Integrative multiomics analysis and CRISPR screening identify functional noncanonical translation loci in the mouse immune system

Abstract

Ribosome profiling has revealed thousands of noncanonical translation events across mammalian genomes, yet functional characterization has overwhelmingly focused on proliferative fitness in cancer cell lines. Here, we present a comprehensive survey of noncanonical translation in the mouse immune system and its functional consequences in macrophages. By performing a unified Ribo-seq meta-analysis across 20 public mouse leukocyte datasets - spanning macrophages, dendritic cells, neutrophils, B cells, and T cells - we define a compendium of 22,276 noncanonical coding sequences (CDSs), including upstream ORFs (uORFs), downstream ORFs, and ORFs on noncoding RNAs and pseudogenes (ncORFs). Proteogenomic integration with reanalyzed mass spectrometry data prioritizes a high-confidence subset with detectable protein products, including pseudogene-encoded and lncRNA-encoded zinc finger proteins. To move beyond cataloging, we carried out two orthogonal CRISPR screens in immortalized bone marrow-derived macrophages: a fitness screen identifying noncanonical CDSs required for macrophage viability, and a TLR1/TLR2-NFκB reporter screen uncovering CDSs that modulate innate immune signaling. These screens nominate uORFs, several conserved between mouse and human, that exert phenotypic effects on par with their cognate coding sequences. We unexpectedly discovered a family of endogenous retroviral envelope-derived proteins translated in adult myeloid cells. Among these, SYNIR is a full-length syncytin-like membrane glycoprotein that positively regulates NFκB-responsive transcription, while SEMR is a secreted protein with structural homology to the feline leukemia virus accessory protein FeLIX that drives broad transcriptional remodeling of macrophage gene programs upon knockout. Updated single-cell RNA-seq annotations and an interactive UCSC Genome Browser session integrating Ribo-seq, proteomics, and CRISPR screen data are provided as community resources. Together, these findings expand the functional landscape of noncanonical translation in immunity and establish endogenous retroviral proteins as previously unrecognized regulators of macrophage biology.

Introduction

Over the past decade, ribosome profiling (Ribo-seq) has revealed that translation occurs at thousands of sites beyond canonical coding sequences. In mRNAs this includes translation of upstream open reading frames (uORFs) in the 5’ untranslated region (UTR), internal ORFs (iORFs) nested in the coding sequence (CDS), and downstream ORFs (dORFs) in the 3’ UTR. Beyond annotated coding genes, translation of ORFs embedded at loci annotated as pseudogenes and noncoding RNAs (ncORFs) has also been repeatedly documented^1–4. In this report we will refer to the subset of ORFs, which are predicted to be translated, as noncanonical coding sequenes (nCDSs). These nCDSs give rise to a “dark proteome”⁵ whose scope continues to expand with improved detection methods and deeper sequencing coverage, challenging classical definitions of coding potential and raising fundamental questions about which translated sequences carry biological function.

Early genome annotation efforts relied heavily on sequence length and evolutionary conservation, establishing a 100-codon threshold that systematically excluded short ORFs from reference gene sets⁶. As a consequence, there may be thousands of unannotated small proteins and regulatory peptides, arising from misclassified “noncoding” transcripts. Computational rescoring of genomic sequences, combined with advances in Ribo-seq and mass spectrometry-based proteomics, has since identified hundreds of translated nCDSs with important phenotypic effects, including defined roles in inflammation^7–9, skeletal muscle activity^10,11, and cell differentiation¹². These discoveries have motivated efforts to catalog noncanonical translation events and to integrate them into updated genome annotations¹³.

Despite progress in cataloging nCDSs^2,14–16, functional characterization has lagged substantially behind discovery. The vast majority of Ribo-seq-detected translation events remain uncharacterized with respect to biological activity. Moreover, many small, nCDS translation products are likely to be unstable, degrading shortly after production, and not act as functional effectors^17,18. Recent efforts have been made to ascribe biological function to some of the tens of thousands of predicted nCDSs with multiple high-throughput perturbation studies linking these ORFs to phenotypes. However, these screens have predominantly tested essentiality in proliferating cancer cell lines and stem cells^19–23, probing nCDS importance in a relatively narrow context. A notable exception is a targeted Perturb-seq experiment in K562s which provided a rich readout of altered gene programming following nCDS knockout, albeit at reduced throughput compared to pooled screens¹⁹. Thus, while the laudable goal of identifying cancer driving noncanonical translation events has helped contextualize some of these nCDSs, the field has not witnessed similar efforts to apply high throughput screens in other biological contexts.

Immune cells represent a particularly compelling context in which to explore noncanonical translation²⁴. Leukocytes must rapidly adapt their proteomes in response to inflammatory cues, pathogen encounter, and tissue damage. This plasticity is regulated in part at the translational level, with global reprogramming of ribosome occupancy occurring within hours of immune activation^25,26. uORFs have emerged as key regulatory elements in this context, modulating translation of downstream CDSs for cytokines, transcription factors, and stress-response genes^27–33. Moreover, a subset of long noncoding RNAs and pseudogenes are induced in inflammatory contexts^34,35, and the translational potential of these loci remains largely unexplored.

A related underexplored dimension of noncanonical translation concerns endogenous retroviruses (ERVs), remnants of ancient retroviral integrations that comprise roughly 8–10% of mammalian genomes^36,37. Most ERV sequences have degenerated beyond coding capacity, but a subset retains intact ORFs that can be transcribed and translated³⁸. The best-characterized examples are the syncytins, envelope-derived (Env) genes that have been co-opted for placental development and trophoblast fusion across placental mammalian lineages^39,40. Whether additional ERV ORFs are translated outside the placenta, and whether they contribute to immune function, remain open questions.

Addressing these gaps requires an integrative approach that combines detection of translation events and high-throughput functional assays. Here, we present a unified Ribo-seq meta-analysis of all public mouse leukocyte datasets available at the outset of this study to define a comprehensive catalog of nCDSs translated in the immune system. We merge computational translation predictions from two complementary callers with a custom transcriptome annotation and highlight hits with confident proteomics support. We then carry out two orthogonal CRISPR screens in immortalized bone marrow derived macrophages (iBMDMs) - an essentiality screen for fitness effects and a TLR1/TLR2-NFκB reporter screen for inflammatory signaling - to identify nCDs with measurable biological activity. This strategy nominates translated uORFs and ncORFs that modulate macrophage viability and innate immune activation, including several conserved between mouse and human. Finally, we characterize a family of translated Env proteins expressed in myeloid cells, including a syncytin-like membrane protein that tunes NFκB-responsive transcription and a secreted Env fragment with structural homology to the feline leukemia virus accessory (FeLIX) protein. Together, these findings expand the functional landscape of noncanonical translation in immunity and establish resources - including updated single-cell annotations and an interactive genome browser - for continued exploration of the noncanonical ORFeome.

Results

A unified leukocyte Ribo-seq meta-analysis defines a noncanonical ORF compendium

We analyzed 20 publicly available Ribo-seq datasets spanning mouse leukocyte lineages (neutrophils⁴¹, macrophages^7,42–45, B cells^46–49, CD4^49–54 and CD8 T cells⁵⁵, and dendritic cells^56–58) and kidney tissue⁵⁹ treated with the inflammatory ligand lipopolysaccharide (LPS) (Fig. 1A,B). After preprocessing to remove contaminating reads (Supplementary Fig. 1A, Methods), the majority of retained reads mapped in-frame to annotated coding sequence (CDS) P-sites, supporting expected ribosome occupancy (Supplementary Fig. 1B,C).

Figure 1. Integrated Ribo-seq and proteomics pipeline across immune cells.

(A) Overview of the noncanonical CDS (nCDS) discovery pipeline. Ribo-seq reads were aligned with STAR to a custom transcriptome constructed from GENCODE vM31, RefSeq lncRNAs, and a Nanopore de novo assembly. Translation events were inferred independently using RiboTISH and PRICE. The predicted nCDS peptides were merged with Gencode protein annotations to generate a custom spectral search database for MS-GF+ and ms2Rescore analysis of publicly available proteomics datasets.

(B) Publicly available Ribo-seq datasets analyzed in this study.

(C) Publicly available mass spectrometry datasets.

(D) Shared predictions between RiboTISH- and PRICE-derived nCDSs at each tool's Benjamini Hochberg adjusted p-value < 0.01 cut-off.

(E) Representative nCDS categories identified by the pipeline. Canonical CDS entries and their N-terminal extensions were not considered in subsequent analysis.

(F) Composition of the aggregated nCDS catalog across all datasets.

(G) Reidentification of previously documented nCDSs.

CDS: coding sequence

UTR: untranslated region

lncRNA: long noncoding RNA

uORF: upstream open reading frame

iORF: internal open reading frame

dORF: downstream open reading frame

ncORF: noncoding RNA open reading frame

We generated a custom transcriptome that merges GENCODE vM31, RefSeq lncRNAs, and a de novo Nanopore direct RNA assembly⁶⁰, to enable comprehensive translation discovery (Fig. 1A). Translation events were inferred independently using two complementary callers - RiboTISH⁶¹ and PRICE⁶² (Fig. 1A; Tables S1,S2) - and predicted protein sequences were merged with GENCODE proteins to generate a tailored spectral search database for peptide spectra matching (Fig. 1A,C, Methods). Aggregating predictions across all datasets yielded a catalog of 22,276 nCDSs spanning uORFs, iORFs, dORFs, and ncORFs, while excluding annotated CDS and N-terminal extensions from the noncanonical set (Fig. 1E,F, Table S3). UORFs made up the largest fraction of nCDS calls, consistent with previous reports^19,22,63 and unsurprising given their position upstream of canonical CDSs and preferential access to the pre-initiation complex (Fig. 1F). At matched confidence thresholds (adjusted p-value [padj] < 0.01), PRICE proved more sensitive, identifying a larger number of ORFs and reidentifying 10 out of 10 previously identified nCDSs compared to 8 out of 10 reindentifications with RiboTISH (Fig. 1D,G, Table S4).

Consistent with expectations for noncanonical translation, predicted nCDSs were predominantly short, with many below the historical 100-codon cut-off used in earlier coding sequence annotation⁶ (Supplementary Fig. 2A). In silico localization predictions with DeepLoc⁶⁴ further suggested that proteins encoded by predicted nCDSs differ from canonical proteomes, including an increased propensity for mitochondrial localization and reduced propensity for membrane embedding, consistent with prior nCDSs predictions¹⁸(Supplementary Fig. 2B, Tables S5–7).

Proteogenomic evidence strongly supports translation of pseudogenes and lncRNA-encoded zinc finger proteins

To determine which Ribo-seq-inferred translation events yield detectable protein products, we reanalyzed public mass spectrometry datasets with our predicted nCDS peptide database using MS-GF+⁶⁵ for initial search and MS2Rescore⁶⁶ to recalculate assignments (Fig 1A; Tables S10,S11). High confidence peptide spectrum matches (PSMs) were reported (q-value < 0.01; posterior error probability < 0.1) and we highlighted nCDSs supported by either robust total peptide-spectrum matches (PSMs; ≥5) or multiple unique PSMs (≥2) and visualized their expression across immune lineages using uniformly processed, batch-corrected gene abundance estimates (Fig. 2A; Table S12) (Methods).

Figure 2. Proteomics supports translation of select noncanonical loci.

(A) Heatmap of nCDS with either a minimum of 5 total peptide spectra matches (PSMs) or 2 unique PSMs. Gene expression is uniformly processed and batch corrected with Limma to calculate comparable counts per million (CPM). Final expression for a given cell type is the mean of experimental replicates.

(B) PSMs of unique nCDSs found in macrophage cell lysate mass spectrometry dataset IPX0001245000 which includes macrophages from many tissues. The density cloud represents 10,000 randomly selected PSMs that mapped to protein coding genes. All displayed PSMs are significant at q-value < 0.01, posterior error probability < 0.1 cut-offs.

(C,D) Uniquely mapped bulk RNA and Ribo-seq at ENSMUSG00000121501 (or Gvin-ps7) and Gm12250, interferon induced pseudogenes with strong evidence of transcription and translation.

PSM: peptide spectra match

CPM: counts per million

CDS: coding sequence

Among the most strongly supported examples were interferon-induced loci annotated as pseudogenes, including ENSMUSG000000121501 (Gvin-ps7) and Gm12250. Despite their pseudogene annotation, these loci exhibit uniquely mapped bulk RNA and robust Ribo-seq signal across their dominant ORFs, alongside proteomic evidence of translation in macrophages (Fig. 2B–D). We also identified a cluster on chromosome 13 in which multiple transcripts annotated as long noncoding RNAs neighbor the protein-coding zinc finger gene Zfp825. Once again, these loci showed convergent evidence of translation: uniquely mapped Nanopore reads, bulk RNA and Ribo-seq coverage, unique PSMs, and predicted KRAB/zinc finger domains consistent with ZFP-like proteins (Supplementary Fig. 3A–E). Together, these cases argue that a subset of “noncoding” or “pseudogene” annotations reflect genuine protein synthesis, supported by orthogonal readouts spanning transcription, ribosome occupancy, and peptide detection.

While proteomic support provides especially compelling evidence for translation, we emphasize that shotgun mass spectrometry remains biased toward abundant, stable proteins and is expected to miss many translated products - particularly small, lowly translated, and unstable proteins. In line with this limitation, while our pipeline reidentified multiple previously published translated ORFs with Ribo-seq (Fig 1G, Table S4), we did not identify confident PSMs that matched the predicted proteins from any of these 10 nCDS. Thus, proteogenomic integration serves primarily to prioritize a subset of translation events, rather than to define the full scope of biologically meaningful noncanonical translation.

A pooled macrophage essentiality screen identifies ORFs that contribute to cellular fitness

To functionally prioritize nCDSs, we designed a targeted CRISPR library against candidates with evidence of transcription and Ribo-seq translation in macrophages (Table S13) using the CRISPRware guide RNA library design package⁶⁷. We targeted 2,165 uORFs and 63 dORFs across 984 protein-coding genes and 1,093 ncORFs across 400 ncRNAs and pseudogenes. Each ORF was targeted by at least 4 independent guide RNAs (gRNAs). In some cases, ORFs overlapped due to out-of-frame translation and gRNAs, could not be uniquely associated with a single target; in these cases the co-targeted ORF IDs were concatenated, e.g. Emsy uORF1:Emsy uORF2, and treated as a single targeted unit (Fig. 3C). For genes harboring targeted uORFs, we additionally included gRNAs against the cognate downstream CDS to enable paired comparisons at shared loci (Methods). GRNAs were cloned into lentivirus and transduced at low multiplicity of infection into our previously published immortalized bone marrow-derived macrophage lines, which constitutively express Cas9 and carry an NFκB-GFP reporter cassette⁶⁸ (Fig. 3A, 4A). Following selection and outgrowth in three distinct clonal lines, we quantified changes in guide abundance and inferred gene-level effects on cellular fitness using MAGeCK-MLE⁶⁹.

Figure 3. Dropout screen against nCDSs with predicted translation in bone marrow derived macrophages (BMDMs).

(A) CRISPRware designed a gRNA library against nCDSs on transcripts expressed in BMDMs (>0.1 TPM). Each ORF was targeted with at least 4 guide RNAs. Where uORFs were targeted, the canonical coding sequence (CDS) was also included. After pooled transduction into immortalized CRISPR-Cas9 BMDMs, cells were selected with puromycin for seven days, and genomic DNA was taken 14 days subsequent to selection.

(B) MAGeCK-MLE dropout results for nCDS and canonical CDS hits from the pooled CRISPR dropout screen. ORFs with padj < 0.05 and Wald-p-value < 0.05 were considered significant hits. To enhance visualization, beta scores < −1.5 were capped at −1.5 and Wald p-values < 1E-6 were set to 1E-6. Full values are available in the associated table.

(C) Comparison of CDS and uORF effects identifies genes where disrupting the upstream ORF and/or the canonical CDS contributes to fitness.

(D,E) NCDS loci illustrating gRNA dropout (mean from triplicates), Ribo-seq support, ORF structure, and conservation with human sequence. Mouse Ribo-seq is combined from all studies (Figure 1). Human THP-1 Ribo-seq was mapped to Hg38 and lifted to mouse Mm39. (D) Kat6a both uORF and cognate CDS passed our threshold for significance. (E) For Zdhhc5 the uORF passed our significance threshold (padj = 2.3E-10, Wald p-value = 9E-3) but the cognate CDS did not (Wald p-value > 0.05).

TPM: Transcript per million

iBMDM: immortalized bone marrow derived macrophages

MOI: multiplicity of infection

Figure 4. TLR1/2-NF-κB sorting screen results.

(A) Activation of the Tlr1/Tlr2-NFkB inflammatory cascade induces GFP production in NFkB-GFP immortalized bone marrow derived macrophage (iBMDM) line. Cells were sorted into GFP high (top 20%) and GFP low (bottom 20%) buckets by GFP expression after stimulation with Pam₃CSK₄. Guide RNA and gene level z-score were calculated with MAUDE (Mean Alterations Using Discrete Expression). Negative z-scores indicate positive regulators of NFkB signaling, positive z-scores indicate negative regulators.

(B) The MAUDE calculated z-scores of known positive and negative regulators of Tlr1/Tlr2-NFkB signaling cascade.

(C) ORFs are considered hits if the direction of the effect is the same for all replicates, the absolute values of the combined z-score are > 2.5 and the padj. < 0.05. nCDS combined z-scores across triplicates calculated with Stouffer’s method. Z-scores are log transformed to improve visualization.

(D) The results of targeting nCDS across both screens. The majority of significant hits are screen-specific. To enhance visualization, beta scores < −2 were capped at −2 and z-scores beyond +/−10 were capped at −10 and 10. Full values are available in the associated table. (E,F) NCDS loci illustrating gRNA combined z-score and dropout (Stouffer’s mean from triplicates), Ribo-seq support, ORF structure and conservation with human ORF where applicable. Mouse Ribo-seq is combined from all studies (Figure 1). Ranges are raw signal counts. (E) Tirap uORF 4 is grouped with negative regulators of NFkB (mean z-score > 2.6, padj < 3.2E-2) while Tirap primary coding sequence (mean z-score < −5.3, padj < 1.8E-6) is grouped with positive regulators. Tirap uORF 3 is highly translated, but does not show a consistent effect across replicates. Signal track is log-transformed to improve visualization.

(F) Ints3 uORF was a significant hit in both screens (dropout: Wald p-value < 0.02, padj. < 5E-4 ; sorting: mean z-score > 7.8, padj < 9.3E-14) as a growth suppressor and negative NFkB regulator, while Ints3 was not a significant hit in either screen.

Encouragingly, an inspection of the top gene hits included common essential genes such as Polrmt and Rpl27(depmap.org)⁷⁰ as well as genes that have demonstrated essentiality specifically in macrophages: Spi1⁷¹, Bach1⁷², Ikbkb⁷³, and Zeb2⁷⁴. On the other hand, pro-apoptotic Bcl2l11⁷⁵ and tumor suppressor Rbm47⁷⁶ genes were significantly enriched targets (Fig. 3B). To minimize false positives, we defined confident hits using dual criteria (permutation padj. < 0.05 and Wald p-value < 0.05), a threshold that proved sufficiently stringent to exclude genes with described effects on viability, such as the pro-apoptotic factor Bbc3, from the high-confidence set (Fig. 4B; Tables S14,S15). Two previously reported noncanonical ORFs, Dleu2 ncORF3 and Ppp1r15a uORF1 (Fig. 1F, Table S4), were recovered as significant dropout hits (Fig. 4B), nominating a potential role in macrophage viability in addition to previously ascribed functions^12,33.

Next we leveraged the paired design to compare uORF-targeting gRNAs with gRNAs against cognate CDS regions: in 19 loci, disruption of either the uORF or the CDS significantly impaired viability, whereas in 8 loci, uORF targeting produced a significant dropout phenotype in the absence of a significant CDS effect (Fig. 3C). We examined four representative uORF hits - Emsy uORF 1, Kat6a uORF 2, Zdhhc5 uORF 2, and Brd2 uORF 4 - and observed consistent negative log fold changes across their targeting guides, consistent with contribution to macrophage fitness (Fig. 3E,F; Supp. Fig. 4A,B). Each of these uORFs is conserved in the orthologous human gene, suggesting selective constraint on the translated element. In three of the cases, published Ribo-seq reads from the human monocytic THP-1 line, when lifted to mouse genome build mm39 coordinates (Methods), recapitulated the ribosome occupancy patterns observed at the corresponding mouse uORFs (Fig. 3E,F; Supp. Fig. 4A), supporting conserved translation across species and strengthening the case that these uORFs are important elements in myeloid biology broadly.

A TLR1/TLR2-NFκB sorting screen reveals noncanonical ORFs that tune innate immune signaling

The same transduced cells were used for an orthogonal screen that directly reads out NFκB pathway activity. At day 14, we stimulated cells with Pam₃CSK₄, a proinflammatory bacterial lipopeptide ligand that binds the TLR1/TLR2 heterodimer, triggering canonical NFκB activation and inducing GFP expression via an NFκB-GFP expression element^68,77. We then performed fluorescence-activated cell sorting to isolate GFP-high and GFP-low populations and sequenced gRNAs in each bin, enabling inference of nCDS and CDS elements that influence this signaling pathway (Fig. 4A).

We analyzed each replicate independently using MAUDE⁷⁸ to estimate a z-score for each targeted element, then filtered candidate hits to require directional consistency across replicates (all z-scores either positive or negative for a given ORF), after which replicate-level information was aggregated (Methods). We defined significant NFκB modulators as those with |z-score| > 2.5 and padj < 0.05 (Fig. 4C; Tables S16-S18). Many top-scoring hits corresponded to established regulators of TLR-NFκB signaling, supporting assay fidelity (Fig. 4B). Importantly, the thresholds remained conservative enough to exclude several likely pathway components that did not meet significance, including Nfkb2 (z-score = 2.21, padj = 0.068), Traf1 (z-score = −2.19, padj = 0.069), and Irak3 (z-score = −2.05, padj = 0.080), indicating that the reported set reflects stringent filtering (Table S18).

To assess whether generalized dropout phenotypes could present as NFκB regulators, we compared effect profiles between the fitness and sorting screens. Overall, the two hit sets were notably independent, indicating that most NFκB modulators were not simply genes required for growth in culture, and conversely that viability hits did not significantly bias GFP output (Fig. 4D). An exception was a short, 14-codon uORF in Ints3, which emerged as a strong hit in both assays despite the absence of a corresponding phenotype for the cognate Ints3-CDS, suggesting a locus where a small translated element may exert disproportionate regulatory impact (Fig. 4F). Conversely, analysis of Irak1 uORF1, our top nCDS hit in the sorting screen, highlighted an important caution for interpreting short-ORF perturbations: the highly active gRNAs substantially overlapped the embedded microRNA mir-718, emphasizing the need to consider local genomic context and to integrate multiple lines of evidence when attributing phenotypes to a given ORF (Sup. Fig. 4A).

Finally, we highlight two instructive uORF patterns that illustrate how nCDS elements can diverge from their cognate CDSs: a conserved uORF in the transcription factor Nrf1 gene produced a stronger effect on NFκB-GFP output (z = −5.1, padj < 4×10⁻⁶) than the Nrf1 CDS (z = −3.2, padj < 9×10⁻³) (Sup. Fig. 4B), and Tirap uORF4 (z = 2.7) exhibited an effect in the opposite direction of its cognate CDS (z = −5.3). Although Tirap uORF3 showed strong evidence of translation in Ribo-seq, its replicate effects were not directionally consistent and it was therefore excluded from the final hit set, reflecting the conservative filtering used to prioritize robust regulators (Fig. 4E). Together these examples demonstrate that uORFs can exert measurable impact on a critical signaling pathway, with evidence that they sometimes exhibit effects as profound, or in opposite directions, to their cognate coding sequence.

A syncytin-like envelope protein - SYNIR - positively regulates NFκB-responsive transcription

One sorting-screen hit was a 596-codon nCDS on an uncharacterized long noncoding RNA identified as ENSMUSG00000120992 (or D17H6S56E-5) (z-score = −2.8, padj = 0.024). In addition to being a sorting screen hit, translation of this lncRNA was supported by multiple independent lines of evidence: strong Ribo-seq signal across the ORF, robust transcription in macrophages, and 14 unique peptide-spectrum matches that map to the predicted protein sequence (Fig. 2A, 5A).

Figure 5. Synir (ENSMUSG00000120992 ncORF 3) encodes a retrovirus-derived membrane-bound protein with homology to Syncytin-1 and modulates NF-κB signaling in bone marrow derived macrophages (BMDMs).

(A) Synir genomic locus with Ribo-seq, BMDM RNA-seq, peptide spectra matches, and predicted protein domains.

(B) Reference UMAP of leukocyte single-cell RNA-seq colored by cell type for context in panels (C-H).

(C,D) Expression of Synir (596 codon ORF) (ENSMUSG00000120992/D17H6S56E-5) and AI506816 (645 codon ORF) which are translated to produce a transmembrane protein with classic retroviral envelope signatures.

(E) Expression of B230354K17Rik which harbors a translated 493 codon ORF with homology to Friend virus susceptibility protein 1 - a protein thought to arise from retroviral gag protein endogenization.

(F,G) Expression of Brip1os (282 codon ORF) and D430020J02Rik (226 codon ORF) which encode truncated viral proteins with N-terminal signal peptides, but lack the transmembrane viral protein domain.

(H) Expression of Syncytin A, a fusogenic protein derived from retroviral envelope endogenization with expression restricted to placental trophoblasts.

(I) Shared domains and features between SYNIR protein and human Syncytin-1. Signal peptide predicted by Signal6P and DeepTMHMM. Transmembrane domain predicted with DeepTMHMM. Furin cleavage site predicted with ProP. Other domains predicted by Interproscan.

(J) Syncytins are membrane bound proteins with the SU and TM domains secured by a disulfide bond following cleavage at the furin site. Once membrane bound, the proteins form homotrimeric structures.

(K) Gene expression changes following CRISPR Cas9 KO with 3 unique gRNAs. Significant differentially expressed genes (n): padj < 0.01, |log₂(FC)| > 1.

(L) Gene expression changes following CRISPR Cas9 KO with 3 unique gRNAs and stimulated for 6 hours with LPS. Significant differentially expressed genes (n): padj < 0.01, |log₂(FC)| > 1. Enlarged colored dots highlight genes to broad pathways found in literature search.

CPM: counts per million

ISD: immunosuppressive domain.

HR1: heptad repeat 1 domain

HR2: heptad repeat 2 domain

LPS: lipopolysaccharide

We examined predicted protein domains and discovered protein architecture consistent with a retrovirus-derived envelope protein, including a signal peptide, an extracellular surface unit (SU), and a transmembrane domain (TM) (Fig. 5A). Together, these features suggested that the encoded protein is a full-length, membrane-spanning envelope-like (Env) protein. To place this gene in context, we performed a BLASTP⁷⁹ search against mouse and human proteins and identified human Syncytin-1 as the closest characterized homolog (e-value < 5e⁻¹⁰). Syncytins are endogenized retroviral genes that are required for formation of the placenta^39,40,80. Given the structural similarity to Syncytin-1 and the regulation of NFκB signaling, we renamed this protein SYNIR (SYNcytin-like Immune Regulator), used hereafter. A closer inspection of SYNIR uncovered remarkable conservation of hallmark features with Syncytin-1. These included canonical cysteine motifs consistent with disulfide-stabilized SU/TM organization, a predicted furin cleavage site that would separate SU and TM, and motifs indicating a homotrimeric interface and an immunosuppressive domain (ISD) (Fig. 5I,J). These shared structural and functional signatures support a model in which SYNIR represents a retrovirus-derived Env protein that has been retained and repurposed in the mouse genome. We next tested whether SYNIR functionally contributes to inflammatory signaling, consistent with its negative z-score in the TLR1/TLR2-NFκB sorting screen (Fig. 4C). We disrupted the SYNIR ORF using three independent gRNAs in our Cas9 iBMDM line and performed RNA-seq at baseline and following stimulation with the inflammatory ligand LPS (Tables S19, S20). In untreated cells, SYNIR knockout induced a focused transcriptional response (27 differentially expressed genes; padj < 0.01, |log2FC| > 1; Fig. 5K). Notably, three dysregulated genes - Ccr5, Amotl2, and Sult1a1 - have reported connections to retroviral infection raising the possibility that SYNIR may influence susceptibility of host to viral infection^81–83.

Following 6 h LPS treatment, SYNIR disruption again altered the transcriptional state (23 differentially expressed genes; padj < 0.01, |log2FC| > 1; Fig. 5L). Two genes induced in the knockout, Slpi⁸⁴ and Mturn⁸⁵, encode proteins reported to negatively regulate NFκB pathway activity, and their upregulation in SYNIR-deficient macrophages is consistent with the sorting screen result, indicating SYNIR behaves as a positive regulator of the pathway (Fig. 4C). Notably, the sorting screen stimulus (Pam₃CSK₄; TLR1/TLR2) and the validation stimulus (LPS; TLR4) are distinct upstream inputs, suggesting SYNIR influences NFκB signaling at a point of convergence rather than in a TLR receptor-restricted manner. In addition to the direct regulators of NFκB, six of the differentially expressed genes have been linked to alterations in M1-proinflammatory or M2-antinflammatory macrophage polarization programs, and these could further modulate the NFκB pathway. Collectively, these data establish SYNIR as a translated, retrovirus-derived Env protein and demonstrate that SYNIR supports NFκB-responsive transcription in macrophages.

Updated scRNA-seq references expand the landscape of envelope-like translated loci across immune and tissue contexts

The identification of SYNIR as a protein of retroviral origin encouraged us to search for additional Env translation events. Across the aggregated Ribo-seq predictions, we identified four lncRNA loci with ORFs whose predicted proteins exhibit Env features: ENSMUSG00000120992 (SYNIR), D430020J02Rik, Brip1os, and AI506816. One additional Env gene, Gm33887, was previously predicted in the RefSeq proteome database (it is annotated as a lncRNA in Gencode, and our pipeline did not identify it as translated). Because envelope-derived genes are typically discussed in the context of placental formation and fetal development⁸⁶, we sought a broad and unbiased view of expression across tissues and immune lineages. We therefore reanalyzed Tabula Muris⁸⁷ and Tabula Senis⁸⁸ single-cell RNA-seq references using an updated gene annotation that includes lncRNAs, expanding quantification well beyond the protein-coding gene set used in the initial releases (Methods). This reanalysis provides a practical resource for interpreting lncRNA specificity across adult tissues at three ages (3-,18-, and 24-months; >110,000 total cells post-filtering) beyond the present investigation (Data Availability).

In these updated scRNA-seq embeddings, the Env lncRNA loci showed distinct and nonredundant expression patterns across immune subsets and tissue contexts (Fig. 5B–F; Supp. Fig. 6A–F), in contrast with Syncytin-A, which demonstrated virtually no transcription in these non-placental cells (Fig. 5G). Across genes, we observed marked differences in both the magnitude of expression and the cell types in which expression concentrates, supporting a diversified landscape of envelope-derived loci with lineage- and tissue-biased expression in adult mice (Fig. 5C; Supp. Fig. 6A–F).

Long-read Nanopore sequencing resolves unannotated retroviral Env gene models and isoforms

Our custom transcriptome incorporated a de novo Nanopore direct RNA assembly from BMDMs⁶⁰, enabling us to identify genes and isoforms that are difficult to resolve with short reads alone. Within this long-read-informed annotation, we identified additional Env ORFs that were absent in Gencode. Three loci emerged as high-confidence Env genes (Supp. Fig. 6A–D). First, XLOC000481 corresponds to a novel Ctse isoform defined by an alternative second exon that is strongly represented in long-read data relative to canonical Ctse isoforms (Supp. Fig. 6A). A Sashimi plot of short-read junction spanning reads corroborated this isoform structure and relative usage (Supp. Fig. 6B). Second, XLOC012421 is an unannotated, highly transcribed gene antisense to Cnbd2 (Supp. Fig. 6C). Third, XLOC008059 is an extended second exon model of the annotated lncRNA Gm62403 (Supp. Fig. 6D). Unlike the annotated Env genes above, these three loci exhibit extreme sequence similarity exemplified by an absence of uniquely targeting CRISPR guides. We observed only low Ribo-seq signals at the loci and RiboTISH, but not PRICE, identified the Env ORFs at these loci as translation events. We attribute this to the challenges associated with uniquely mapping ~30-nt, single-end Ribo-seq reads to repeated genomic elements. This limitation is overcome with long reads, but we are only able to confidently report on transcriptional, not translational, activity at these loci - a reminder of the limits of current approaches.

An additional retroviral-derived gag-like protein has been discovered on the B230354K17Rik lncRNA.

A final notable novel retroviral-related hit with convincing proteomics support was the 493 codon ORF on lncRNA B230354K17Rik (Supp. Fig. 8A). The encoded protein is predicted to have a Friend virus susceptibility domain, homologous to the protein encoded on Fv1, a gene that is known to control infection by Friend leukemia virus and which is derived from an endogenized gag gene⁸⁹. We predicted a structure with Alphafold 3⁹⁰ and found that it produced a high confidence globular domain and a lower confidence extended alpha helix with the two connected by disordered linkers (Supp. Fig. 8B). We performed a FoldSeek⁹¹ structure search on the globular domain and identified Fv1 and an uncharacterized globular protein, 2410002F23Rik as closely related hits. We again looked at global expression and observed that both B230354K17Rik and 2410002F23Rik were widely expressed (Supp. Fig. 9A–D). As retroviruses typically exhibit strong preference for infecting specific cell types, this wider expression may indicate functions that diverge from a purely anti-viral, defensive role.

Brip1os encodes a secreted envelope-like protein with FeLIX homology and broad paracrine transcriptional impact.

Among the Env candidates, D430020J02Rik and Brip1os were notable outliers in predicted protein length: both encoded proteins less than 300 amino acids long, less than half the length of classical Env proteins. Domain prediction shows that these proteins preserve an N-terminal signal peptide and SU-like envelope domain, but lack a transmembrane segment, consistent with secreted, rather than membrane-bound, proteins (Fig. 6A). Mirroring our approach with SYNIR we used BLASTP to search for homologs, but did not find characterized proteins in mouse or human. Next, we used Alphafold to predict structures and found that both produced high-confidence β-sandwich-like architectures. We used Foldseek to search structural databases and identified an ortholog in FeLIX, a previously described feline retrovirus-derived accessory protein with documented roles in modulating the ability of feline leukemia virus to infect host cells^92,93 (Fig. 6B).

Figure 6. Brip1os encodes a retrovirus-derived secreted protein, SEMR, in bone marrow derived macrophages (BMDMs) with immunomodulatory function and homology with FeLIX protein.

(A,B) Predicted domains and AlphaFold structure prediction of Feline Leukemia virus accessory protein (FeLIX), SEMR (Brip1os ncORF7), and D430020J02Rik ncORF 1. Foldseek search reveals close structural homology between the three proteins.

(C) Brip1os genomic locus with Ribo-seq, BMDM RNA-seq, peptide spectra matches, and predicted protein domains.

(D) Unique peptide-spectrum matches (PSMs) from macrophage secretome datasets. PSMs uniquely mapping to the SEMR protein sequence are labelled.

(E) Volcano plot of CRISPR–Cas9 knockout of Brip1os ncORF7 in BMDMs treated with LPS for six hours induces broad transcriptional remodeling.

(F) Number of significant gene changes in immune and metabolic Reactome pathways after SEMR loss.

Significant genes (n): padj < 0.01, |log₂(FC)| > 1.

LPS: lipopolysaccharide

Brip1os, in particular, emerged as an attractive candidate for functional follow-up because of its widespread and consistent expression across tissues (Fig. 5C; Supp. Fig. 5B), and as the most prominent nCDSs hit in macrophage secretome proteomics datasets (Fig. 6D). We asked whether the Brip1os encoded protein has a measurable regulatory impact in cultured macrophages. Using three independent gRNAs, we disrupted the Brip1os ncORF7 in iBMDMs and performed RNA-seq at baseline and after 6 hours of LPS stimulation. KO induced extensive transcriptional remodeling in both conditions (Fig. 6E,F; Supp. Fig. 8A,B; Tables S21,S22), with hundreds to thousands of differentially expressed genes (padj < 0.01, |log2FC| > 1). The genes most prominently affected span core macrophage programs, including metabolic and hypoxia-linked regulators (e.g., Hif1a, Bach1), inflammatory and NFκB-associated transcriptional circuitry (e.g., Nfkb1, Junb, Rel), lineage determinants (e.g., Zeb1, Zeb2), and broad chromatin and transcriptional regulators (e.g., Ctcf, Kmt2a) (Fig. 6E; Supp. Fig. 8A). Pathway-level analysis highlighted enrichment of Reactome⁹⁴ pathways related to metabolism and immune activity (Fig. 6F; Supp. Fig. 8B). Based on this protein’s secretion, derivation from a retroviral envelope protein, and impact on cellular metabolism, we name this protein SEMR (Secreted Envelope-like Metabolic Regulator). Notably, despite the profound changes in this targeted KO experiment, SEMR was not a significant hit in either screen. This observation, along with the overwhelming evidence that the SEMR protein is secreted, implies that it functions as a paracrine signaler, and identifying its requisite receptor(s) presents an intriguing future opportunity.

An interactive genome browser resource enables rapid community interrogation of noncanonical ORFs and functional evidence

A core theme throughout this study is that confident interpretation of noncanonical translation requires synthesis of multiple orthogonal signals: Ribo-seq evidence, peptide-spectrum support, transcription, protein domain predictions, conservation, and evidence from CRISPR screens. To make these layers directly interrogable by the community, we created an interactive public UCSC Genome Browser^95,96 session that aggregates the processed data into a navigable format (Fig. 7A–E).

Figure 7. Interactive and customizable public Genome Browser session for further inquiry.

(A) Example of the Leukocyte ORF Genome Browser session at the Brip1os locus. The session includes (from top to bottom) guide RNA effect sizes, mouse and human Ribo-seq, protein domain predictions, peptide spectra matches, nCDS predictions, Gencode gene annotations, bulk RNA-seq and cross-species conservation

(B) Guide RNA effect sizes for both dropout and sorting screen can be displayed as means or per replicate.

(C) Ribo-seq tracks can be set to show mouse (aggregate leukocyte) and human (THP1) signal tracks from elongation inhibition experiments (cycloheximide) as well as signal from translation initiation inhibition experiments (lactimidomycin or harringtonine).

(D) Proteomics tracks include domain predictions from SignalP (secreted peptides), DeepTMHMM (secreted peptides and transmembrane proteins), Interproscan (protein domain database aggregator) and peptide spectra matches from comparison

(E) NCDS prediction tracks include merged PRICE and RiboTISH calls as well as per-tool tracks and configurable display of transcriptome sources.

The session is built on the current mouse reference genome (mm39) and is designed so that any locus can be queried via standard gene-name or chromosome coordinate search. The browser view integrates, in a single place, gRNA effect sizes from both essentiality and sorting screens (with options to view replicate-level or mean-level tracks), aggregate mouse leukocyte Ribo-seq tracks, including separate entries for elongation (cycloheximide treated) and initiation (harringtonine and lactimidomycin). Human THP-1 Ribo-seq tracks mapped to hg38 and lifted to mm39, proteomics evidence including peptide-spectrum matches, and protein feature predictions such as secretion signals and transmembrane domains from Signal6P⁹⁷, DeepTMHMM⁹⁸, protein family domains from Interproscan/Pfam^99,100 and intrinsically disordered regions from MobiDB-lite3¹⁰¹ (Supp. Fig. 7A–D). By lowering the barrier to locus-level evaluation, this resource is intended to accelerate hypothesis generation and prioritization for follow-up studies on the many additional translated ORFs uncovered in this work but not yet functionally characterized. Hosting on the highly reliable UCSC Genome Browser ensures that this work will remain a stable and accessible long-term resource.

Discussion

Here we present a unified Ribo-seq meta-analysis across mouse leukocyte lineages and implement dual orthogonal CRISPR screens - one measuring cellular fitness, the other NFκB activity - to identify noncanonical coding sequences (nCDSs) with measurable biological activity in macrophages. This strategy nominates uORFs and ncORFs that modulate viability and innate immune signaling, and unexpectedly uncovers multiple retrovirus-derived envelope proteins expressed and translated outside the placenta, including SYNIR, a membrane protein that tunes NFκB-responsive transcription and SEMR, a secreted protein with broad paracrine impact on macrophage gene programs.

Expanding the functional landscape of noncanonical translation beyond proliferative fitness

The finding that uORFs can exert phenotypic effects on NFκB signaling that are on par with, or opposite to, their cognate coding sequences has important implications. Classical models of uORF function emphasize translational repression of the downstream CDS, wherein uORF translation attenuates main-ORF output by sequestering ribosomes and/or triggering nonsense-mediated decay. In most cases, we observed the opposite with CDS and uORF KO, resulting in directionally consistent phenotypic impact (Fig. 3C,D,E ; Supp. Fig 4A,B; Supp. Fig. 5A,B). In these cases, it is possible that gRNAs targeting the 5’ UTR reduce gene transcription, thereby reproducing the phenotype of disrupting the CDS directly. However, the added evidence of cross-species conservation of the uORFs argues that these are bona fide coding sequences producing peptides with independent phenotypic effects. On the other hand, we present two examples which are not conserved in the human genome, Irak1 uORF1 (Supp. Fig. 5A) and Tirap uORF4 (Fig. 4E). Irak1 uORF1 was the most significant noncanonical ORF hit in the sorting screen, but substantially overlaps mir-718, a microRNA known to reduce NFkB signaling¹⁰². However, the gRNAs targeting the Irak1 uORF1 and mir-718 are consistent with targeting a positive regulator of NFκB signaling, complicating interpretation and leaving open the possibility that the uORF itself is an important player in this pathway. Finally, Tirap uORF4 is grouped with the negative regulators, in contrast with the Tirap CDS, perhaps indicating that disruption of the uORF releases a brake on downstream CDS translation. Whether the peptides encoded by these uORFs act in cis (e.g., by modulating co-translational folding or localization of the nascent main-ORF protein) or in trans (e.g., by interacting with cellular machinery independent of the main ORF) remains an open question that warrants targeted biochemical follow-up.

Retrovirus-derived proteins as an underexplored class of effector proteins

Perhaps the most unexpected finding of this study is the identification of widespread transcription and translation of envelope-derived proteins in adult mouse cells. Endogenous retroviruses comprise roughly 8–10% of mammalian genomes, but only a handful of their coding remnants have been functionally characterized, most notably the syncytins, which are required for placental development^39,40,80. Our data implicate ERV-Envs as important genes in adult cells and tissues indicating expanded roles beyond their canonical placental and fetal expression.

SYNIR (ENSMUSG00000120992) emerged from the NFκB sorting screen as a translated lncRNA encoding a 596-amino acid protein with all the hallmarks of a full-length envelope glycoprotein (Fig. 5I). Its closest characterized homolog is human Syncytin-1, a fusogenic protein with expression restricted to placental and fetal tissue. SYNIR, on the other hand, is highly expressed in adult myeloid, lymphoid, mammary gland, and large intestine tissue (Supp. Fig. 5I). CRISPR KO of SYNIR altered both baseline and LPS-induced transcription, with upregulation of known NFκB pathway inhibitors (Slpi, Mturn) upon knockout. This result was consistent with the sorting-screen linking SYNIR with positive regulation of NFκB activity. While we were able to confirm SYNIR as a screen hit, the main goal of this project, we are intrigued by the fusogenic potential of this syncytin homolog (Fig. 5I). The hallmark of syncytin cell fusion is the generation of multinucleated cells - an important part of placental maturation. Notably, macrophages are also known to form multinucleated cells with pathological function: osteoclasts, Langhans giant cells, and foreign body giant cells¹⁰³. The idea that SYNIR (or other Env proteins reported herein) contributes to this phenomenon is intriguing, but support for this possibility is tempered by expression of this gene in HSCs, B cells, and other cell types without documented multinuclear cell formation. Still, we see this as an exciting opportunity for future study.

Notably, the most up- and down-regulated genes in SYNIR knockout: Ccr5 and Sut1a1, respectively, are orthologs to proteins linked to HIV infection in humans. Ccr5, a well-established co-receptor for HIV, is downregulated, while Sut1a1, the human ortholog of which facilitates HIV reverse transcription⁸², is upregulated. This gives us reason to believe that SYNIR is involved in retroviral entry and replication. Furthermore, this observation resonates with a broader theme in ERV biology: that endogenized viral genes can be co-opted to modulate host susceptibility to exogenous retroviruses. Other germane examples of this phenomenon include Fv1, a gag-derived gene that encodes Friend virus susceptibility protein 1, and which is closely related to an nCDS we identified on the lncRNA B230354K17Rik (Supp. Fig. 8); and Feline leukemia virus susceptibility protein (FeLIX), a secreted Env fragment, which, as its name indicates, is known to module leukemia virus infection in felines, and which is a structural ortholog of another of our documented Env proteins, SEMR.

Because FeLIX has only been studied in the context of modulating feline leukemia virus infection, we were surprised to find that knockout of SEMR induced striking transcriptional changes affecting hundreds of genes - including master regulators of inflammation (Nfkb1, Rel, Junb), hypoxia (Hif1a, Bach1), and lineage identity (Zeb1, Zeb2) (Fig. 6E). This impressive KO response demonstrates that SEMR is not narrowly involved in modulating viral infection as we had hypothesized. Furthermore, as retroviruses exhibited strong cell-specific tropism, Brip1os’s near ubiquitous transcription (Supp. Fig. 6E) is inconsistent with targeted restriction of retrovirus infection. This exciting finding demonstrates that ERV proteins can act beyond retroviral infection modulation and influence critical pathways.

Based on the evidence compiled in this report, SEMR is likely involved in paracrine signaling: in a pooled screen, secreted factors are diluted across the culture, masking signaling effects, and consistent with SEMR not being identified as an essentiality hit. The detection of SEMR as the most prominent nCDS protein in macrophage secretome proteomics and its predicted signal peptides supports a role in extracellular signaling. Identification of the cognate receptor(s) for SEMR is a priority for future work and may reveal new axes of immune regulation.

Considerations for interpreting CRISPR screens targeting noncanonical ORFs

Several features of nCDSs complicate loss-of-function interpretation. First, many translated elements are short, often below 100 codons, leaving limited sequence space for gRNA targeting. GRNA tiling will not always be able to distinguish nCDS-specific effects from regulatory DNA perturbation. The Irak1 uORF1 case illustrates this challenge: the top-scoring gRNAs also disrupted mir-718, confounding attribution of the phenotype. Second, overlapping nCDS predictions can preclude assignment of a phenotype to a single translation product. In many cases, our pipeline identified multiple overlapping uORFs at a single locus, and gRNAs could not be uniquely assigned. In both of these challenging cases, careful consideration of local genomic context can resolve issues of interpretation. To facilitate this process, we organized our findings visually as a UCSC Genome Browser session (Resource availability) that aggregates Ribo-seq, RNA-seq, proteomics, gRNA effect size, and more into a navigable interface.

Limitations and future directions

Several limitations of the present study warrant consideration. First coverage across immune lineages is uneven: macrophages and CD4 T cells dominate the catalog, with fewer datasets representing other cell types. Second, our proteogenomic integration relies on shotgun mass spectrometry, which is biased toward abundant, stable proteins. Many translated products - particularly small, lowly expressed, or rapidly degraded peptides - likely evade detection. The absence of PSMs for 10 previously validated nCDSs in our reanalysis underscores this limitation (Fig. 1G). Third, the mechanistic basis for most hits remains unknown. For uORFs, distinguishing peptide-mediated effects from translational-regulatory effects will require targeted biochemical and genetic dissection.

Resource availability

A central goal of this study was to lower barriers to community engagement with noncanonical translation data. The interactive UCSC Genome Browser session we provide integrates gRNA effect sizes, Ribo-seq tracks, proteomics evidence, and protein feature predictions in a single navigable interface. By making these data layers directly accessible, we hope to accelerate hypothesis generation and follow-up studies on the thousands of translated ORFs identified here but not yet functionally characterized. https://genome.ucsc.edu/s/emalekos/Leukocyte_ORF_paper

The updated single-cell RNA-seq annotations, incorporating lncRNAs absent from original Tabula Muris releases, further extend the utility of this work for the broader research community. https://github.com/ericmalekos/scRNAseq

Code associated with this manuscript is available at

Conclusions

This study demonstrates that noncanonical translation in immune cells extends well beyond leaky scanning or transcriptional noise: a subset of translated uORFs and ncORFs carries measurable biological activity in macrophage viability and inflammatory signaling. The discovery of retrovirus-derived envelope proteins expressed and translated in myeloid cells, including a syncytin-like membrane protein that modulates NFκB and a secreted FeLIX-like factor with paracrine transcriptional impact, expands the known repertoire of endogenized viral genes with host regulatory function. Together with the community resources released here, these findings establish a foundation for continued exploration of the noncanonical ORFeome in immunity and beyond.

Methods

Custom gene annotation

Comprehensive gene annotation v31 was downloaded from Gencode¹⁰⁴. NCBI RefSeq gene annotation ncbiRefSeq.2020–10-27¹⁰⁵ was downloaded, and transcripts with transcript_biotype “lnc_RNA” were extracted. We previously generated a long-read de novo BMDM transcriptome assembly⁶⁰. These three files: Gencode vM31, NCBI Refseq lncRNAs, and BMDM long-read genes were merged with Another GFF Analysis Toolkit’s agat_sp_merge_annotations utility¹⁰⁶.

Mouse Ribo-seq data

GEO datasets were retrieved and FASTQ files were assessed with FastQC and MultiQC¹⁰⁷. Adapters were removed with TrimGalore!¹⁰⁸. Mouse tRNA and rRNA sequences were downloaded from GtRNAdb¹⁰⁹ and RNACentral¹¹⁰, respectively, and a bowtie2¹¹¹ index was generated. Trimmed reads were mapped to the t/rRNA index with ‘--very-fast-local’ flag, and unmapped reads were retained. Retained reads were mapped to mouse genome assembly mm39 with STAR¹¹² with splice sites defined in the custom transcriptome and the following arguments: --outFilterMultimapNmax 1, --outSAMattributes All, --outFilterMismatchNmax 3, --alignEndsType EndToEnd, --outFilterIntronMotifs RemoveNoncanonicalUnannotated --twopassMode None. Alignments to chrM and nonstandard chromosomes were discarded, and alignments that overlapped Gencode vM31 snoRNAs were filtered out with bedtools¹¹³ intersect -v -s -f 0.33.

Aggregate signal tracks were formed by separately merging elongating alignments (cycloheximide-treated) and initiating alignments (lactimidomycin/harringtonine-treated) across datasets with bamtools merge. Alignments were converted to Bigwig with the deepTools bamCoverage and split by strand. The Ribo-seq signal was not normalized.

Human Ribo-seq data

THP-1 Ribo-seq FASTQ files from GSE208041¹¹⁴ were downloaded and processed as above, with reads mapping to human r/t/snoRNAs discarded. Remaining reads were mapped to the hg38 human reference genome. Alignments were lifted to mm39 coordinates with the liftover utility and mm39.hg38.chain file provided by the UCSC Genome Browser^95,96 (https://hgdownload.soe.ucsc.edu/downloads.html).

ORF calling

Biological replicates within GEO datasets were merged and filtered for alignments in the 27–32 nucleotide length range. RiboTISH⁶¹’s QC module was run on each library and manually examined for 3-nucleotide periodicity for each set of read lengths. CDSs were called using a hierarchical strategy with both RiboTISH and PRICE⁶² at three levels: (1) replicate-level merging within individual experiments, (2) experiment-level merging within each GEO dataset, and (3) global merging across all datasets. This hierarchical approach maximized sensitivity while maintaining specificity by calling CDSs at multiple scales of evidence.

RiboTISH was run with --alt --longest to identify the longest statistically significant CDS with any near-cognate start codon. PRICE was run with default settings, and Benjamini-Hochberg correction was applied to generate adjusted p-values (padj). NCDSs from both tools were retained if they reached padj < 0.01 in at least one hierarchical test and encoded peptides >10 amino acids. ORFs sharing start or stop codons with Gencode protein-coding genes were excluded to focus on truly novel translation events.

The final nCDS set was generated by merging PRICE and RiboTISH calls, consolidating entries that (1) shared a stop codon and (2) where one amino acid sequence was a complete subsequence of the other, retaining the longer variant.

Annotating novel ORFs

Amino acid sequences from the final set of novel nCDSs were annotated for functional domains with SignalP6⁹⁷(signal peptides), DeepTMHMM⁹⁸ (signal peptides and transmembrane domains), MobiDB-lite3¹⁰¹ (intrinsically disordered regions), and InterProScan5/Pfam^99,100 (protein families) with default settings. Localization predictions were determined using DeepLoc2⁶⁴.

Proteomics data

Database construction:

Cell-type-specific proteome databases were generated for macrophages, B cells, and T cells by combining nCDSs with Gencode vM31 proteins from transcripts expressed (TPM > 0) in the following RNA-seq datasets:

• Macrophages: PRJNA482293¹¹⁵, GSE120762⁷, GSE99787⁴⁵

• T cells: GSE94671¹¹⁶

• B cells: GSE94671¹¹⁶

For immunopeptidomics¹¹⁷ analysis of the mouse tissue atlas, a pan-tissue database containing all Gencode proteins plus all nCDSs was used. MaxQuant common contaminant sequences were appended to all databases.

Peptide assignment:

Mass spectrometry data from,

• Macrophages: IPX0001245000^118,119

• Macrophage secretomes: PXD001905¹²⁰,PXD011780¹²¹,PXD017320¹²²,PXD029155^123,124

• B cells: PXD022099^123,125

• T cells: PXD022099^123,125

• Immunopeptidomics: PXD008733^117,123

Were searched against cell-type-matched databases using MS-GF+⁶⁵ with a reverse-decoy database for false discovery estimation. Carbamidomethylation of cysteine was set as a fixed modification and methionine oxidation as a variable modification. For TMT-labeled dataset PXD029155, TMT (+229.16) was included as a static modification on N-termini and lysines.

Rescoring and filtering:

MS2Rescore 2.0⁶⁶ was used to rescore MS-GF+⁶⁵ output with ms2pip¹²⁶ and percolator¹²⁷. Trained models were selected based on fragmentation method: “HCD2019” for HCD datasets, “Immuno-HCD” for immunopeptidomics HCD, “TMT” for TMT-labeled samples, and “CID” for all other acquisition types. PSMs matching Gencode proteins were excluded, and retained PSMs matched at least one nCDSs. NCDS PSMs with q-value < 0.01 and posterior error probability < 0.1 were considered high-confidence hits.

short read RNA-seq

RNA expression for cell-type comparison (Figure 2A) was quantified with Salmon¹²⁸ (--validateMappings --seqBias --gcBias) against Gencode vM31 using datasets GSE60927, GSE210222, GSE88987, GSE120762, PRJNA482293, GSE141745. Transcript-level quantifications were aggregated to gene-level counts with tximport¹²⁹, log-transformed, and batch-corrected using limma's removeBatchEffect function¹³⁰.

RNA-seq visualization tracks

For locus-level visualization (e.g., Figure 2C,D), paired-end RNA-seq reads from GSE94671 (100 bp), GSE195678 (101 bp), GSE201870 (150 bp), GSE210419 (150 bp), and GSE120762 (76 bp) were aligned with STAR¹¹², with multimapping reads discarded (--outFilterMultimapNmax 1 --twopassMode Basic). Alignments were converted to CPM-normalized BigWig tracks with deepTools¹³¹ bamCoverage.

Long read RNA-seq

We previously published Nanopore direct RNA data from wild-type C57BL/6 BMDMs (untreated and 6-hour LPS) that were basecalled and aligned to mouse build mm10⁶⁰. We lifted the alignments to mm39 with the liftover utility and mm10.mm39.chain file provided by the UCSC Genome Browser^95,96 (https://hgdownload.soe.ucsc.edu/downloads.html) for visualization.

sgRNA Library design

In order to generate a gRNA library specific to nCDSs relevant to macrophage biology we used the utilities and workflow from the CRISPRware⁶⁷ software package. Novel ORFs were filtered to include only those on RNAs expressed at >0.1 TPM in BMDM-specific RNA-seq datasets (GSE60930, GSE81250, GSE89184, GSE99787, GSE120762) and called at the individual condition level (i.e., before merging across conditions) in at least one macrophage Ribo-seq dataset, ensuring high-confidence ribosome engagement in macrophages. GTFs were generated for both PRICE and RiboTISH ORFs, and candidate gRNAs with cut sites (−4 bp from PAM) intersecting these ORFs were enumerated. GRNAs were filtered to exclude those with MIT specificity score¹³² <50, Ruleset2 cleavage efficiency¹³³ <30, or which overlap with annotated CDS regions. Up to 10 gRNAs were selected per nCDS, prioritizing guides targeting the 5' half of nCDSs >200 bp, the 5’ 75th percentile of nCDSs 100–200 bp, or any position for nCDSs <100 bp to balance expected phenotypic impact with guide availability for short sequences. When >10 suitable guides were available, those with highest specificity scores were selected. For nCDSs with overlapping translated regions (arising from alternative transcripts or frameshifts), gRNAs were assigned to combined “ORF-units” for analysis.

Controls:

Up to 8 sgRNAs targeting non-nCDS regions of each lncRNA were included to distinguish ORF-specific effects from transcript-level regulation. For each targeted uORF, the cognate downstream CDS was also targeted with 6 sgRNAs. Positive controls we included 100 essential genes from our previous genome-wide screen⁶⁸ and 236 NFκB pathway regulators (138 from our prior NFκB-GFP sorting screen and 98 from GO terms¹³⁴ GO:0007250, GO:0038061, GO:0004704, and GO:0046696). Protein-coding gene gRNAs were designed using CRISPick^133,135. The 1,500 non-targeting sgRNAs with the lowest measured dropout from our prior screen served as negative controls.

Selected gRNA sequences were 20 nucleotides in length, and the 5’-nucleotide was replaced with “G” to enhance transcription. The final library consisted of 39,748 gRNAs.

Lentiviral transduction and selection

The gRNA library was synthesized as an oligonucleotide pool by Twist Bioscience and cloned into a lentiviral vector containing puromycin resistance and BFP reporter cassettes as previously described^68,135. HEK293T cells were transfected with library plasmid and packaging vectors, and the supernatant was harvested at 72 hours. Three distinct clonal lines of NFkB-GFP⁺, Cas9⁺ immortalized bone marrow-derived macrophages were infected with lentivirus at MOI < 0.3 at day −10. After three days (day −7), lentiviral media was replaced with DMEM +10% FBS media and 10 ug/ml puromycin. After seven days of puromycin selection, each clonal line showed >90% BFP positive cells by flow cytometry, and day 0 samples were collected and stored in freezing media (90% FBS, 10% DMSO).

Essentiality screen

For 14 days post-selection, cells were split every ~3 days to maintain confluency between 60–70% while ensuring >2000X gRNA library coverage. On day 14, cells from each clonal line were collected at >1000X and stored in freezing media.

Sorting CRISPR screen

At day 14, cells from each clonal line were treated with 200 ng/mL Pam₃CSK₄. After 24 hours, cells were transferred to FACS buffer to prepare for sorting on a BD FACSAria II as previously described⁶⁸. In brief, GFP was excited using a 488-nm laser and detected using a 525/50-nm filter. For each clonal line, the top 20% and bottom 20% of cells were collected as GFP-high and GFP-low samples, with >100X coverage maintained in each sorted sample.

Screen sequencing and analysis

Genomic DNA was collected from cell pellet and PCR amplified as previously described^68,136. Libraries were sequenced, 23-nucleotide single-end, on an Illumina NextSeq 2000 sequencer operated by the UCSC Paleogenomics Lab.

Sequenced reads were matched to gRNAs with MAGeCK⁶⁹ count with the arguments (--trim-5 1 --sgrna-len 19). In cases where gRNAs targeted overlapping nCDS regions, arising from alternative transcripts or out-of-frame translation, gRNAs were matched to “ORF-units”, e.g. Emsy uORF1:Emsy uORF2. nCDS-loci with less than 4 gRNAs were filtered out and Mageck MLE was used to test for essentiality hits between the day 0 and day 14 samples (--control-sgrna negctrl_ids.txt --norm-method control --permutation-round 10 --remove-outliers --sgrna-efficiency gRNA_scores.tsv), where negctrl_ids.txt is the list of 1,500 negative control gRNAs and gRNA_scores.tsv is the DeepHF¹³⁷ cleavage efficiency prediction for each gRNA.

The gRNA counts from the high and low GFP sorting experiments were normalized with DESeq2¹³⁸ calculated size factors adjustment, and processed in MAUDE⁷⁸ as clonal replicates with output reported by the geneLevelStats() function. This resulted in three dataframes, one for each clonal replicate, with z-scores assigned to each nCDS-loci and canonical CDS. The three dataframes were then filtered to a final combined set which removed entries that were not directionally consistent across all three datasets, reducing the number of tested nCDS-loci from 4,891 to 1,434. In this filtered set, replicate aggregated p-values and z-scores were calculated with Stouffer’s method and new padj values were assigned using the Benjamini-Hochberg correction.

novel ORF conservation

NCDS nucleotide sequences were BLATed¹³⁹ against the human genome hg38 build on the UCSC Genome Browser. In cases where an nCDS was identified on the orthologous human gene, the translated sequences were aligned and visualized in the Jalview application¹⁴¹.

CRISPR KO Experiments

Three gRNAs were chosen to target SYNIR and SEMR ORFs.

Three SYNIR gRNAs: 1- TGGAGTCTTGATCCCGTCTG, 2- GGGTGAACTGAACTGTAATG, 3-TATGCACAGAAGTATGACTC

SEMR gRNAs: 1- GAAACTAAGATACAGTACCA, 2- AACGTGGCAGGTGATATCAG, 3-ACCAGCCACGCAACCTAACG

And three non-targeting gRNAs were used as negative controls. Complementary oligos were ordered from IDT, reconstituted and ligated into pU6 expression vector. The vector also contained a puromycin resistance cassette and mCherry reporter. Lentivirus was prepared as described above and iBMDM samples were each infected with a single gRNA. Infected cells were puromycin-selected for 7 days when the mCherry population reach >90%. Each gRNA-iBMDM line was treated with 200ng/ml LPS for 6 h and samples were collected for RNA processing.

RNA sequencing

Total RNA was isolated from cells using TRIzol reagent (SigmaAldrich, T9424) and Direct-zol RNA Miniprep Plus kit (Zymo Research, R2072), following manufacturer’s instructions. RNA samples were sent to Novogene for 150×150 bp stranded library preparation and were sequenced on a NovaSeq X Plus sequencer.

Differential gene expression

Reads were quantified with Salmon¹²⁸ (--validateMappings --gcBias --seqBias) against Gencode vM31. Transcript-level abundance estimates were summarized to gene level using tximport with length-scaled TPM counts (countsFromAbundance = “lengthScaledTPM”). Differential expression was performed with DESeq2¹³⁸ using default parameters and a two-factor design comparing samples with targeting- and non-targeting-gRNAs. Genes were considered significantly differentially expressed at adjusted p-value < 0.01 and |log₂(fold-change)| > 1. Volcano plots were generated in R using ggplot2.

Tabula muris/senis scRNASeq reprocessing

To expand the scope of genes available for interrogation in these single-cell atlases, we reprocessed 158,420 cells with reads sequenced with the SMART-Seq2¹⁴⁰ protocol - a sensitive, full transcript sequencing approach, not restricted to the 3’ end originally published in the Tabula Muris⁸⁷ and Tabula Senis⁸⁸ projects. For quantifying expression, we used Salmon¹²⁸ with (--validateMappings --gcBias --seqBias) against Gencode annotation vM38. We filtered out cells with fewer than 500 unique genes or less than 50,000 total gene counts, leaving ~110,000 cells with roughly equal distribution across mouse ages 3-, 18-, and 24-month. Cell annotations, including cell ontology ids, tissue of origin, mouse of origin, etc. were matched from the original publications, and the cell annotation approach is described therein. Gene counts were normalized as ln(1 + counts per million). For clustering, 3000 high-variance genes were selected with the scran¹⁴² package and used to generate 50 principal components at separately at distinct mouse ages and globally, yielding four Rdata object, 3-,18-,24-months, and all ages combined. For each age and for the global set UMAPs were generated with scater¹⁴³ with n_neighbors 30 and min_dist 0.5. Full pipeline description and scripts are available in the GitHub repo (Data Availability).

Envelope protein structure homology search

The Alphafold 3⁹⁰ web browser interface was used to predict structures for SEMR B230354K17Rik ncORF8, and D430020J02Rik ncORF1 proteins. For SEMR and D430020J02Rik ncORF1, the predicted signal peptide was not included in the prediction, as this would be cleaved to generate a functional protein. For B230354K17Rik ncORF8, one subdomain was folded with high confidence, while the rest of the protein was predicted at low confidence. The high-confidence structure predictions were searched against the AFDB50 database with FoldSeek⁹¹.