Biology got three new search engines this week: enzymes, malaria clocks, and tumor nucleosomes

Three new ways to search biology

Three preprints built the same kind of thing for three different corners of biology: a search engine. One retrieves enzymes from a metagenomic haystack of 8.2 billion associations. One indexes the daily rhythms of the malaria parasite. One maps nucleosome landscapes across cancers and reads tumor signals in blood. All three are un-peer-reviewed preprint findings, and all three ship as public resources. m:…6352 m:…1797 m:…3899

The enzyme retriever

EzHit, the enzyme paper's system, operates at a scale that's hard to picture: 8.20 billion retained enzyme-reaction associations across 30.27 million predicted enzyme sequences from the OMG_Prot50 dataset, the authors report. Its job is retrieval — given a reaction, find the enzyme. In the authors' reported comparisons, EzHit outperforms representative enzyme reaction matching baselines on enzyme annotation, orphan-reaction retrieval, and family-specific fine-tuning. And it survived contact with reality: experimental validation confirmed the activity of EzHit-prioritized metagenomic enzyme candidates for a sparsely annotated reaction. m:…6352

The use case in the paper's title is metagenomic biocatalyst discovery: mining environmental DNA for enzymes that do chemistry nobody has characterized yet. The orphan-reaction benchmark is the telling one — reactions with almost no annotated enzymes are exactly where retrieval beats memorization. m:…6352

The confidence machinery is the detail worth noting. The authors report that Mahalanobis distance-based geometric confidence provides more useful reliability estimates than probability-derived uncertainty metrics or evidential learning. A retriever's value lives or dies on knowing when it's unsure — here the answer is geometry, not probabilities. m:…6352

The malaria clock

PlasmoRhythm is a database built around the malaria parasite's biological rhythms — the cycles that govern when the parasite is active and, by extension, when it's vulnerable. It integrates 67 transcriptomic, proteomic, and metabolomic datasets from 15 studies, the authors report. It supports cross-study interrogation with four standard rhythmicity tools — JTK_CYCLE, ARSER, Lomb-Scargle, and meta2d — and, the authors say, enables the identification of rhythmic virulence factors and antimalarial drug interactors. Time as a drug-discovery input: find the genes and proteins that keep a schedule, then time the intervention. m:…1797

The cancer atlas

The third paper maps nucleosome landscapes across cancers and puts them to work in liquid biopsy. The authors report that nucleosome occupancy provides regulatory information complementary to chromatin accessibility for tumor stratification. The headline numbers are in cell-free DNA: tissue-derived nucleosome signatures in cfDNA enable classification among three cancer types at an AUROC of 0.877, and colorectal cancer detection using those signatures reaches an AUROC of 0.930. The resource behind it, OncoNucAtlas, is a publicly accessible database for exploring pan-cancer nucleosome landscapes. m:…3899

Why it matters

None of these is a company. They're preprints and public databases. But the pattern is the one that built the first generation of bio-AI infrastructure: structured, searchable biological data becoming a platform others build on. EzHit turns billions of enzyme associations into retrievable candidates. PlasmoRhythm turns scattered malaria omics into a queryable clock. OncoNucAtlas turns nucleosome landscapes into a public atlas. Retrieval layers on top of raw biological data have a proven playbook — and these three look like the next layer of it. That framing is mine; the numbers above are the authors'. m:…6352 m:…1797 m:…3899