know which sequence features, variants, or regulatory elements drive the prediction. An uninterpretable model may achieve high accuracy but provides no biological insight, which limits its scientific value and makes it difficult to validate or trust its predictions for clinical applications. Interpretability methods (visualizing learned filters, computing feature importance, performing in silico mutagenesis) can reveal which sequence motifs the model has learned, whether they correspond to known biology, and whether the model is using biologically meaningful features or exploiting artifacts. This is essential for building trust and generating testable hypotheses.

Clinical applications raise the stakes further. A variant pathogenicity predictor used in genetic diagnosis must be interpretable enough for clinicians to understand and evaluate its reasoning. Black-box predictions, however accurate, face barriers to clinical adoption if they cannot be explained.