Real Research Applications

The AlphaFold Protein Structure Database

The most immediate application of AlphaFold was simply releasing its predictions, free and open, to the world. The AlphaFold Protein Structure Database, run jointly by DeepMind and EMBL-EBI, launched in July 2021 with predictions for ~365,000 proteins spanning the human proteome and 20 model organisms.

By July 2022 the database had been expanded to 214 million structures — covering essentially every sequence in UniProt. This included proteins from organisms across the tree of life: bacteria, archaea, fungi, plants, parasites, and animals. Many of these proteins had never had any experimental structural information at all.

Within the first year of the full 214M database, it had been accessed by over 1 million users across 190 countries, and its structures had appeared in thousands of published papers.

Drug discovery: structure-based design at scale

The traditional path to a drug target looks like this: identify a protein involved in disease → determine its 3-D structure experimentally → find a small molecule that fits into its active site like a key into a lock → optimise that molecule until it's a drug. The bottleneck has always been step 2: experimental structure determination.

AlphaFold removes that bottleneck for many targets. Now researchers can get a starting structure in hours rather than years, and use it immediately for structure-based drug design (SBDD) — computational docking of millions of candidate molecules, followed by prioritised experimental testing.

Isomorphic Labs

In 2021, Demis Hassabis co-founded Isomorphic Labs, a DeepMind spinout (now part of Google) dedicated to using AI for drug discovery. They use an expanded internal version of AlphaFold (including AlphaFold 3 capabilities) to predict how potential drugs bind to disease targets. In 2024, Isomorphic announced drug discovery partnerships with Eli Lilly and Novartis worth up to $3 billion combined — the largest AI drug discovery deals at that time.

Malaria vaccine research

Malaria kills approximately 600,000 people per year, mostly children under five in sub-Saharan Africa. The parasite Plasmodium falciparum has a particularly complex life cycle and has evolved numerous tricks to evade vaccines. AlphaFold predictions of malaria parasite proteins have helped researchers identify new vaccine antigen candidates by revealing previously hidden structural features that the immune system could target.

AlphaMissense: classifying genetic variants

In September 2023, DeepMind published AlphaMissense in Science — an AlphaFold-based model that predicts the pathogenicity (disease-causing potential) of all possible missense variants in the human genome.

To understand why this matters, we need to explain what a missense variant actually is.

What is a missense variant? (Explained from scratch)

Your DNA contains about 3 billion base pairs. The genetic code works like this: every three consecutive DNA bases (a "codon") specifies one amino acid. For example, the codon GCU codes for alanine, GAU codes for aspartate.

Sometimes a single DNA base is changed — a "T" becomes a "C", for instance. This is called a single nucleotide polymorphism (SNP). Most SNPs occur in non-coding regions and have no effect. But some fall within protein-coding regions and change the codon.

Three outcomes are possible:

• Synonymous (silent): The codon changes but still codes for the same amino acid (because the genetic code is redundant). GCU → GCC — both alanine. No change to the protein.
• Nonsense: The codon changes to a stop codon. The protein is truncated. Usually very harmful.
• Missense: The codon changes and codes for a different amino acid. GCU (Ala) → GAU (Asp). The protein now has one different amino acid at that position.

What AlphaMissense does

AlphaMissense takes an AlphaFold-based architecture and fine-tunes it to predict whether a given missense variant is "likely pathogenic" (disease-causing), "likely benign," or "uncertain." It does this by:

1. Using the evolutionary information (MSA) to assess whether the amino acid change is tolerated by nature — if no known organism uses that amino acid at that position, it's likely harmful.
2. Using the structural representation to assess the structural impact — does the mutation disrupt a hydrogen bond? Bury a charged residue? Destabilise the hydrophobic core?

AlphaMissense classified 71 million possible missense variants in the human proteome — providing a pathogenicity score for every single one. This is transformative for clinical genetics: when sequencing a patient's genome and finding a VUS, clinicians can now check AlphaMissense's prediction as one line of evidence about its likely significance.

Neglected tropical diseases

Neglected tropical diseases (NTDs) affect over 1 billion people, predominantly in low-income tropical countries. They include diseases like schistosomiasis (blood flukes), leishmaniasis (sand-fly parasites), Chagas disease (Trypanosoma cruzi), and sleeping sickness (Trypanosoma brucei). Despite their enormous disease burden, they receive a fraction of the global R&D investment compared to diseases prevalent in wealthy countries.

AlphaFold has been particularly impactful here because NTD pathogens are often poorly characterised — their proteomes are sequenced but few structures exist experimentally. The AlphaFold database provided structures for thousands of NTD pathogen proteins overnight, giving researchers a structural foundation for drug discovery that previously would have taken decades.

The Wellcome Trust and DNDI (Drugs for Neglected Diseases initiative) have used AlphaFold predictions to identify druggable pockets in schistosome and trypanosome proteins that had no prior structural characterisation.

Antibiotic resistance

Antimicrobial resistance (AMR) is projected to cause 10 million deaths per year by 2050, making it one of the most serious public health threats globally. Bacteria evolve resistance to antibiotics through several mechanisms, all involving changes to protein structure.

AlphaFold has been used to:

• Characterise resistance mechanisms: Predict structures of mutant proteins in resistant bacteria to understand exactly how mutations alter the antibiotic-binding site.
• Identify new drug targets: Survey the full proteomes of drug-resistant bacteria (like MRSA or extensively drug-resistant tuberculosis) for unexploited structural vulnerabilities.
• Design new antibiotics: Use predicted structures of essential bacterial proteins that have no human homologue (so drugs won't affect human cells) as templates for de novo drug design.

Agriculture & the environment

Heat-tolerant crops: the Rubisco problem

Rubisco (ribulose-1,5-bisphosphate carboxylase/oxygenase) is the enzyme responsible for carbon fixation in photosynthesis — arguably the most abundant protein on Earth, present in every plant and cyanobacterium. It's also notoriously inefficient: it's slow, it makes errors, and its activity declines at high temperatures.

Engineering more efficient or heat-tolerant versions of Rubisco has been a goal of agricultural biotechnology for decades. The challenge: Rubisco is a large, complex, multi-subunit enzyme that's very difficult to work with experimentally. AlphaFold predictions of Rubisco variants from extremophile organisms (bacteria that thrive in hot environments) have provided structural templates for engineering more thermostable versions in crop plants — with enormous potential implications for food security under climate change.

Plastic-degrading enzymes

In 2016, scientists discovered a bacterium (Ideonella sakaiensis) that could eat PET plastic (used in bottles) using an enzyme called PETase. The discovery was exciting but PETase was too slow for industrial use. Engineering it to be faster and more thermostable became a priority.

AlphaFold structures of PETase and related bacterial enzymes have enabled rational engineering: understanding which parts of the structure control substrate binding and catalysis, then making targeted mutations to improve activity. Teams at the University of Portsmouth and UT Austin have used this approach to create engineered PETases that work significantly faster than the wild-type enzyme — a step toward biological plastic recycling.