What we’re reading, March 20, 2026
Science policy debates, housing innovation myths, and the Jones Act suspension
Happy spring! Here’s what caught our attention this week:
Ryan Hill and Carolyn Stein have a new working paper on how AlphaFold has affected protein structure research in the five years since its release. They ask three primary questions:
Did experimental (i.e., non-computational) structure determination decline?
Did basic research shift toward proteins that previously had no structural information?
Did early-stage drug development start targeting those newly solved proteins?
Pre-register your guesses for each, before reading item #3 for a discussion of what Hill and Stein found. — Jordan Dworkin
While you’re thinking about that, let’s stay with science. Does MAGA Actually Want American Science to Win? is the latest from Ari Shulman, writing in The New Atlantis. Shulman argues that MAGA’s critique of the failings of status quo American science are broadly on point (and part of a tradition that predates President Trump), but that the policy conclusions it draws from that critique will not make America stronger. Dramatic cuts to research budgets are not enough (I would add that restrictions to skilled immigration don’t help either). The positive vision for science that Shulman grapples most with is a revitalization of scientific ideals championed by NIH Director Jay Bhattacharya, where there are no taboos, no appeals to authority, and free debate carries us to truth. My take? The deep challenge of organizing science is that only specialized knowledge can assess the quality of scientific work.1 That means peer approval is a fundamental part of the scientific process, and from there it’s not far from appeals to authority and deferring to experts. It’s tough! But that’s not to say the status quo was optimal and things can’t be better. Indeed, I’m a fan of some of the metascience innovation we’ve seen under this administration, from the NSF tech labs proposal to the streamlining of research bureaucracy. — Matt Clancy
Got your guesses about Alphafold ready? Here’s what Hill and Stein found.
Did experimental (i.e., non-computational) structure determination decline? The rate of experimental structure determination is unchanged, and the relevant papers are still being published in top journals; AlphaFold predictions are complementing rather than replacing experimental research (though Hill & Stein note that this lack of substitution isn’t necessarily efficient, or permanent).
Did basic research shift toward proteins that previously had no structural information? Basic research on previously unsolved proteins is up 15-35% relative to proteins with known structures, with initial shifts taking place in protein function research and more recent (and larger) shifts among expression, disease, and interaction studies.
Did early-stage drug development start targeting those newly solved proteins?There is no corresponding increase in early-stage drug R&D on newly solved proteins vs. previously solved ones (though the authors note that private pharma work would not be captured here).
Overall, it seems that structure was a meaningful bottleneck for basic science, but not the binding one for drug discovery. — Jordan Dworkin
Back in 2023, we asked Michael Wiebe to take a look at a 2021 article by Enrico Moretti titled “The Effect of High-Tech Clusters on the Productivity of Top Inventors.” We were interested because one of the inputs to our grantmaking decisions is a quantitative estimate of the social impact of a grant, and this paper was important for our estimate of how housing affects innovation. Unfortunately Wiebe found a lot of problems with the paper, prompting us to substantially downgrade the weight we put on the innovation effects of new housing. Other (non-innovation) benefits of housing now form a much larger part our estimate of its impact (though we’ll revise estimates again as new evidence emerges). Wiebe has now written up his conclusions on his substack, and a comment on the paper will appear in the journal in which it was originally published. — Matt Clancy
Brian Potter examines the elusive cost savings of the prefabricated home, and as someone now living through the third modular construction hype cycle of my career, the piece is a useful corrective. There’s a reason we keep coming back to off-site construction: HUD-code manufactured housing does deliver real savings in the US; the Sears “kit home” panelized model was a legacy success in the US; and contemporary international examples (especially Germany and Sweden) show factory-built housing can work at scale. But again and again, modular construction in the US has failed to sustainably deliver on its cost promises, with high-profile bankruptcies in the UK telling a similar story. The benefits tend to be schedule, predictability, and quality control rather than price. As we politely take part in the latest wave of enthusiasm, it’s worth being clear-eyed that this has been tried before, and is not a substitute for growth control reform. — Alex Armlovich
Separately, Jerusalem Demsas writes about how urban divestment is (still, as Pete Saunders reminds us) a much larger problem in the US than urban “gentrification”, but compositional & mimetic drivers of un-representative stylized facts crowd out the underlying statistical & material reality of neighborhood conditions in most American cities. — Alex Armlovich
It’s official: the Trump administration is suspending the Jones Act, which bans non-US flagged vessels from traveling between US ports, for sixty days in an attempt to keep gas prices low amid the Strait of Hormuz crisis. I doubt the suspension will do much toward that specific purpose, but it’s as good a time as any to think about the extreme costs that law imposes on the US economy. Colin Grabow, the premier Jones Act foe at the Cato Institute, highlights some oil/gas specific costs, like the fact that California gets fuel from the Bahamas as opposed to Gulf Coast US states because of a lack of tankers that can ship oil directly from Texas and Louisiana. — Dylan Matthews
One of the big challenges in permitting reform is that it’s really hard to forecast how changing one piece of the massive US legal/regulatory permitting system will play out. At a more micro-level, it’s also a headache for developers to try to figure out what environmental requirements their project actually triggers. Specifically, when contemplating a new project, developers do not know if the Clean Water Act applies until the Army Corps of Engineers completes an assessment. In a new working paper, Greenhill, Walker, and Shapiro train a deep learning model to evaluate whether Clean Water Act provisions apply to a given parcel, using Army Corps of Engineers determinations tunder several recent rulemakings. They find their model is 65 times better at identifying regulated sites than the leading geophysical approach. This helps in analyzing past regulatory shifts (the authors find Sackett deregulated roughly a third of all previously regulated waters), helps in generating high-quality projections of proposed regulations before they’re implemented, and could potentially help developers better predict if they’re actually subject to CWA restrictions. — Willow Latham-Proenca
Invention is different in this sense; specialized knowledge is often not required to tell if a technology delivers what it promises.