The exact text we saved

This is the version exactly as it was saved, character for character. The SHA-256 hash below is a 64-character code worked out from that text. Run the same text through SHA-256 and you get the same code; change a single letter and the code comes out completely different. Each version also stores the previous version’s hash, which chains them together so no version can be swapped out quietly.

TRUVACE RECORD VERSION
record: TRV-2026-2494
version: 1
kind: certified
reason: Certified into the record
timestamp: 2026-11-12T07:42:32.746854Z
status: published
lens: g_space
sector: science
headline: How Artificial Intelligence Shapes Science: Evidence from AlphaFold
dek: We study how a frontier AI model affects scientific discovery by examining the release of the AlphaFold2 algorithm and its impact on structural biology and related fields of science. Structural biology is the field of science concerned with understanding the structure and function of proteins. Researchers in this field historically devoted substantial time and resources to experimentally solving three-dimensional protein structures. AlphaFold can predict these structures without running experiments. In July 2021…
gain_title: A working paper comparing proteins with and without a pre-2021 experimental structure found that basic research on proteins lacking a structure rose 15 to 40% after AlphaFold2's July 2021 release.
problem_title: (none)
trace_subject: (none)
gain_reading: A working paper comparing proteins with and without a pre-2021 experimental structure found that basic research on proteins lacking a structure rose 15 to 40% after AlphaFold2's July 2021 release.
gain_evidence: basic research on proteins that had no structure information prior to AlphaFold increases by 15 to 40% relative to proteins that already had a structure
problem_reading: (none)
problem_evidence: (none)
quick_read: Economists Ryan Hill and Carolyn Stein studied what happened to protein science after Google DeepMind released AlphaFold2 predictions in July 2021. In a difference-in-differences comparison of proteins that had no experimental structure with those that already had one, basic research on the unsolved proteins rose 15 to 40% (16 to 25% in the paper's main estimates).

The paper treats AlphaFold as a test of how AI changes research. Solving one protein structure by experiment can take months to years and cost $100,000 or more. Free predictions for hundreds of thousands of proteins removed that bottleneck for the roughly 92% of curated SwissProt proteins that had no experimental structure, and attention moved toward them.

The effect has not yet reached drug discovery. The authors saw no change in early-stage drug research on the newly illuminated proteins in the three years after release, though they call the results noisy. Experimental structure determination has also barely slowed, so predictions are so far a complement to lab work, not a replacement.
limitation: The drug-development finding rests on public ChEMBL data over three years and is described as noisy; private pharmaceutical bioactivity work is not captured, so activity there could differ.
tag: Evidence-backed gain
key_points: Only 8% of the curated SwissProt proteins had an experimental structure before AlphaFold, so most proteins received a large shock of new structural information. | The authors' main estimates put the rise in research on previously unsolved proteins at 16 to 25% relative to solved ones, with larger effects late in the sample period. | New papers on previously unsolved proteins were much more likely to cite core AlphaFold papers, which the authors read as evidence that the AI model drove the shift. | The panel covers 546,646 proteins from 2017 to 2024; only 14% were linked to any non-structure paper in that period. | Experimental structure deposits to the Protein Data Bank showed no decline after July 2021; over 60% of deposits near the end of the sample had data collected after AlphaFold's release. | Use of predicted structures as starting templates in molecular replacement rose sharply after AlphaFold, concentrated among structures without a similar experimental structure. | No similar rise appeared in ChEMBL drug-binding experiments on unsolved proteins in the three years after release; only 1% of SwissProt proteins link to any ChEMBL activity.
rundown: The authors built a protein-level panel from SwissProt, a curated database of 546,646 proteins with AlphaFold predictions, linking each to curated papers, citations in OpenAlex and drug-binding data in ChEMBL. They dropped papers about structures themselves to measure research beyond structure solving. AlphaFold released 365,000 predicted structures in July 2021, 560,000 more in December 2021 and over 200 million in July 2022.

On experimental work, the paper finds little substitution. Using 61,638 Protein Data Bank deposits from 2017 to early 2024, researchers kept depositing structures at an indistinguishable or slightly higher rate, and the structures kept appearing in journals including Cell, Nature and Science. The predicted confidence of structures being solved dropped about one point, against a standard deviation above ten, so scientists did not pivot to proteins AlphaFold handles poorly.

The authors say predictions are not perfect stand-ins. They cite a comparison of about 100 structures in which roughly 10% of very confidently placed amino acids were meaningfully misplaced, and note that AlphaFold typically predicts a protein's default state. They also caution that continued lab work may reflect researcher incentives, not efficiency, and that AI could replace experiments as models improve.

In an example from the paper, Andrea Pauli's lab used an AlphaFold structure of a previously unsolved sperm protein, Tmem81, to hypothesize that three sperm proteins form a complex that recognizes the egg protein Bouncer. They validated the idea with experiments and published it in Cell in 2024.

On drugs, the authors say it is too early to expect new medicines. They argue that more basic science may be needed before the new structures become drug targets, and that activity could still emerge later.
sources:
- peer_reviewed | NBER Working Paper | https://doi.org/10.3386/w35143 | 2026-05-01
prev: 0000000000000000000000000000000000000000000000000000000000000000

SHA-256 hash

7d0937b49b91f6ea3fa80540a711bff5cfce78668563d58c06719a79fa04c98f

Previous version’s hash

0000000000000000000000000000000000000000000000000000000000000000