AI closes in on biology's last structural puzzles

Overview

Determining the three-dimensional shape of a single protein used to take a graduate student an entire career. In March 2026, Google DeepMind released AlphaFold 4, a system that predicts the structure of multi-protein complexes in minutes rather than hours or days, extending a streak of capability gains that began when the original AlphaFold won a protein-prediction competition in 2018. Weeks earlier, DeepMind's drug-discovery spinoff Isomorphic Labs disclosed a proprietary model called IsoDDE that more than doubles AlphaFold 3's accuracy on key drug-design benchmarks, while a team at the National University of Singapore published D-I-TASSER, a tool that outperforms both AlphaFold 2 and AlphaFold 3 on difficult multi-domain proteins.

In This Story

3 Videos

5 People

3 Orgs

AI-designed drugs reach patients within five years

3 Context

Human Genome Project (1990-2003)

Key Indicators

214M+

Protein structures predicted

AlphaFold's public database now covers virtually every catalogued protein known to science

IsoDDE accuracy gain over AlphaFold 3

Isomorphic Labs' proprietary model more than doubles AlphaFold 3's accuracy on protein-ligand benchmarks

2M+

Researchers using AlphaFold

Scientists across 190 countries have used the AlphaFold database for published research

$3.3B

AI drug discovery funding in 2024

Venture capital invested in artificial intelligence-driven drug discovery startups in a single year

Interactive

Exploring all sides of a story is often best achieved with Play.

Ever wondered what historical figures would say about today's headlines?

Debate Arena

Two rounds, two personas, one winner. You set the crossfire.

People Involved

Demis Hassabis

Chief Executive Officer, Google DeepMind and Isomorphic Labs (2024 Nobel laureate in Chemistry; leads both DeepMind's research and its drug-discovery spinoff)

John Jumper

Senior Research Scientist, Google DeepMind (2024 Nobel laureate in Chemistry; leads AlphaFold development)

David Baker

Director, Institute for Protein Design, University of Washington (2024 Nobel laureate in Chemistry; leads open-source protein design tools)

Mohammed AlQuraishi

Computational biologist, Columbia University (Developing open-source alternatives to AlphaFold)

Zhang Yang

Professor, National University of Singapore (Lead developer of D-I-TASSER protein prediction pipeline)

Organizations Involved

Google DeepMind

AI Research Lab / Internal Security Team

Status: Developer of AlphaFold series; released AlphaFold 4 in March 2026

Alphabet's artificial intelligence research laboratory, responsible for developing the AlphaFold series of protein structure prediction models.

Isomorphic Labs

Biotech / AI Drug Discovery

Status: Released proprietary IsoDDE model in February 2026; partnerships with major pharmaceutical companies

An Alphabet-owned drug discovery company that commercializes protein structure prediction technology developed by Google DeepMind.

European Molecular Biology Laboratory - European Bioinformatics Institute (EMBL-EBI)

Research Institution

Status: Hosts the AlphaFold Protein Structure Database

A European intergovernmental research organization that partners with Google DeepMind to host and distribute the AlphaFold database of over 214 million predicted protein structures.

Timeline

AlphaFold 4 predicts protein complexes in minutes
Breakthrough

Google DeepMind released AlphaFold 4, which predicts the structure of multi-protein complexes in minutes rather than the hours or days required by its predecessor, using an improved neural network architecture and expanded training data.
2 days ago
D-I-TASSER outperforms AlphaFold on complex proteins
Research

A team at the National University of Singapore published D-I-TASSER, a hybrid deep-learning and physics-based tool that predicts complex multi-domain protein structures about 13 percent more accurately than AlphaFold 3.
February 25th, 2026
Isomorphic Labs unveils IsoDDE, a proprietary 'AlphaFold 4'
Breakthrough

Isomorphic Labs released a technical report describing IsoDDE, a proprietary drug design engine that more than doubles AlphaFold 3's accuracy on protein-ligand benchmarks. Unlike previous AlphaFold versions, IsoDDE is not publicly available.
February 10th, 2026
Boltz-1 matches AlphaFold 3 performance as fully open model
Release

An MIT lab released Boltz-1 under an MIT license with full training code and model weights, matching AlphaFold 3 and Chai-1 across key protein prediction benchmarks.
November 1st, 2024
Nobel Prize in Chemistry awarded for protein prediction
Recognition

Demis Hassabis and John Jumper shared half the Nobel Prize in Chemistry for AlphaFold; David Baker received the other half for computational protein design.
October 9th, 2024
Chai-1 released as open-source AlphaFold competitor
Release

Chai Discovery released Chai-1 under an Apache 2.0 license, providing an open-source alternative to AlphaFold 3 for academic and commercial use.
September 1st, 2024
AlphaFold 3 predicts multi-molecule interactions
Breakthrough

Google DeepMind and Isomorphic Labs released AlphaFold 3, capable of predicting how proteins interact with DNA, RNA, small molecules, and ions, with at least 50 percent better accuracy than existing methods.
May 8th, 2024
AlphaFold database expands to 200 million structures
Release

DeepMind and EMBL-EBI uploaded predicted structures for nearly every catalogued protein known to science, covering approximately one million species.
July 28th, 2022
Isomorphic Labs founded
Corporate

Demis Hassabis launched Isomorphic Labs as an Alphabet-owned company to apply AlphaFold technology to commercial drug discovery.
November 4th, 2021
AlphaFold database launches with 350,000 structures
Release

Google DeepMind and EMBL-EBI launched the AlphaFold Protein Structure Database, initially covering the entire human proteome and 20 model organisms.
July 22nd, 2021
AlphaFold 2 solves the protein folding problem
Breakthrough

AlphaFold 2 achieved atomic-level accuracy at CASP14, predicting protein structures with a median error of less than one angstrom. Scientists declared the 50-year-old protein folding problem effectively solved.
November 30th, 2020
AlphaFold 1 wins CASP13
Breakthrough

DeepMind's AlphaFold placed first in the 13th Critical Assessment of Structure Prediction competition, outperforming all other computational methods for predicting protein structures from amino acid sequences.
December 2nd, 2018

Scenarios

AI-designed drugs reach patients within five years

Discussed by: Nature Biotechnology, pharmaceutical industry analysts, Isomorphic Labs leadership

If AlphaFold 4 and IsoDDE deliver on their accuracy claims at scale, AI-driven drug candidates could move from target identification to Phase I clinical trials in under 18 months, as Insilico Medicine has already demonstrated with its fibrosis candidate. Major pharmaceutical partners like Eli Lilly and Novartis are already integrating these tools into their pipelines. Under this scenario, the first wave of AI-designed drugs would reach regulatory approval by 2030, potentially cutting average development costs from $2 billion to under $1 billion per approved medicine.

Open-source models close the gap, keeping drug design accessible

Discussed by: Mohammed AlQuraishi (Columbia University), Boltz-1 developers, open-science advocates in Nature and Science commentary

Open-source alternatives like Boltz-1, Chai-1, and OpenFold 3 already match AlphaFold 3's performance. If the open-source community can reverse-engineer the architectural improvements behind IsoDDE and AlphaFold 4, academic labs and smaller biotech firms would retain access to state-of-the-art drug design tools. This would distribute innovation more broadly but reduce the competitive advantage that companies like Isomorphic Labs derive from proprietary models.

Accuracy ceiling limits real-world drug design impact

Discussed by: Computational chemistry researchers, independent benchmarking studies in Journal of Chemical Information and Modeling

Protein structure prediction and actual drug design are different problems. Even with high accuracy on static structure prediction, these models may struggle with protein dynamics, the behavior of molecules in cellular environments, off-target binding, and the complex pharmacokinetics that determine whether a drug actually works in patients. Under this scenario, AI accelerates early-stage research but fails to meaningfully reduce late-stage clinical trial failure rates, which currently hover around 80-90 percent.

Proprietary AI creates a two-tier research ecosystem

Discussed by: Nature editorial board, open-access research advocates, university computational biology departments

If Isomorphic Labs' proprietary model represents the new frontier and no open-source equivalent emerges, drug discovery would increasingly concentrate among well-capitalized pharmaceutical companies and their AI partners. Academic researchers and institutions in lower-income countries, which benefited enormously from the free AlphaFold database, would work with less capable tools. This would accelerate commercial drug development but slow the kind of basic research that drives unexpected breakthroughs.

Historical Context

Human Genome Project (1990-2003)

1990-2003

What Happened

A 13-year, $2.7 billion international effort sequenced the roughly 3 billion base pairs of human DNA. The project competed with Craig Venter's privately funded Celera Genomics, which used a faster shotgun sequencing method and announced a draft genome simultaneously with the public effort in 2000.

Outcome

Short Term

Both the public consortium and Celera published draft genomes in February 2001. The public data was made freely available, while Celera initially restricted access through a subscription model.

Long Term

The open-access model won. Free genome data enabled thousands of studies that Celera's proprietary model could not have supported. The project catalyzed the genomics industry, drove DNA sequencing costs from $100 million per genome in 2001 to under $200 by 2024, and established a precedent for open science in biology.

Why It's Relevant Today

The tension between Isomorphic Labs' proprietary IsoDDE and the open AlphaFold database mirrors the Celera-versus-public-consortium dynamic. The genome project showed that open access to foundational biological data produces more aggregate innovation, a lesson that shapes the current debate over whether AI drug design tools should be public goods.

CRISPR gene editing revolution (2012-present)

2012-present

What Happened

Jennifer Doudna and Emmanuelle Charpentier published their discovery that a bacterial immune system component, CRISPR-Cas9, could be reprogrammed to cut any DNA sequence with precision. The technology reduced the cost and time of genetic editing from months and thousands of dollars to days and a few hundred dollars.

Outcome

Short Term

The discovery triggered a patent battle between the Broad Institute and UC Berkeley, a 2020 Nobel Prize in Chemistry for Doudna and Charpentier, and rapid commercialization by companies like Editas Medicine and CRISPR Therapeutics.

Long Term

In December 2023, the United States Food and Drug Administration approved the first CRISPR-based therapy, Casgevy, for sickle cell disease. The technology moved from laboratory discovery to approved medicine in roughly 11 years, a timeline that AI-driven drug design aims to compress further.

Why It's Relevant Today

Like CRISPR, AI protein prediction represents a platform technology that dramatically lowers the barrier to biological manipulation. Both compress timelines from years to days and raise questions about equitable access. The CRISPR patent wars offer a preview of the intellectual property disputes likely to emerge as proprietary AI models become central to drug development.

X-ray crystallography and the first protein structure (1958)

1958

What Happened

John Kendrew and Max Perutz at Cambridge used X-ray crystallography to determine the three-dimensional structure of myoglobin, a protein that stores oxygen in muscle tissue. The work required growing protein crystals, bombarding them with X-rays, and painstakingly interpreting the resulting diffraction patterns. It took Kendrew roughly two decades of effort.

Outcome

Short Term

Kendrew and Perutz received the 1962 Nobel Prize in Chemistry. X-ray crystallography became the dominant method for determining protein structures for the next six decades.

Long Term

By 2020, researchers had determined approximately 170,000 protein structures through experimental methods. That number represented a tiny fraction of the estimated 200 million known proteins, illustrating the bottleneck that AlphaFold ultimately broke.

Why It's Relevant Today

The contrast in scale is the point. What took Kendrew 20 years for a single protein, AlphaFold 4 now accomplishes for multi-protein complexes in minutes. The roughly 170,000 experimentally determined structures over six decades compare to 214 million computationally predicted structures released in a single database expansion.