Take the interaction count apart first. The pulldowns covered the proteins encoded by 100 autism-linked genes and returned more than 1,800 interactions [5][6], which averages roughly 18 partners per bait [1]. Of those, 87% had not been reported before [6], so about 1,566 edges are new to the literature and only around 234 had prior support behind them [2]. A novelty share that high is what you get when mass spectrometry goes where nobody had pulled down before, and it is also the condition a map is in before anyone has tried to break it with a second method. The account of the paper attaches no false-discovery estimate or independent validation rate to the 1,800 [3].
The summary and the methods work at different scales. Hundreds of genes and dozens of mutations are described as converging on a small number of shared networks [3], while the experiment started from 100 genes [5]. The larger number counts partners the atlas found rather than genes it tested [6], and that distinction is where the convergence lives: risk genes that share no sequence can still share a complex.
FOXP is the strongest part of the argument that this convergence is mechanical rather than statistical. Mutations at entirely different positions inside FOXP1 all disrupt the same FOXP1-FOXP4 contact [10], and FOXP2 mutations land on that interaction too [9]. AlphaFold placed the lesions at the interface, and Xenopus embryos plus human forebrain organoids supplied the downstream readout [7]: cortical neurons develop prematurely and circuits become more excitable [9]. Pairing a whole-embryo model with human-derived tissue is a defensible design, because it separates "this interface matters for development" from "this interface matters in human cells".
The thing it does not tell you is magnitude. No effect sizes accompany the organoid phenotypes in the write-up [4], and a dish has no behaviour to measure. GEN puts the relevance of the findings at approximately 30% of individuals with profound autism, many carrying rare high-impact mutations in established risk genes [16]; the phrasing leaves it unclear whether 30% is a share of profound autism or the share of autism that is profound, and that difference sets how many people any single-complex drug could reach.
The hub's remit is small molecules that modulate disrupted complexes [13], and Nevan Krogan of UCSF frames the atlas as resolving targets "down to the interfaces we can target with a drug" [14]. No compound, series or assay appears in the announcement [5]. My read is that the map earns its funding as infrastructure whether or not FOXP1-FOXP4 proves druggable, since interface-resolved networks are reusable and a failure on one complex leaves the other edges standing. The narrower milestone that tests ARIA's actual bet is a molecule that restores the FOXP1-FOXP4 contact and shifts the organoid phenotype it was picked for.
