2026-08-27
Weissman and Mann labs screened 52,979 viral smORFs and found >2,000 active microproteins; AlphaFold 3 showed EBV BNLF2b mimics host PKA inhibitors to drive epithelial growth.
Viral genomes pack overlapping frames, non-ATG starts, and programmed frameshifts into a few kilobases, so standard gene finders miss a lot of coding sequence. Ribosome profiling has already turned up thousands of non-canonical ORFs in herpesviruses, influenza, SARS-CoV-2, and HIV. The open question is how many of those small ORFs (smORFs) actually make functional microproteins, usually under 120 amino acids.
HIV annotates 7 of 15 genes as microproteins; SARS-CoV-2 annotates 10 of 29. Short proteins such as Vpu and Vpr previously pointed the field to host nodes like tetherin and DCAF1. Across the human-infecting virome those short proteins had not been mapped at functional scale. Groups led by Weissman, Mann, and Elledge built a phenotype-first funnel to do that.
They pulled every start-stop smORF from 502 human-infecting viral genomes in VirusHostDB (38 families, all seven Baltimore classes). Oligo synthesis at the time capped at 400 nt, so they kept sequences under 121 amino acids. The library has 52,979 elements: 1,606 annotated, 47,525 unannotated (at least 24 aa, ATG or CTG start, 89.7% of the library), and 3,848 negative controls. Of the annotated set, 74% have UniProt annotation scores of 3 or less.
Toxic ORFs die during lentiviral packaging. The pANDORA vector puts a Tet-On cassette in antisense orientation with almost no basal leak. Titers drop about 50-fold; SARS-CoV-2 Nsp1 and similar toxins survive until induction.
The funnel has three layers. A 10-day growth screen in K562 and RPE-1, calling hits at least two standard deviations from matched negative controls, returned 4,329 primary hits (8.2%). An enriched library of 4,107 elements (toxic in both lines, pro-growth in either, plus every annotated gene) was rescreened and yielded 2,028 hits at FDR < 0.05. Downstream assays added MHC-I surface levels, survival under ER stress, Perturb-seq on the enriched library, affinity-enrichment mass spectrometry (about 2,304 runs), and AlphaFold 3 models kept only at interface confidence ipTM > 0.75.
Perturb-seq used a 10x Flex probe workflow on RPE-1 and scored transcriptional phenotypes for 3,412 smORFs. Function was assigned by guilt-by-association: cluster on transcriptional Pearson distance plus a minimum-distortion embedding, then layer sequence homology and prior literature. All 18 transcription-factor positive controls passed the energy-distance test with more than 10 DEGs; every negative control stayed under 5 DEGs.
Toxic hits correlated across the two cell lines (Pearson R = 0.78). Pro-growth hits appeared almost only in non-transformed RPE-1. In the secondary screen, 70% of hits remained unannotated, 55% overlap annotated genes, and 34% start at a non-ATG codon. Among annotated smORFs, 58% (939/1,606) had no growth phenotype, which is why MHC-I and ER-stress screens were added.
| Assay | Scale | Hits |
| Primary growth | 52,979 | 4,329 (8.2%) |
| Enriched growth | 4,107 | 2,028 (FDR<0.05) |
| MHC-I surface | 8,806 | 748 |
| ER-stress survival | enriched library | 602 |
| Perturb-seq energy test | 3,412 | 1,185 (691 with >10 DEGs) |
Forty percent of MHC-I hits missed the growth screens. Known TAP inhibitors HSV ICP47, EBV BNLF2a, and cowpox CPXV012 came back. Five MHC-I-only candidates profiled by AE-MS hit different steps: synthesis (AKMV OPG024 binds STAT1; AF3 ipTM = 0.78), peptide transport, ER maturation, trafficking, and degradation. Many hits are bipartite: a transmembrane helix for localization plus a short domain or short linear motif (SLiM) that does the work.
Perturb-seq defined 24 clusters (433 smORFs) and 752 singletons. Three patterns sit side by side: homologs with shared function, non-homologs that converged, and homologs that split. HPV E7 and adenovirus E4 ORF6/7 share no sequence homology, yet both bind the Rb-E2F repressor and drive S-phase entry. A separate cluster groups HIV Vpr, SARS-CoV-2 ORF6, monkeypox MPXVgp154, YLDV132R, and an unannotated HMO astrovirus smORF: all blunt IFN-α/γ reporters by disrupting nucleocytoplasmic transport. AF3 placed C-terminal residues of ORF6, MPXVgp154, and YLDV132R against the Rae1 cargo pocket; an MSA recovered a shared DDXXMEID-like SLiM. Methionine-to-alanine mutants (ORF6 M58A, MPXVgp154 M69A, YLDV132R M75A) lost the IFN-inhibition phenotype. Vpr-Rae1 models scored poorly (ipTM < 0.4); Vpr tracks with importins instead.
The astrovirus smORF is overprinted in the +1 frame of capsid ORF2, with a start codon only 19 bp downstream of ORF2. Close relatives lack that start, consistent with de novo birth. More distant astrovirus clades independently grew different smORFs at the same locus, with different sequences and functions.
The hardest mechanism is EBV BNLF2b. It speeds growth in RPE-1 and normal oral keratinocytes and supports anchorage-independent colonies in soft agar, a phenotype shared with E7 and not with the other pro-growth candidates. AE-MS ranked PKA catalytic subunits PRKACA/PRKACB at the top. BNLF2b inhibits PKA; the host inhibitor PKIA phenocopies the fitness gain. Downstream signaling is MAPK: pERK and pMEK rise, and the RAS inhibitor RMC-7977 or the MEK inhibitor PD0325901 wipe out the BNLF2b and PKIA advantage without touching E7. AF3 on a disorder-truncated BNLF2b-PKA complex scored ipTM = 0.79, with BNLF2b occupying the substrate cleft via an RRNAL pseudosubstrate motif that mirrors PKIA/PRKAR1A RRXAI, plus an adjacent proline-rich motif. Alanine/lysine scanning and targeted mutants collapse both binding and growth. PKIA cannot drive soft-agar colonies, and PKA-binding-defective BNLF2b mutants keep a partial anchorage-independent phenotype. RNA-seq supplies the second arm: ECM/EMT genes (RUNX2, COL1A1, FN1 and others) go up; anoikis-linked genes go down. The paper calls this mimicry-plus: copy a host kinase inhibitor to turn on MAPK, and run a separate transformation program.
The default that unannotated short ORFs are mostly noise does not survive this screen. Seventy percent of the 2,028 secondary hits lack annotation; 75% of strong Perturb-seq phenotypes are unannotated or barely annotated. For virology and protein design, SLiMs plus transmembrane helices are portable parts. Rae1 and TAP keep being hit by distant viruses, which marks them as weak points in the host network.
BNLF2b also closes a clinical gap. Anti-BNLF2b antibodies were already a nasopharyngeal-carcinoma screening marker with no assigned function. In epithelial cells the protein runs PKA-MAPK, which is a reason to test MAPK-targeted drugs in EBV-associated epithelial cancers, still untested in tumor models. Once Perturb-seq scales, the pan-viral proteome can be reorganized by function rather than sequence. For virtual-cell work that wants perturbations other than CRISPR, viral ORFs push cells into states those tools do not.
The atlas undercounts. Microproteins that need other viral cofactors drop out; EBV BDLF3.5, which acts inside the viral pre-initiation complex to drive late lytic genes, scored in none of the screens. Oncogenic phenotypes are acutely cell-type and media dependent. K562 barely shows pro-growth hits, so the same mechanism in another model could be missed. Growth as the funnel entrance already biases the set: 40% of MHC-I hits had no growth phenotype.
Ectopic expression is its own confounder. The pTRE promoter may fail to accumulate short, unstable peptides (false negatives) and can overshoot physiology (artifacts). Length- and biophysics-matched negative controls bound the background; they do not replace infection. Host-virus AF3 models lack coevolution signal. The ipTM > 0.75 cutoff throws away many real interfaces and still leaves predictions. Mutants back the Rae1 and PKA keys; most other molecular keys do not have that support. BNLF2b tumor evidence stops at overexpression in RPE-1 and oral keratinocytes. There is no EBV-infection transformation model and no causal data from patient tissue.