Divide the index by its contents and you get about eight predicted spectra for every molecule: more than 800 million spectra covering more than 100 million compounds [14]. The write-up does not say what the other seven capture, and it does not report how often a retrieved match is the right one [19]. What it does establish is that the spectra were generated before anyone searched them and then filed into a browsable resource [3], so the expensive model inference was paid once, by the authors, instead of per query by every lab.
Run the thousandfold claim backwards and the incumbent looks thin. A thousandfold expansion of a space containing more than 100 million molecules [2][5] puts existing experimental coverage on the order of 100,000 compounds, which is about 0.1% of that space [15]. That is consistent with the paper's own accounting, which puts collective library coverage at under 1% of known compounds [4]. Spectra without a close library relative usually just stay unannotated [18], which is the mechanical reason the 80% figure is as large as it is [1].
The design choice that matters is what comes out alongside the predicted spectrum. DeepMS2Reasoner enumerates physically plausible fragmentation steps using symbolic chemical rules, uses a neural network to assign each step a likelihood, and returns an annotated account of how the molecule came apart [7]. A library hit gives a chemist a similarity number to trust or distrust. A fragmentation route gives them something to disagree with, which is a different and more useful kind of output.
The demonstration's denominator deserves attention. Several thousand chemical features separated germ-free mice from mice with normal gut microbiota, and most of the abundant ones defeated standard identification [10]. The team queried the index with 111 of them, selected as the most abundant unidentified microbiota-dependent compounds [11]. Roughly a third returned close predicted matches with related structural candidates [11], which is about 37 compounds; the other 74 or so came back only as molecular neighborhoods of structurally related compounds [12][16]. Two compounds went further, with spectra pointing to polyamine derivatives, according to Frank Schroeder of the Boyce Thompson Institute and Cornell [13][8].
So the tested cases were a selection of the loudest signals in the data set, the most abundant unidentified compounds rather than a random draw from the unannotated pile, and the account does not say how retrieval behaves on fainter features [11][19]. Nor does the ranking change the cost of being sure: by the authors' own description, pinning down a single unknown structure runs days to months of iterative analysis and experimental validation [6]. A neighborhood of candidates shortens the reasoning, not the bench work.
Sizing how much of that 80% AIMe actually recovers requires retrieval precision measured against spectra whose structures are already known. The announcement reports coverage [19].
