Sign in

Mengyi Sun

@sunmengyi.bsky.social
236 followers 1.2K following 20 posts
PostsRepliesMedia
Mengyi Sun @sunmengyi.bsky.social · 08/04/2026
On a personal note, as a computational biologist, AlphaFold has also been one of the most inspiring developments to track. Just trying to understand how it works has been a major learning experience for me, and part of what made this project so exciting to work on. 15/15
020
Mengyi Sun @sunmengyi.bsky.social · 08/04/2026
And www.innovationgrowthlab.org/wp-content/u... from @iglglobal.bsky.social
innovationgrowthlab.org
120
Mengyi Sun @sunmengyi.bsky.social · 08/04/2026
Also glad to see complementary work emerging in this space, including the great work papers.ssrn.com/sol3/papers.... by Carolyn Stein and @ryanreedhill.bsky.social 13/15
papers.ssrn.com
How Artificial Intelligence Shapes Science: Evidence from AlphaFold
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 rela
100
Mengyi Sun @sunmengyi.bsky.social · 08/04/2026
If gains are concentrated in data-rich regions, AI may reinforce the canon. If gains are greatest in sparse, uncertain regions, AI may expand the frontier. AlphaFold2 appears closer to the second case. 12/15
100
Mengyi Sun @sunmengyi.bsky.social · 08/04/2026
We propose a simple mechanism: The impact of AI on science depends not only on average performance, but on where its informational gain is greatest. 11/15
100
Mengyi Sun @sunmengyi.bsky.social · 08/04/2026
We also find increased attention to historically understudied human genes. Relative to the 2020 baseline, research on these genes rises by 29.0%. 10/15
100
Mengyi Sun @sunmengyi.bsky.social · 08/04/2026
The effect propagates downstream too. Among papers consuming structural knowledge, engagement with genes lacking any experimental structure rises sharply after 2021. By 2025, the increase is 5.58 percentage points above the pre-trend. 9/15
100
Mengyi Sun @sunmengyi.bsky.social · 08/04/2026
We also see a broader reallocation of attention. AF2-citing PDB papers engage 38.3% more novel protein sequences than contemporaneous non-citing papers. And concentration of attention across protein clusters stops increasing after AF2. 8/15
100
Mengyi Sun @sunmengyi.bsky.social · 08/04/2026
This isn’t just a generic time trend. The novelty shift is concentrated: --in papers that cite AlphaFold2 --in targets with high-confidence AF2 predictions --not in the low-confidence regime 7/15
100
Mengyi Sun @sunmengyi.bsky.social · 08/04/2026
Main result: A long-running decline in the novelty of experimental structural targets stops after AF2. Before AF2, novelty declines by 1.2 percentage points per year. After AF2, the slope becomes statistically indistinguishable from zero. 6/15
100
Mengyi Sun @sunmengyi.bsky.social · 08/04/2026
Then the key question: Did AI shift what scientists chose to study? We measure novelty by asking whether a structure targets a protein sequence cluster that had not been studied before. 5/15
100
Mengyi Sun @sunmengyi.bsky.social · 08/04/2026
First, AF2 diffused very quickly. By 2024, 20.5% of new PDB structures are associated with papers citing AlphaFold2. Among cryo-EM entries, the adoption rate is even higher: 38.7%. 4/15
110
Mengyi Sun @sunmengyi.bsky.social · 08/04/2026
We analyze: --245,396 experimental structures in the Protein Data Bank --87,656 associated papers --248,191 downstream biomedical papers that consume structural knowledge 3/15
100
Mengyi Sun @sunmengyi.bsky.social · 08/04/2026
We study this using the 2021 release of AlphaFold2 as a quasi-experimental opportunity. Why AlphaFold2? Because it was one of the first AI-for-science tools to diffuse rapidly and visibly into real experimental workflows. 2/15
100
Mengyi Sun @sunmengyi.bsky.social · 08/04/2026
Many people worry that AI will make science more productive but also more conservative — pushing attention toward already well-studied areas. @yianyin.bsky.social Sukwoong, and I ask: can AI instead help expand the scientific frontier? www.biorxiv.org/content/10.6... 1/15
174
Mengyi Sun @sunmengyi.bsky.social · 19/04/2025
If diffusion can't—and shouldn't—be constrained, perhaps it's time we simply embrace it. 3/3
000
Mengyi Sun @sunmengyi.bsky.social · 19/04/2025
Yet, more profoundly, as Thomas Jefferson observed: "He who receives an idea from me, receives instruction himself without lessening mine; as he who lights his taper at mine, receives light without darkening me." 2/3
100
Mengyi Sun @sunmengyi.bsky.social · 19/04/2025
Really glad that this is out! Fantastic collaboration with @honglin-bao.bsky.social and @innovation.bsky.social! Key insight: Knowledge diffusion is nearly impossible to constrain—where there's a will, there's a way. From the atomic bomb to ChatGPT, determined minds always find a path. 1/3
direct.mit.edu
Where there’s a will there’s a way: ChatGPT is used more for science in countries where it is prohibited
Abstract. Regulating AI is a key societal challenge, but effective methods remain unclear. This study evaluates geographic restrictions on AI services, focusing on ChatGPT, which OpenAI blocks in seve...
194
Mengyi Sun @sunmengyi.bsky.social · 19/11/2024
🙋‍♂️
110
Mengyi Sun @sunmengyi.bsky.social · 19/11/2024
do the list still have some capacity?if yes,please add me
100