Machine Learning in Genomics

Core Idea

Machine learning (ML) in genomics applies computational models to learn patterns from large biological datasets and make predictions. Applications include variant effect prediction (classifying variants as pathogenic or benign), gene expression prediction from DNA sequence, cell type classification from scRNA-seq data, protein structure prediction (AlphaFold), drug response prediction, and regulatory element identification. Deep learning models — particularly convolutional neural networks (CNNs) for sequence motif detection and transformers for long-range sequence dependencies — have achieved breakthroughs where handcrafted features and classical statistics fall short. Interpretability methods (attention maps, DeepLIFT, in silico mutagenesis) extract biological insights from trained models.

How It's Best Learned

Train a simple CNN to predict transcription factor binding from DNA sequence using a published ChIP-seq dataset. Visualize the learned convolutional filters and compare them to known binding motifs. Then deliberately overfit the model (too many parameters, no regularization) and observe how training versus validation performance diverges — this builds intuition for the bias-variance tradeoff in a genomics context.

Common Misconceptions

• A model with high accuracy on a test set is not necessarily biologically meaningful — it may have learned batch effects, label leakage, or confounders rather than true biological signal.
• Deep learning does not always outperform simpler methods; for small datasets or well-understood problems, logistic regression or random forests may perform equally well with better interpretability.

Explainer

Genomics generates datasets of a scale and complexity that strain traditional statistical methods. A human genome contains 3 billion positions, each of which could harbor a variant. A scRNA-seq experiment profiles 20,000 genes across 50,000 cells. An epigenomic atlas maps dozens of histone marks across hundreds of cell types. Machine learning provides the computational tools to find patterns in this data that manual analysis or classical statistics cannot.

Classical ML approaches — random forests, support vector machines, logistic regression, gradient boosting — remain widely used for structured genomic data. Variant pathogenicity prediction (tools like CADD) uses dozens of hand-engineered features (conservation scores, protein impact predictions, regulatory annotations) fed into ensemble classifiers. Gene expression prediction from genotype data uses penalized regression (LASSO, elastic net). Cell type classification from scRNA-seq uses random forests or SVMs on selected marker genes. These methods are interpretable, well-understood, and effective when the features are well-defined and the dataset is modest in size.

Deep learning has transformed problems where the raw data (DNA sequence, protein sequence, microscopy images) contains patterns that are difficult to capture with hand-engineered features. DeepBind and DeepSEA pioneered the use of CNNs for learning regulatory sequence grammar directly from ChIP-seq data. Enformer (a transformer architecture) predicts gene expression from 200 kb of surrounding DNA sequence, capturing distal regulatory effects that CNNs cannot reach. AlphaFold2 used a bespoke architecture to solve protein structure prediction. In each case, deep learning succeeded by learning representations from data rather than relying on human-specified features, and the learned representations often revealed new biology — motif syntax, regulatory grammar, and structural constraints that had not been previously recognized.

The critical challenge in genomic ML is evaluation and generalization. Genomic data has strong structure: genes are related by evolution, variants are correlated by linkage disequilibrium, and regulatory regions share sequence features. Naive random splitting of data into training and test sets can produce inflated performance estimates because related examples leak between splits. Proper evaluation requires biologically aware splitting: by chromosome (no chromosomal overlap), by gene family (no homologs in both sets), or by time (training on older data, testing on newer). Beyond prediction accuracy, interpretability methods — attention weights, saliency maps, in silico mutagenesis (systematically mutating input positions and observing the effect on prediction) — are essential for extracting biological insights and building confidence that the model has learned genuine biology rather than artifacts.