AlphaFold Is Free. Patient-Led Labs Are Using It.

In November 2025, DeepMind reported that AlphaFold was being used by more than three million researchers across 190 countries, and that roughly 30 percent of papers citing the underlying Nature articles were focused on disease research. The same month marked roughly eighteen months since the May 2024 release of AlphaFold 3 and its companion AlphaFold Server, which made structural prediction of protein, DNA, RNA, ligand, and antibody interactions free for academic and non-commercial use. In October 2024 the underlying science earned Demis Hassabis and John Jumper the Nobel Prize in Chemistry, alongside David Baker of Rosetta.

The headline numbers are correct and the Nobel is real. The more interesting story is what happens when a structural-biology tool that previously required a cryo-EM facility, a dedicated wet lab, and a postdoc to drive it becomes a free web form anyone with an institutional email can use. The answer, increasingly, is that patients and the small academic groups they fund are running the searches themselves.

This is a use-case profile of what that looks like in practice. The cases below are documented in peer-reviewed papers, New England Journal of Medicine case reports, and ARPA-H grant announcements. They are not "AlphaFold will cure cancer." They are narrower than that, and more interesting because of it.

The wedge: open weights, open database, free server, $0 marginal cost

Before unpacking specific cases, the structural change worth naming clearly.

AlphaFold Protein Structure Database (the open database hosted by EMBL-EBI) covered roughly 214 million unique protein sequences by the 2024 Nucleic Acids Research update, with predicted structures freely downloadable and re-usable under CC-BY-4.0. That database alone removed the single largest barrier to entry in structural biology for any protein of interest: the first thing a rare-disease parent researcher used to need was a structure to look at, and getting one experimentally cost anywhere from $50,000 for a tractable globular protein up through low millions for a difficult membrane complex. The cost is now zero, the wait is now seconds, and the file format is the same PDB that every downstream tool reads.

AlphaFold 3 (May 2024) extended the prediction surface beyond proteins-alone to proteins with DNA, RNA, ions, ligands, and post-translational modifications. The AlphaFold Server runs it through a browser. There is a daily request cap and a non-commercial license, but for a patient advocacy group running ten or twenty searches a week against candidate drug compounds, the cap is not the binding constraint.

The third piece is the open-weights ecosystem that grew around AlphaFold. ColabFold (Mirdita et al., Nature Methods 2022) made AlphaFold 2 runnable on a free Google Colab GPU. OpenFold, ESMFold, and RoseTTAFold each provide alternative paths to similar outputs, often with different speed/accuracy tradeoffs that suit different rare-disease use cases (long disordered regions, membrane proteins, antibody loops). The total cost of running a structure prediction against a candidate variant for a patient's specific mutation is now measured in pennies of compute, not dollars of facility time.

That is the wedge. Whether anyone outside of pharma actually picked it up is the empirical question.

Case one: the IAHSP/Alsin work — small lab, rare disease, AlphaFold-driven mechanism hunt

Infantile-Onset Ascending Hereditary Spastic Paralysis is a motor neuron disease so rare it has fewer than a hundred documented cases worldwide. It is caused by recessive mutations in ALS2, the gene encoding Alsin, a large multi-domain protein that acts as a guanine nucleotide exchange factor for Rab5 and plays a role in endosomal trafficking in neurons. The clinical course is brutal — onset before age two, progressive paralysis, full quadriplegia by adolescence. There is no approved disease-modifying therapy.

In December 2021, a research group led by Sara Pellegrini and colleagues published a perspective in Drug Discovery Today titled "AI-based protein structure databases have the potential to accelerate rare diseases research: AlphaFoldDB and the case of IAHSP/Alsin." The paper made an argument that has aged well: Alsin had never been experimentally crystallized, no structural data existed in the PDB for the full-length protein, and the dozens of patient-derived missense mutations scattered across its four domains could not be interpreted mechanistically without one. AlphaFoldDB had quietly dropped a predicted full-length structure into the public database. The Pellegrini paper walked through what that single free file enabled: domain-by-domain flexibility profiling, mapping of every known IAHSP mutation onto a structural model, and a first-pass hypothesis ranking for which variants likely destabilized the protein versus disrupted specific protein-protein interfaces.

That work has continued. In 2025, an ACS Omega paper from a different group (Rampini et al.) used the AlphaFold model as the input for molecular dynamics simulations of the R1611W mutation in the VPS9 domain — the same variant identified in the Pellegrini perspective as one of the most studied IAHSP-causing mutations. The work characterized how the mutation alters Alsin's oligomeric state and disrupts its GEF function on Rab5. That is the kind of mechanistic detail that drug discovery programs use to decide which patient subset to target, which downstream pathway to drug, and whether a small molecule or an antisense oligonucleotide is the right modality.

Neither of those papers came from a billion-dollar biopharma. They came from academic groups working on a disease that has fewer patients than a single mid-sized US town, using a freely downloaded structure file that did not exist three years earlier.

Case two: the n-of-1 ASO pipeline — Stargardt, Batten, and the ABCA4 variant

N-of-1 antisense oligonucleotide therapies are the most aggressive expression of patient-led rare disease research. The model was made public by the Milasen case at Boston Children's, where Timothy Yu's lab designed an ASO targeted at one girl's private splicing variant in the CLN7 gene causing Batten disease, shipped it under an FDA expanded-access pathway, and treated her within roughly a year of identifying the variant. The cost was largely borne by the family and a foundation they assembled.

The Milasen model is now being extended to other ultrarare splicing variants. A 2024 Nucleic Acid Therapeutics case report described preclinical development of ASOs to rescue aberrant splicing caused by an ultrarare ABCA4 variant in a child with early-onset Stargardt disease. AlphaFold structures are not the centerpiece of every n-of-1 pipeline — splice-modulating ASOs care more about the RNA sequence than the protein structure — but predicted structures help in two specific ways:

• When the variant is a missense rather than a splice site, the predicted structure tells the researcher whether the residue sits at a folding core (likely loss-of-function), an active site (likely catalytically dead), or a protein-protein interface (potentially druggable with a different modality).
• When a candidate small molecule is being considered to compensate for the loss-of-function, the predicted structure of the wild-type protein in complex with the candidate, generated through AlphaFold 3, is the cheapest possible first filter before committing wet-lab time.

The honest framing is that AlphaFold does not turn a parent into a drug developer. It removes one of the steps that used to require institutional infrastructure, so the parent-funded research group can ask better questions of the contract research organizations they hire downstream. That is a real shift in who gets to participate.

Case three: the Castleman / Every Cure adjacency — AI-driven repurposing, parallel model

The case worth naming carefully is David Fajgenbaum's work at Every Cure. Fajgenbaum is a physician-scientist at Penn who nearly died of idiopathic multicentric Castleman disease as a medical student and saved his own life by repurposing sirolimus, an immunosuppressant originally approved for kidney transplant patients. He has been in remission for more than a decade and has built a research program around the proposition that thousands of already-approved drugs have hidden indications nobody has tested.

In February 2025, his team published a case report in the New England Journal of Medicine describing the use of an AI platform to screen roughly 4,000 approved drugs against iMCD biology. The system identified adalimumab (Humira), a TNF-alpha inhibitor approved for rheumatoid arthritis and Crohn's disease, as a high-ranked candidate. A patient with treatment-refractory iMCD entering hospice was treated with adalimumab and entered remission. The patient is now approximately two years into remission as of the published report.

The adjacency to AlphaFold is worth being precise about. The Every Cure platform is built primarily on a knowledge graph approach — drug-target-disease relationships pulled from the literature and ranked by machine-learning models — not on de-novo structural prediction. AlphaFold and AlphaFold 3 are complementary to that pipeline, not the centerpiece of it. The Castleman / adalimumab result is the proof-of-concept that AI-driven drug repurposing can save the life of a specific patient with a specific rare disease, working from a specific data substrate. It is not a result driven by AlphaFold structures. The reason to include it in this profile is that it establishes the operating model — patient-led research organization, AI platform, repurposing-first hypothesis, NEJM case report, expanding clinical trial pipeline — that AlphaFold-driven groups are now copying.

Every Cure received a three-year, $48.3 million ARPA-H contract announced in February 2024 to scale this work into the MATRIX program. The funding is public and the trial pipeline includes a JAK1/2 inhibitor study for iMCD planned for 2025. This is what the patient-led model looks like when it succeeds at attracting federal infrastructure.

Case four: the small-molecule hit identification proof of concept

The first peer-reviewed paper showing AlphaFold structures driving end-to-end small-molecule drug discovery against a target that had no experimental structure was published by Insilico Medicine in Chemical Science in January 2023. The paper described the identification of a novel hit compound against CDK20, a kinase implicated in hepatocellular carcinoma, using AlphaFold's predicted structure of CDK20 as the input to their generative chemistry pipeline. The molecule was synthesized and showed activity in biochemical assays.

Insilico is not a patient-led group — they are a venture-backed AI drug discovery company that raised more than $400 million in private financing. The reason their CDK20 paper matters for this profile is methodological. It established that the workflow of "predict structure with AlphaFold, generate candidate molecules with a generative chemistry model, synthesize the top few, test in vitro" is a real workflow that produces real hits. The barrier to running that workflow at a small academic lab is the generative chemistry model, not AlphaFold. Open-source equivalents — REINVENT, DiffDock, and several others — have closed that gap meaningfully in the eighteen months since.

A small academic lab supported by a rare-disease foundation can now, in principle, replicate the Insilico workflow against a rare-disease target for tens of thousands of dollars in compute and synthesis costs, where the same pipeline would have required millions in pharma infrastructure five years ago. Whether they should is a separate question — the dropout rate of in-vitro hits going forward to drugs is brutal — but the price of taking a shot has fallen by roughly two orders of magnitude.

What AlphaFold cannot do yet — read this part twice

The cases above are real and verifiable. They are also bounded. AlphaFold solves one part of a long pipeline, and a structural prediction in the database is not a drug. The honest list of what the tool does not do, as of mid-2026:

• Binding affinity prediction is unreliable. AlphaFold 3 predicts the geometry of a protein-ligand complex reasonably well when the conformational change on binding is small. When the conformational change exceeds roughly five angstroms RMSD, the predictions degrade sharply, and the model does not output a binding affinity number that correlates well with experimental KD. The April 2025 bioRxiv assessment paper (a comprehensive evaluation of AlphaFold 3 in drug discovery) is explicit about this limitation.
• Dynamic conformations are largely invisible. Many drug targets — including GPCRs, kinases in different activation states, and intrinsically disordered proteins — exist as ensembles of conformations rather than single fixed structures. AlphaFold predicts a single static structure, and the alternative-conformation predictions from sub-sampling techniques are an active research area, not a solved problem.
• Induced fit is not modeled. When a ligand binds and reshapes the binding pocket as it docks, AlphaFold does not capture that reshaping. The structure you get is the apo (unbound) prediction, and the rearrangement that drugs actually undergo during binding has to be modeled separately with physics-based methods.
• The structure is a starting point, not a drug. A predicted structure tells you what to test. Synthesis, in-vitro validation, ADMET profiling, animal studies, IND-enabling toxicology, and clinical trials all remain. The median timeline from confirmed hit to FDA approval is still measured in years for repurposed drugs and in over a decade for novel chemical matter. AlphaFold compresses the front of that pipeline. It does not compress the back.

Why the patient-led model is structurally durable

The argument for taking the patient-led-research model seriously is not that any individual patient organization will discover a billion-dollar drug. Almost none will. The argument is that the cost structure has changed in a way that makes the experiment worth running at vastly more rare-disease targets than commercial drug discovery economics will ever justify.

Roughly seven thousand rare diseases are catalogued in databases like Orphanet and the NIH's Genetic and Rare Diseases Information Center. Fewer than ten percent of them have an FDA-approved treatment. Pharma economics route capital toward the diseases with patient populations large enough to support a $2 billion R&D bill at 20-percent gross margin, which leaves most rare diseases permanently outside the funding envelope. The patient-led model — a foundation funds a small academic group, the group runs AI-accelerated discovery against a defined target, the foundation pays for early in-vitro work, a clinical-stage partner picks it up if it works — fits diseases where the patient-pool economics will never close on their own.

AlphaFold did not invent this model. The Castleman Disease Collaborative Network, the Cystic Fibrosis Foundation, Cure SMA, and a dozen other rare-disease research consortia were running variations of it before AlphaFold existed. What AlphaFold and the free open-database ecosystem around it changed is the unit economics of the front of the pipeline. The first protein structure used to cost the price of a postdoc-year. It now costs a Google Drive download.

The honest expectation

Most rare-disease groups running AlphaFold-based projects will not produce a drug. Most will produce a slightly better mechanistic understanding of their disease, a stronger basis for grant applications, a more compelling case for a contract research organization, and a small contribution to the academic literature. That outcome is not a failure. It is the realistic median, and it is still a dramatically better baseline than the alternative of having no structural information at all.

The cases that succeed — adalimumab for iMCD, the Insilico CDK20 hit, the ASO programs for ultrarare splicing variants — will continue to emerge one paper at a time, in case reports and small clinical series rather than blockbuster announcements. The combined effect over a decade is meaningful even if no single result rewires the industry.

The wedge to remember is not the Nobel Prize or the 200-million-structure database. The wedge is that the cost to attempt this kind of research, at the bench scale a small patient-funded lab can afford, has collapsed. The volume of attempts is what produces the rare wins. The volume of attempts is what is now finally possible.