AI
“Scientists spent 50 years solving 194,000 protein structures by hand, while AlphaFold predicted 214 million protein structures in 3 years, covering nearly every known protein on Earth.”
Plain restatementOver roughly 50 years of experimental work, scientists deposited about 194,000 protein structures in the Protein Data Bank. Within about three years of AlphaFold2's breakthrough, the AlphaFold Protein Structure Database contained roughly 214 million predicted protein structures, covering nearly all protein sequences catalogued in UniProt.
The numbers in this post are real and come from the people who run the databases involved. The AlphaFold Protein Structure Database, run jointly by EMBL's European Bioinformatics Institute and Google DeepMind, does hold over 214 million predicted protein structures, grown from about 360,000 at its July 2021 launch, and DeepMind described its July 2022 release as covering nearly all catalogued proteins known to science. The Protein Data Bank, founded in 1971, did hold roughly 194,000 experimentally determined structures around 2022. What the post glosses over is that these are two different kinds of thing: the 214 million are computer predictions with varying confidence, which researchers treat as strong hypotheses rather than finished answers, while the 194,000 are structures determined by laboratory experiment. The phrase "by hand" is also inaccurate, since experimental structures come from X-ray crystallography, NMR and cryo-electron microscopy, all computation-heavy techniques. The 194,000 figure is also several years old, and the experimental archive now lists over 260,000 structures. The scale of AlphaFold's contribution is genuine and large, but the side-by-side comparison makes it look like a direct like-for-like replacement, which it is not.
[drifted from the evidence:] Scientists spent 50 years [drifted from the evidence:] solving 194,000 protein structures [drifted from the evidence:] by hand, while AlphaFold [drifted from the evidence:] predicted 214 million protein structures [drifted from the evidence:] in 3 years, covering nearly [drifted from the evidence:] every known protein [drifted from the evidence:] on Earth.
[added by the neutral restatement:] Over roughly 50 years [added by the neutral restatement:] of experimental work, scientists deposited about 194,000 protein structures [added by the neutral restatement:] in the Protein Data Bank. Within about three years of AlphaFold2's breakthrough, the AlphaFold [added by the neutral restatement:] Protein Structure Database contained roughly 214 million [added by the neutral restatement:] predicted protein structures, covering nearly [added by the neutral restatement:] all protein [added by the neutral restatement:] sequences catalogued in UniProt.
Red-tinted words in the claim drifted from the evidence. Green-tinted words are what a neutral restatement needs.
The trace / claim to source
- The figure of roughly 214 million predicted structures is real and comes from the database operators themselves, not from a promotional estimate. The database holds over 214 million predicted protein structures, up from roughly 300,000 in 2021.
- The time frame of about three years is broadly right for how the database reached that size. Archiving began with the July 2021 release of over 360,000 structures and the 214 million figure is recorded as of September 2023, a span of about two years from launch and about two years and ten months from the November 2020 CASP14 result.
- The coverage statement reflects what the developers said at the time. The July 2022 release was described by DeepMind as predicted structures for nearly all catalogued proteins known to science.
- The experimental side of the comparison is close to the contemporaneous record. A 2022 account put the Protein Data Bank at more than 194,000 structures built up over decades, and the archive was founded in 1971 and held more than 190,000 structures in 2022, so "about 50 years" fits the period from 1971 to 2021 or 2022.
- The claim's own word choice is careful in one important respect: it says AlphaFold "predicted" rather than "solved," which is the correct distinction.
- The sentence places 194,000 and 214 million side by side as if they were the same kind of object. They are not. The first are experimentally determined structures; the second are computational models whose reliability varies by protein and by region, described in the refereed literature as valuable hypotheses that accelerate but do not replace experimental structure determination. A reader is invited to conclude that AI produced a thousand times more finished knowledge, which the evidence does not support.
- The confidence distribution is left out. For the human proteome, 58% of residues were covered by a confident prediction and 36% of all residues by a very high confidence one, and an unrefereed whole-proteome assessment reports that 30% to 40% of amino acids in model organisms fall in the low-confidence range. Those figures are scoped to the proteomes studied, not to all 214 million entries, but they show the tally is not uniform in quality.
- "nearly every known protein on Earth" means nearly every protein sequence catalogued in UniProt at a particular moment. The 200 million structures were predicted from UniProt sequences as they stood in April 2021, which is a snapshot of a catalogue, not the full diversity of proteins that exist. The operators have since updated the database to the UniProt 2025_03 release and added isoform coverage.
- The 194,000 figure describes the experimental archive around 2022 and is presented in an October 2026 post as the total of human experimental effort. The RCSB PDB homepage snapshot now lists 260,626 experimental structures, so experimental determination did not stop when AlphaFold arrived.
- The phrase "by hand" describes the method inaccurately. Experimental structures are determined with X-ray crystallography, nuclear magnetic resonance and cryo-electron microscopy, each involving heavy computation, rather than by manual work. Traditional methods are time consuming and costly experimental techniques such as X-ray crystallography and cryo-electron microscopy.
- Counting units differ on both sides. PDB entries include nucleic acids and complexes as well as proteins, and many entries are repeat structures of the same protein, while AlphaFold database entries are one model per sequence, including large numbers of near-identical sequences across species. Neither number is a count of distinct proteins solved or predicted.
- The surrounding slide framing, "8 Things AI Did That No Human Ever Could" and "CRACKED PROTEIN FOLDING," sits awkwardly with the fact that AlphaFold2 was trained on the experimental structures humans produced, which is the archive the first half of the claim is counting.
- The precise provenance and date of the exact figure 194,000 could not be pinned to a dated Protein Data Bank holdings snapshot. Two 2022 publications support a number in that range, one giving more than 194,000 and one giving more than 190,000, but whether the post's number refers to total PDB entries, protein-only entries, or a specific month is not determinable from the post.
- What fraction of all 214 million predictions are high confidence is not established here. The confidence figures retrieved apply to the human proteome and to sets of model organisms, not to the full database.
- Whether the post's "3 years" is measured from CASP14, from the AlphaFold2 paper, or from the database launch is not stated, and the three starting points give spans ranging from about two years to about three years.
The AlphaFold database figure is real and is documented by the people who run the database. The AlphaFold Protein Structure Database holds over 214 million predicted protein structures, expanded from the roughly 300,000 structures released in 2021, and these predictions have been integrated into resources including the PDB, UniProt, Ensembl and InterPro. Data archiving for AlphaFold DB began with an initial release in July 2021 holding over 360,000 structures for 20 model organism proteomes, and as of September 2023 it archived over 214 million predicted protein structures. The coverage statement traces to the developer's own July 2022 announcement. DeepMind, with EMBL's European Bioinformatics Institute, released predicted structures for "nearly all catalogued proteins known to science," one year after releasing and open sourcing AlphaFold and creating the database. All 200 million plus structures were made available for bulk download, and most pages on UniProt were to carry a predicted structure. Contemporary reporting described the release in similar terms, covering some 200 million proteins from 1 million species, described as covering nearly every known protein on the planet. The 194,000 figure matches the experimental archive as it stood around 2022. One named outlet wrote at the time that decades of slow experiments had revealed the structure of more than 194,000 proteins, all housed in the Protein Data Bank, and a 2022 review of the archive recorded that the PDB, established in 1971 as the first open-access digital data resource in biology, housed more than 190,000 three-dimensional structures of proteins, DNA and RNA and their complexes, and was then in its 51st year of operation. That archive has kept growing: the RCSB PDB homepage snapshot lists 260,626 experimental structures, 405 integrative structures and 1,062,058 computed structure models, with latest entries dated 7 October 2026. On what the predictions are: they are models with per-residue confidence, not experimental determinations. For the human proteome specifically, the dataset covers 58% of residues with a confident prediction, of which 36% of all residues have very high confidence, against a baseline where only 35% of human proteins mapped to a PDB entry, in many cases covering only a fragment of the sequence. An unrefereed preprint assessing whole proteomes reports that for model organisms, 30% to 40% of all amino acids fall into the low-confidence pLDDT range. A refereed commentary frames the predictions as valuable hypotheses that accelerate but do not replace experimental structure determination. The database is also a snapshot, not a live mirror of known proteins. The AlphaFold database provides 200 million structures predicted by AlphaFold2 and released in 2022 from UniProt sequences dated April 2021, and of the 20,504 full-length human structures in it, 631 entries conflict with the June 2025 UniProt release, with similar discrepancies for other species. The operators have since updated it: the 2025 update aligns the database with the UniProt 2025_03 release and extends structural coverage to include isoforms plus underlying multiple sequence alignments.
Complete reasoning
The reply is formatted for pasting into the thread where the claim is circulating.
Ask this case
Answers come only from the case file above; nothing is added.
Is the claim that AlphaFold predicted 214 million protein structures accurate?
Yes. The AlphaFold Protein Structure Database, run by EMBL's European Bioinformatics Institute and Google DeepMind, held over 214 million predicted structures as of September 2023, up from about 360,000 at its July 2021 launch.
Did scientists really spend 50 years solving 194,000 protein structures by hand?
The 194,000 figure and roughly 50-year timespan match the Protein Data Bank, founded in 1971, as it stood around 2022, but the structures were not solved by hand. They came from computation-heavy lab techniques like X-ray crystallography, NMR, and cryo-electron microscopy.
Are the 194,000 experimental structures and 214 million AlphaFold structures really comparable?
No. The 194,000 are experimentally determined structures, while the 214 million are computer predictions of varying confidence that researchers treat as hypotheses rather than finished answers, not a like-for-like replacement.
Does AlphaFold's database really cover nearly every known protein on Earth?
DeepMind described its July 2022 release that way, but this means nearly all protein sequences catalogued in UniProt at that time, specifically sequences dated April 2021, which is a snapshot rather than the full diversity of proteins that exist.
How reliable are the AlphaFold predictions themselves?
Reliability varies: for the human proteome, 58% of residues had a confident prediction and 36% had very high confidence, and an unrefereed study found 30% to 40% of amino acids in model organisms fall into a low-confidence range.