AI in Bioinformatics

Biology as an information problem

Bioinformatics treats DNA, RNA, and proteins as sequences of symbols — four-letter alphabets for nucleic acids, twenty-letter alphabets for amino acids — and applies computational methods to extract meaning from them. This framing makes biology tractable for AI: comparing two gene sequences becomes a string-matching problem, predicting a protein's shape becomes a structured prediction problem, and finding disease-associated genetic variants becomes a pattern-detection problem over enormous datasets.

The field exploded in importance once sequencing became cheap: the first human genome cost roughly three billion dollars to sequence in 2003, and today a genome can be sequenced for a few hundred dollars. That flood of data is exactly what modern machine learning needs to be useful, and it's why bioinformatics has become one of AI's most consequential application domains.

Sequence alignment: the oldest problem in the field

Sequence alignment asks: given two sequences (say, a gene from a human and the equivalent gene from a mouse), what is the best way to line them up to reveal their similarities, allowing for insertions, deletions, and substitutions? This matters because evolutionary relationships, shared function, and disease mechanisms are often revealed by how closely sequences match.

The classical solution is dynamic programming: algorithms like Needleman-Wunsch (global alignment) and Smith-Waterman (local alignment) build a scoring matrix cell by cell, where each cell represents the best possible alignment score up to that point, and trace back through it to recover the optimal alignment. This guarantees the mathematically best alignment under a given scoring scheme, but it's computationally expensive — true dynamic programming alignment scales with the product of the two sequence lengths, which is too slow for searching a sequence against an entire genome database.

The protein folding problem

A protein is built from a linear chain of amino acids, but it only becomes functional once it folds into a specific three-dimensional shape — and that shape determines almost everything about what the protein does, from catalyzing reactions to binding other molecules. For fifty years, predicting a protein's 3D structure directly from its amino acid sequence (without expensive lab techniques like X-ray crystallography or cryo-electron microscopy) was one of biology's grand unsolved problems, known as the protein folding problem.

The core difficulty is the astronomical number of possible foldings: a modest-sized protein chain has more possible configurations than there are atoms in the observable universe, yet real proteins reliably fold into the same shape in fractions of a second — a paradox (Levinthal's paradox) that hinted the folding process follows strong underlying rules, even if those rules weren't understood well enough to simulate directly.

How AlphaFold cracked it (conceptually)

AlphaFold, developed by DeepMind, achieved a breakthrough at the 2020 CASP competition (the field's standardized blind-test benchmark), reaching prediction accuracy competitive with experimental methods for the first time. At a conceptual level, it works by treating structure prediction as a learned pattern-matching problem rather than a physics simulation: it takes a protein's amino acid sequence, searches large databases for evolutionarily related sequences across many species (a multiple sequence alignment), and uses co-evolutionary signal — the statistical fact that amino acids which end up physically close in the folded structure tend to mutate together across evolutionary history — as a strong clue about which parts of the chain fold near each other.

A deep neural network architecture (built around attention mechanisms similar to those in transformers) processes this evolutionary and geometric information jointly, iteratively refining a 3D structure prediction along with a confidence estimate for each predicted region, rather than simulating the physical folding process step by step.

Experimental vs. computational structure determination

Genomics applications beyond structure

Structure prediction is the highest-profile result, but AI touches most of modern genomics. Variant calling pipelines use machine learning to distinguish real genetic mutations from sequencing errors in raw sequencer output. Gene expression analysis uses clustering and dimensionality reduction (similar to techniques covered in PCA and clustering) to identify cell types from single-cell RNA sequencing data. Drug discovery pipelines use AI models to predict how candidate molecules will bind to a target protein's predicted structure, dramatically narrowing the search space before expensive lab synthesis and testing. And disease risk prediction models combine genomic variants with clinical data to estimate individual risk for conditions with a genetic component.

What's next

AlphaFold's architecture leans heavily on attention mechanisms originally developed for language. The Transformers module covers that underlying architecture in depth — useful context for understanding why the same core idea keeps resurfacing across language, vision, and now biology.