Computational Biology

A dynamic-programming algorithm for global alignment of two biological sequences.

The first proper algorithm of computational biology; conceptual ancestor of every aligner.

Dynamic programming basics

Optimal sequence alignment is solvable in O(mn) time with a single DP table.

Adapts Needleman–Wunsch to local alignment by allowing zero scores in the DP table.

A small modification with huge consequences for finding conserved motifs and homology.

Needleman–Wunsch

Local alignment is just global alignment with a non-negativity constraint.

A heuristic algorithm trading some sensitivity for orders-of-magnitude speed over Smith–Waterman.

The most-used algorithm in all of biology; the workhorse of every database search.

Smith–Waterman, basic indexing

A statistically-grounded heuristic can be the right answer when exact is too slow.

Builds a hidden Markov model per protein family for sensitive remote-homology searches.

The standard probabilistic framework for protein family detection.

BLAST, HMM basics

A probabilistic model of a family beats a single best representative.

Indexes the reference genome via the Burrows–Wheeler transform for fast NGS read alignment.

One of the most-used aligners in modern genomics — the BWT made it possible.

BLAST, suffix arrays

A clever string-data structure makes a previously intractable problem trivial.

Another BWT-based aligner with a focus on memory footprint and speed.

A useful contrast to BWA; both papers together define the modern read-alignment toolbox.

BWA, suffix arrays

There's often more than one good way to apply a powerful new data structure.

Uses de Bruijn graphs to assemble genomes from millions of short sequencing reads.

The de Bruijn graph idea is the conceptual core of every modern assembler.

Graph algorithms, basic genomics

Genome assembly is an Eulerian path problem on a graph of k-mers.

A spliced aligner that handles RNA-seq reads spanning exon-exon junctions efficiently.

The de facto standard RNA-seq aligner — read it before any expression analysis paper.

BWA, basic transcriptomics

RNA-seq needs splice-aware alignment, not just genome-aware alignment.

A statistical framework for differential expression with shrinkage estimators for low-count noise.

The most-cited paper in modern transcriptomics — methodologically influential beyond its field.

Negative binomial distribution, basic GLMs

Empirical-Bayes shrinkage gives stable inferences from noisy count data.

A dimensionality-reduction algorithm with a manifold-learning theoretical foundation.

Quietly displaced t-SNE in single-cell biology and elsewhere — a major workflow shift.

PCA, basic topology useful

Better assumptions about the data manifold give better low-dimensional embeddings.

A statistical framework for integrating single-cell datasets across conditions and modalities.

The standard analysis pipeline for one of the most-active areas of biology.

PCA, basic single-cell biology

Anchor-based integration lets you compare cells across very different experiments.

Demonstrates droplet-based single-cell RNA sequencing at hundreds of thousands of cells per run.

The technology paper that made single-cell biology a mainstream field.

RNA-seq basics

Massive throughput on cells, not just reads, opens new biological questions.

A geometry-aware Transformer architecture (Evoformer + structure module) predicting 3D protein structure to atomic accuracy.

A scientific landmark of historic magnitude; the protein-folding problem largely solved.

Transformers, basic structural biology

Hard scientific problems crack when you bake the right geometric and biological priors into the model.

Uses a 15B-parameter protein language model (ESM-2) to predict structure directly from sequence.

Faster, simpler alternative to AlphaFold for many use cases; demonstrates the power of protein LLMs.

AlphaFold, protein language models

A big enough language model implicitly learns structure from sequence alone.

A three-track network reasoning over sequence, distance, and structure simultaneously.

A useful counterpoint to AlphaFold; same problem, related but distinct architecture.

AlphaFold, basic graph nets

Structure prediction benefits from reasoning over multiple representations in parallel.

A graph neural network that designs protein sequences to fold into a desired backbone.

A modern reference for inverse folding and the design half of computational protein engineering.

AlphaFold, GNNs

Inverse design (give me a sequence for this shape) is the dual problem to folding.

A diffusion model that generates novel protein backbones for specified geometric constraints.

The "Stable Diffusion of biology" — generative protein design now actually works.

Diffusion models, protein design basics

Diffusion is a remarkably general way to generate structured biological objects.

A Transformer pretrained on 30M single-cell transcriptomes and fine-tuned for downstream biology.

Foundation-model thinking finally hitting cell biology.

Transformers, single-cell RNA-seq

Pretraining on rank-encoded gene expression yields strong biological transferability.

Extends AlphaFold to predict protein–nucleic-acid, protein–ligand, and protein–protein complexes.

The current state of the art; redefines what computational structural biology can do.

AlphaFold, diffusion models

A single foundation model can handle proteins, RNA, DNA, and small-molecule interactions.

A protein language model designs novel functional proteins beyond known evolutionary distributions.

A glimpse of foundation-model-driven exploration of biological design space.

ESM, protein design basics

Language models trained on natural proteins can imagine functional unnatural ones.