AdaptiveFlow opens a 69-billion-compound docking library to labs without supercomputer budgets

The engineering worth studying here is subtractive. Docking all 69 billion compounds against a single target is a bill the paper itself names as a major bottleneck, one that worsens as vendor libraries push past a trillion molecules ^{[7] [14]}. AdaptiveFlow's answer is to make chemical space addressable: 18 molecular properties serve as the axes of a grid, cells of that grid are ranked, and only the promising subspaces get docked, with active learning available to sharpen the choice ^[3]. The reported saving is orders of magnitude of compute ^[3].

That design carries a cost the abstract does not price. A property grid assumes potency clusters in property-defined neighbourhoods. For enzyme active sites that is a reasonable bet. The paper's own motivation is the harder case, targets such as protein-protein interfaces and allosteric sites, where pockets are often shallow, dynamic or poorly defined ^[13]. The thing the abstract does not tell you is recall: of the actives an exhaustive dock of the full library would surface, how many survive prioritization.

Some arithmetic for scale sense. A typical high-throughput screen covers hundreds of thousands of compounds ^[11]; taking 500,000 as the figure, the dock-ready library is roughly 138,000 times larger ^[16]. Measured the other way, against on-demand chemistry that the authors put at up to 3 trillion molecules across vendors, 69 billion is about 2.3 percent ^{[17] [14]}. And against the estimate that synthetically accessible drug-like molecules exceed 10^60 ^[10], every screening claim, this one included, samples a fraction on the order of 10^-49 ^[18]. That is the honest frame for library size as a selling point.

Which is why the co-crystals do more work than the 5.6-million-CPU run. Near-linear scaling in the AWS cloud is a real engineering result ^[5], but it is the compute reduction, not the millions of cores, that makes the capability portable. The structures of FSP1 complexes, which the authors say gave mechanistic insight into inhibition ^[9], are the part that indicates the docking rank corresponded to physical binding. Two targets, FSP1 and PARP1 ^[6], is a demonstration rather than a hit rate, and "nanomolar" spans three orders of magnitude; the abstract reports no potency values.

The quieter contribution is preparation. The authors ship the Enamine REAL Space in screening-ready form, described as the largest ready-to-dock drug-like library they know of, and additionally in SELFIES ^[2], which is what turns a vendor catalogue into something a graduate student can queue up. More than 1,500 docking protocols come wired in, including GPU-accelerated and machine-learning methods ^[4], against a field where many platforms remain CPU-only ^[8]. Whether that translates into a production win is a different measurement: cost per experimentally confirmed hit, including synthesis of on-demand compounds and assay time. This paper leaves that number unreported, and a compute-reduction figure is no stand-in for it.