A microprotein atlas of the human frontal cortex in Alzheimer’s disease

Brendan Miller; Eduardo Vieira de Souza; Calvin Lau; Joan M Vaughan; Victor J Pai; Servando Giraldez; Andréa Rocha; Jolene K Diedrich; Clodagh C O’Shea; David A Bennett; Alan Saghatelian

doi:10.1038/s43587-026-01207-x

. 2026 Sep 14;6(9):1979–1997. doi: 10.1038/s43587-026-01207-x

A microprotein atlas of the human frontal cortex in Alzheimer’s disease

Brendan Miller ^1,^✉, Eduardo Vieira de Souza ¹, Calvin Lau ¹, Joan M Vaughan ¹, Victor J Pai ¹, Servando Giraldez ², Andréa Rocha ¹, Jolene K Diedrich ³, Clodagh C O’Shea ², David A Bennett ⁴, Alan Saghatelian ^1,^✉

PMCID: PMC13577902 PMID: 42736449

Abstract

Understanding the molecular basis of neurodegeneration requires a comprehensive map of the genome’s protein-coding output. Although thousands of small open reading frames (ORFs) are translated in the human brain, proteomic evidence for their encoded microproteins (MPs) (≤150 amino acids (aa)) remains limited. Here, we present a brain MP atlas that integrates transcriptomics, mass spectrometry and deep-learning-predicted spectra across more than 600 postmortem frontal cortex samples with and without Alzheimer’s disease (AD). We identified 1,067 MPs absent from reviewed UniProtKB entries with high-confidence spectral support. A subset is differentially expressed in AD independently of the annotated main ORF at the same locus. A small ORF expressed by MKKS encodes a 63-amino-acid MP that is the locus’s predominant translation product and is downregulated in AD; its loss impairs microglial mitochondrial respiration, implicating it in microglial bioenergetics. This atlas expands the annotated brain proteome and provides a resource for studying MPs in aging and neurodegeneration.

Subject terms: Ageing, Alzheimer's disease, Gene expression

A new atlas identifies 1,067 unannotated microproteins in the human frontal cortex with strong spectral grades. A subset is differentially expressed in Alzheimer’s disease independently of their canonical genes, including an MKKS-derived microprotein required for normal microglial mitochondrial respiration.

Main

The molecular mechanisms of healthy cognitive aging remain incompletely understood. Multi-omics studies of the postmortem human brain have provided key insights but depend on reference genome databases that do not represent small open reading frames (smORFs). These smORFs are genomic elements that, when transcribed and translated, can encode short polypeptides, often termed microproteins (MPs) (typically fewer than 100–150 aa)^1–3. Because of their historical exclusion, the role of MPs and the genes that encode them remain largely unexplored in omics-based neurobiology research.

Approximately 10% of the entries in UniProtKB’s reviewed section (Swiss-Prot) are proteins shorter than 150 aa (ref. ⁴), but this fraction may be an underestimate⁵. Recently, the TransCODE consortium documented over 7,000 actively translated smORFs across human cell types, nearly a quarter with mass spectrometry (MS) evidence, termed ‘peptideins’ to denote expression evidence without an established biological function⁶. Should these peptideins be experimentally validated as functional proteins, they would nearly double the number of annotated entries shorter than 150 aa in the reviewed UniProt/Swiss-Prot database.

Collaborative efforts, such as the TransCODE consortium, aim to enhance the annotation of MPs by overcoming historical detection challenges. MPs have been difficult to detect because their short length yields few proteolytic peptides, limiting their identification using MS. In addition, since a large fraction of smORFs is embedded within or adjacent to annotated protein-coding ORFs, conventional RNA sequencing workflows cannot infer expression of smORFs alone. Recent advances in ribosome profiling, proteogenomics and MS-based technologies have begun to address these challenges.

Ribosome sequencing (Ribo-seq) enables the identification of thousands of translating smORFs by sequencing ribosome-protected fragments^7,8. MS-based proteomics provides evidence that smORF translation indeed yields an MP^9–11. Once MPs are detected, machine learning and computational approaches are often used to infer their biological function^12–14. Characterized MP functions range from mitochondrial membrane components to secreted signaling peptides and immunopeptidome modulators¹⁵, positioning them within aging and neurodegeneration pathways. Despite their emerging functional relevance, a comprehensive brain MP atlas has not been built. Such a resource would provide an experimental entry point for investigating smORFs and their MPs in fundamental neurobiology and neurodegeneration.

To address this gap, we built a transcript-to-MP atlas integrating native transcriptomics with deep proteomics across more than 600 postmortem frontal cortex tissues with and without Alzheimer’s disease (AD). This atlas provides protein-level expression evidence in the form of tandem MS (MS/MS) spectra graded against deep-learning-predicted references. We report over 1,000 MPs absent from the reviewed Swiss-Prot catalog whose spectra closely match predicted spectra. One of these is an MP encoded by the MKKS locus that is significantly downregulated in AD, disrupts mitochondrial oxidative phosphorylation upon knockout (KO) and diverges in expression from the canonical MKKS ORF-encoded protein, a decoupling also seen in other MPs, replicated across cohorts and linked to RNA-binding protein (RBP) motifs near splice junctions. Data including MS/PROSIT plots, grades, expression profiles, sequences and formatted files are accessible via a browsable application, which can be used to filter sequences based on user-specific interests (https://huggingface.co/spaces/brmiller/brain-microprotein-atlas-app).

Results

Unique MS evidence of MPs in aged human brains

To identify MPs with unique MS evidence, we generated a custom dorsolateral prefrontal cortex (DLPFC)-derived protein search database from 12 native transcriptomes¹⁶. This database consists of all putative smORFs and their resulting MP sequences appended to the UniProt reference proteome (UP000005640). UniProt organizes entries into two curation tiers: Swiss-Prot (manually curated, experimentally supported) and TrEMBL (computationally annotated, awaiting review)⁴. We designated Swiss-Prot-matched sequences (≤150 aa) as reviewed and the combined set of TrEMBL-matched sequences plus transcriptome-reconstructed smORFs not annotated by Ensembl as unreviewed.

We then used this search database to identify MPs in tandem mass tag (TMT)-MS data from 610 DLPFC samples from the Religious Orders Study (ROS) and the Memory and Aging Project (MAP) across the AD neuropathological spectrum. ROSMAP TMT-MS data are derived from 400 samples across 50 batches and 24 fractions (1,200 RAW files) and 210 samples across 14 batches and 48 fractions (672 raw files)¹⁷. Our analysis yielded 4,321 MPs passing protein identification with a false discovery rate (FDR) < 0.01, including 3,217 unreviewed and 1,104 reviewed (Fig. 1a–d and Supplementary Tables 1–3), each with at least one unique peptide-spectrum match (PSM).

The 3,217 unreviewed MPs map broadly across all autosomes, mostly encoded near or within genes producing a larger canonical protein (Fig. 1b), with a median aa length nearly half that of the reviewed MPs (median 58 aa versus 117; Fig. 1c). Their encoding smORFs are located upstream and downstream, within the coding sequence (CDS) of annotated ORFs, or are derived from transcripts annotated as noncoding. Downstream ORFs (dORFs) were the largest class (19.9% of all 4,321 MPs), followed by short isoforms (14.3%) and TrEMBL/altORFs (14.1%), with internal ORFs (iORFs) (9.2%), long noncoding RNA (lncRNA)-hosted smORFs (6.4%), upstream ORFs (uORFs) (5.7%) and pseudogene ORFs (psORFs) (5.0%) comprising the remainder (Fig. 1d).

MPs are historically challenging to detect using MS because of limited peptide recovery from short length and relatively low abundance⁹. Indeed, reviewed MPs generated roughly an order of magnitude more spectral counts per sequence than unreviewed MPs (Fig. 1e); 11% of unreviewed MPs exceeded mean abundance by more than one standard deviation, versus 48% of reviewed MPs (Fig. 1f, and Supplementary Table 4). Unreviewed MPs also showed substantially weaker protein–mRNA concordance than reviewed MPs (ρ = 0.06 versus ρ = 0.36; Fig. 1g).

Because MPs may fall below MS detection thresholds, we integrated Ribo-seq from 43 non-AD DLPFC samples with ShortStop, a machine learning framework that classifies translated smORFs resembling annotated UniProt/Swiss-Prot MPs^5,12 (Fig. 1h). This identified 665 of 1,769 (37.6%) reviewed and 1,592 of 4,809 (33.1%) translated MPs lacking MS evidence in ROSMAP TMT-MS (Fig. 1i). Among MPs with unique ribosome coverage evidence, MS-detected reviewed entries showed higher ribosome occupancy than Ribo-seq-only reviewed entries (median log₁₀ RPKM = 3.70 versus 3.38; P = 8.58 × 10⁻¹²) and unreviewed MPs (2.72; P = 6.34 × 10⁻⁴⁵), indicating that MS identification is associated with higher unique Ribo-seq coverage for reviewed sequences with a functional annotation (Fig. 1j). Data-independent acquisition (DIA) MS across four samples recovered only modest additional sequences (35 reviewed, 30 unreviewed) (Supplementary Table 1), suggesting that limited recovery reflects abundance, not a technical limitation.

Integrating Ribo-seq, MS and PROSIT provides an evidence framework for confidently identifying brain-expressed MPs

Large-scale proteomics studies are inherently susceptible to false positives¹⁸. Therefore, we assessed the spectral quality of the 3,217 unreviewed MPs that passed protein identification (FDR < 0.01) by using PROSIT, a deep learning framework predicting theoretical fragmentation spectra for comparison against observed spectra¹⁹. Each MP was assigned to one of four confidence tiers (strong, moderate, weak, insufficient) based on its combination of PROSIT-compared spectral angle (SA) score (0–1 scale, normalized) and fragment-ion coverage, with thresholds derived from previously reported MP grading criteria (Fig. 2a,b)²⁰. This scoring system was cross-validated against two independent manual reviewers (Extended Data Fig. 1).

Fig. 2 — a, Peptide grading criteria. Criterion 1: SA agreement between observed and PROSIT-predicted spectra. Criterion 2: fragment-ion coverage (percentage of b/y-ions observed). b, Peptide grading matrix (n = 3,594 peptides uniquely mapped to unreviewed MPs). Each cell reports the count of peptides in each combination of normalized SA bin (≥0.778/0.500–0.778/0.222–0.500/≤0.222) and fragment-ion coverage bin (sparse <25%/partial 25–50%/majority >50%). Cell labels indicate the assigned grade (strong/moderate/weak/insufficient). c, Per-protein SA distributions (median SA across the peptides of each unreviewed MP; n = 3,001), stratified according to confidence grade (insufficient n = 158, weak n = 343, moderate n = 1,433, strong n = 1,067). Descriptive distribution only; no statistical test was performed. d, MP peptide grade according to smORF type (stacked bar chart; n = 3,581 proteins total). Each protein is assigned its highest confidence uniquely mapped tryptic peptide grade. Strong/moderate/weak/insufficient tiers are shown for reviewed MPs (n = 580), dORFs (n = 816), TrEMBL/altORFs (n = 568), short isoforms (n = 544), iORFs (n = 379), lncRNA-ORFs (n = 262), uORFs (n = 231) and psORFs (n = 201). e, Theoretical tryptic unique peptide capacity for MPs amenable to PROSIT examination. The bar plots show the fraction of reviewed (n = 580) and unreviewed (n = 3,001) MPs expected to yield 0, 1, 2, 3 or ≥ 4 tryptic peptides based on the sequence. Descriptive distribution only. f, Observed unique tryptic peptide support according to confidence tier, split into 1-peptide and ≥2 tryptic peptide categories for unreviewed (n = 3,001) and reviewed (n = 580) MPs. g,h, Representative PROSIT mirror plots of an iORF from the *PTOV1* locus graded moderate (g) and a uoORF from the *NYAP1* locus (h) graded strong. The parent gene schematic and Ribo-seq track are shown above each spectrum (y axis, Ribo-seq CPM). Representative of a single identified spectrum.

Source data

Extended Data Fig. 1 — Balloon plots comparing manually assigned tiers against the PROSIT spectral-angle (SA)-derived tier for Reviewer 1 (a) and Reviewer 2 (b) with inter-rater agreement (c). Unit of study = individual spectrum; n = 50 spectra per panel. Bubble area is proportional to the number of spectra at each coordinate pair (maximum n = 50); bubble shade reflects the grading tier. Statistics: two-sided Spearman rank correlation. (a) Reviewer 1 versus SA tier: ρ = 0.584, P = 8.32 × 10⁻⁶. (b) Reviewer 2 versus SA tier: ρ = 0.647, P = 3.79 × 10⁻⁷. (c) inter-rater agreement (Reviewer 1 versus Reviewer 2): ρ = 0.955, P = 7.01 × 10⁻²⁷.

Source data

Of 3,001 unreviewed MPs amenable to PROSIT interrogation, 1,067 (35.6%) were strong, 1,433 (47.8%) moderate, 343 (11.4%) weak and 158 (5.3%) insufficient. Over 90% of MPs graded as strong had more than 50% fragment-ion coverage and a high normalized SA score over 0.78 (Fig. 2c). Similarly, over 95% of MPs graded as moderate had fragment-ion coverage over 50% but lacked the high SA concordance; nevertheless, these may reflect genuine peptides, warrant follow-up and should not be discarded. Overall, across all grades, MP detection was consistent for all smORF classes (Fig. 2d).

Conventional proteomics typically requires two unique tryptic peptides per protein for confident identification²¹, a threshold most MPs cannot meet given their short length (Fig. 2e). Conservation-based approaches such as PhyloCSF offer an alternative but are also poorly suited to MPs because many that have been documented are evolutionarily young²²; accordingly, nearly three-quarters of MS-evidenced MPs in our dataset showed negative PhyloCSF scores (Extended Data Fig. 2). Consequently, relying on peptide count or conservation alone would exclude many bona fide MPs.

Extended Data Fig. 2 — n = 1,530 Reviewed and n = 4,052 Unreviewed microproteins, spanning both MS-detected and Ribo-seq/ShortStop-detected, with a computed mean PhyloCSF score (biologically independent sequences). Violin plots show the score distribution with embedded box plots (center line, median; box bounds 25th and 75th percentiles; whiskers extend to the most extreme data points within 1.5× the interquartile range). Statistics: two-sided Wilcoxon rank-sum test; W = 4,864,549, P < 2.2 × 10⁻¹⁶ (below the floating-point floor reported by R, asterisks denote ****). Reviewed microproteins show significantly higher PhyloCSF scores (median 3.90 vs −5.49), consistent with stronger evolutionary conservation.

Source data

Our MS-based gradient system and ribosome profiling integration workflow helps recover MPs that peptide-count or conservation filters alone would exclude. Among 1,067 strong-grade unreviewed MPs, 814 had just one observed tryptic PSM (Fig. 2f), but these can be further assessed using ribosome profiling (Fig. 2g–h and Supplementary Tables 3–5). For example, an MP encoded by an upstream overlapping ORF (uoORF) within NYAP1 illustrates this: despite a negative PhyloCSF score (−3.4) and only one detectable tryptic peptide, it showed clear ribosome occupancy and a strong PROSIT grade (Fig. 2h).

Differential MP abundance in AD using TMT-MS quantification

TMT labeling enables multiplexed peptide quantification. We performed differential abundance analyses on TMT-MS data from 480 ROSMAP DLPFC samples, grouped into three diagnostic categories per prior criteria: non-AD, asym-AD (pathology without symptoms) and sym-AD (pathology with symptoms), and then carried out a differential abundance analysis between non-AD and sym-AD donors. Given the MP detection challenges, we used a stringent model requiring detection in at least half of samples and a relaxed model using all quantified proteins, each reporting high-confidence (FDR < 0.05) and exploratory (0.05 ≤ FDR < 0.2) tiers (Fig. 3a,b, Extended Data Fig. 3 and Supplementary Table 2).

Fig. 3 — Symbols denote Benjamini–Hochberg FDR: + < 0.2; * < 0.05; ** < 0.01; *** < 0.001. a,b, Differentially abundant MPs, sym-AD versus non-AD, according to smORF type, including reviewed MPs. The bars split into high-confidence (FDR < 0.05, light orange and light teal) and exploratory (FDR = 0.05–0.2, dark orange and dark teal) tiers. Groups: non-AD (n = 111) versus sym-AD (n = 187). Two-sided Student’s t-test, Benjamini–Hochberg FDR, corrected within each stringency model. a, Stringent: detected in more than 50% of donors (n = 580 MPs). b, Relaxed: detected in one or more donors per condition (n = 3,179 MPs). c–f, Normalized abundance, non-AD versus sym-AD, for two reviewed and two unreviewed sequences: SST (c), MDK (reviewed) (d), BCL3-iORF (e) and RAB3C N-terminal isoform (unreviewed) (f). Donor n (non-AD/sym-AD): SST 111/187; MDK 109/178; BCL3-iORF 69/100; RAB3C 87/162. Violin plots with embedded box plots (center line, median; box, 25th–75th percentiles; whiskers, most extreme values within 1.5× the interquartile range (IQR)). Two-sided Student’s t-test, Benjamini–Hochberg FDR, stringent model (≥ 50% detection); values are t-statistic, log₂(fold change) (95% CI), Cohen’s d, exact P and FDR. c, SST: t₍₂₉₆₎ = −6.01, log₂(fold change) = −0.236 (−0.314 to −0.159), d = −0.720, P = 5.59 × 10⁻⁹, FDR = 3.08 × 10⁻⁷ (***). d, MDK: t₍₂₈₅₎ = 14.07, log₂(fold change) = 1.342 (1.155 to 1.530), d = 1.712, P = 1.63 × 10⁻³⁴, FDR = 6.05 × 10⁻³¹ (***). e, BCL3-iORF: t₍₁₆₇₎ = −2.40, log₂(fold change) = −0.103 (−0.188 to −0.018), d = −0.375, P = 0.0177, FDR = 0.0595 (+). f, RAB3C: t₍₂₄₇₎ = 3.01, log₂(fold change) = 0.097 (0.034 to 0.161), d = 0.401, P = 0.00284, FDR = 0.0148 (*). g, Heatmap: smORF-containing genes mapped to snRNA-seq cell types using summary statistics (Mathys et al.²⁴), showing log(CPM) and log(fold change) for SST, MDK, BCL3 and RAB3C across neuronal and glial subtypes. On the log(fold change) row, filled circles mark significant global AD pathology association (summary statistics adjusted P < 0.05); open circles mark trend-level association (0.05–0.1); unmarked cells reached neither threshold. log(CPM) carries no significance marking. h,i, PROSIT mirror plots, both strong quality, representative of one spectrum per peptide: BCL3-iORF (peptide AAAPQAPAAR) (h) and RAB3C N-terminal isoform (peptide MASAQDAR) (i). j, Confocal microscopy of HMC3 microglia overexpressing C-terminal ALFA-tagged BCL3-iORF and RAB3C N-terminal isoform. Representative of three independent transfections. 4′,6-diamidino-2-phenylindole (DAPI) (blue) and ALFA (orange); BCL3-iORF localizes to the nucleus.

Source data

Extended Data Fig. 3 — Multidimensional scaling (MDS) illustrates correction steps for 210 samples (a, ROSMAP TMT-MS Round 2) and 400 samples (b, ROSMAP-TMT Round 1): starting from raw log2 abundance, then normalized to the global internal standard (GIS), followed by Tunable Approach for Median Polish of Ratio (TAMPOR) batch correction. (c) Additional correction across rounds (green = Round 1; blue = Round 2 samples). (d) Cumulative microprotein quantification rate (x axis) versus cumulative fraction of proteins (y axis) with quantitative relative abundance prior to TAMPOR.

Source data

In the stringent model, 99 reviewed MPs (≤ 150 amino acids) were differentially abundant at FDR < 0.05. For example, decreased SST and elevated MDK recapitulated prior ROSMAP findings²³, validating the framework for interpreting unreviewed MPs (Extended Data Fig. 4). This same framework identified 22 unreviewed MPs at FDR < 0.05, rising to 40 at the exploratory tier (Fig. 3a). The relaxed model recovered 258 unreviewed MPs at FDR < 0.2, spanning all smORF classes (Fig. 3b). Effect sizes reflect directional associations rather than large shifts, as median absolute log₂ (fold change) among significant stringent-model MPs was 0.0969, consistent with cellular heterogeneity in bulk postmortem tissue²³.

Extended Data Fig. 4 — Each point represents a Swiss-Prot-annotated protein quantified across cohorts. The x-axis shows the log₂ fold change (AD vs. Control), and the y-axis indicates –log10 q from multiple hypothesis correction. Selected proteins with strong differential expression and statistical significance are labeled, including SMOC1, MDK, NTN1, and APP, replicating previous findings from Johnson et al. The horizontal line marks the FDR q-value = 0.2 threshold. Statistics: two-sided Student’s t-test, Benjamini-Hochberg-adjusted FDR (q-value), same stringent model (≥50% detection) as Fig. 3a.

Source data

Proteins that are differentially abundant as revealed through TMT-MS can be further interpreted in their cellular context (Fig. 3c–f). We mapped each MP to its encoding genomic locus and queried published single-nucleus RNA sequencing (snRNA-seq) summary statistics from the same ROSMAP cohort²⁴ (Supplementary Table 6). Among Swiss-Prot reviewed proteins shorter than 150 aa, SST, predominantly expressed in inhibitory neurons, was negatively associated with AD pathology across all major cell types including glia (Fig. 3g). MDK showed widespread neuronal and glial expression but was negatively associated with AD only in oligodendrocytes (oligos) despite elevated protein AD brains.

Among unreviewed MPs, we highlight two microglia-enriched examples with altered abundance in AD. BCL3-iORF (Fig. 3e,h) was lower in sym-AD tissue at the exploratory tier of the relaxed differential expression model (FDR < 0.2), whereas the RAB3C N-terminal isoform was significantly higher in the stringent model (FDR < 0.05). BCL3-iORF is most highly expressed in microglia, whereas the RAB3C N-terminal isoform shows microglia-specific expression that is negatively associated with global AD pathology. For both MPs, transient overexpression in HMC3 human microglial cells confirmed stable expression. BCL3-iORF localized to the nucleus, whereas RAB3C showed no nuclear overlap (Fig. 3j).

A subset of smORFs diverges in expression from the main ORF of the encoding gene

Nearly three-quarters of MS-detected MPs are encoded by loci already annotated to produce a larger canonical protein. Therefore, we asked whether smORF-containing transcripts are regulated independently of those containing the gene’s main canonical ORF (mORF). After quantifying reads mapping specifically to the smORF coding sequence (CDS) versus the mORF CDS, we generated coexpression profiles pairing each gene’s smORF CDS with the same gene’s mORF CDS.

We observed a bimodal expression architecture. Half of all smORF-mORF pairs showed near-perfect coexpression (mean |r| = 0.97), consistent with shared transcripts, while the other half were more weakly correlated (mean |r| = 0.52), indicating partial independent regulation for many smORF-harboring transcripts (Fig. 4a and Supplementary Table 7).

Fig. 4 — Unless noted, comparisons use ROSMAP (n = 90 non-AD/142 sym-AD) and MSBB (n = 57 non-AD/179 sym-AD); distinct brain donors; unit of study = individual smORF CDS. The asterisks denote Benjamini–Hochberg FDR: *P < 0.05; **P < 0.01; ***P < 0.001. NS, not significant. a, Per-pair Pearson correlation of smORF versus mORF transcript expression across ROSMAP donors (n = 4,326 smORF-mORF pairs; each faint orange line = one pair’s linear fit). Pairs are split at the median correlation (|r| = 0.868, 50th percentile) into strongly correlated (left; n = 2,163, mean |r| = 0.97) and weakly correlated (right; n = 2,163, mean |r| = 0.52) halves. The black dashed line represents a descriptive pooled least-squares fit per subset. b, Dysregulated smORF types (RNA-seq FDR < 0.05), according to class, colored according to additive (orange) versus independent (red) mORF expression. Two-sided LRT (additive model), ROSMAP; n = 390 smORF-mORF gene pairs; 152 decoupled (LRT P < 0.05, uncorrected). c–i, Volcano plots (log₂(fold change) versus −log₁₀(FDR), sym-AD versus non-AD), according to smORF class: dORF (c), short isoform (d), TrEMBL/altORF (e), lncRNA (f), iORF (g), uORF (h) and psORF (i). Two-sided DESeq2 Wald test, Benjamini–Hochberg FDR; ROSMAP. j–l, Three dORF cases—EMC10 (j), RNF41(k) and GNAQ (l). Variance stabilizing transformed (VST)-normalized smORF expression, non-AD versus sym-AD (ROSMAP). Box plots: center indicates the median; boundaries indicate the 25th–75th percentiles; and whiskers indicate the most extreme points within 1.5× the IQR. Two-sided DESeq2 Wald test, Benjamini–Hochberg FDR; log₂(fold change): EMC10 = 0.336, FDR = 4.04 × 10⁻² (*); RNF41 = 0.392, FDR = 1.66 × 10⁻² (*); GNAQ = 0.188, FDR = 2.63 × 10⁻² (*). Parent mORFs were not significant for all three (FDR: EMC10 = 0.41, RNF41 = 0.36, GNAQ = 0.62). Independence from the parent mORF was significant through a two-sided additive LRT (EMC10 P = 1.16 × 10⁻²; RNF41 P = 4.22 × 10⁻³; GNAQ P = 6.27 × 10⁻³). m, smORF-mORF coexpression, ROSMAP versus MSBB (parahippocampal gyrus); 4,323 shared pairs. Cross-cohort r = 0.813. n, NCOA1 smORF versus mORF, ROSMAP and MSBB. Pearson’s r according to diagnosis: ROSMAP non-AD = 0.276, sym-AD = 0.457; MSBB non-AD = 0.407, sym-AD = 0.277. The global linear-model fit is shown as an orange line. o, NCOA1 smORF and mORF expression (VST-normalized), non-AD versus sym-AD, ROSMAP and MSBB. Box plots as in j–l. Two-sided DESeq2 Wald test, Benjamini–Hochberg FDR: ROSMAP log₂(fold change) = 0.249, FDR = 2.01 × 10⁻³ (**); MSBB log₂(fold change) = 0.225, FDR = 7.94 × 10⁻³ (**). p, Representative PROSIT mirror plot and Ribo-seq track of the NCOA1-dORF.

Source data

Within this decorrelated fraction, differential smORF expression in AD may reflect isoform-specific regulation. Applying a likelihood ratio test (LRT) across smORF-containing transcripts differentially expressed in sym-AD versus non-AD, we found that 39% of testable pairs showed disease-associated changes independent of the mORF (RNA-seq FDR < 0.05; Fig. 4b–i). dORFs were disproportionately represented (48%, the highest rate among smORF classes) and enriched for moderate-to-strong MS/MS spectral grades, supporting protein-level relevance. Three dORF loci—EMC10, RNF41 and GNAQ—illustrate this, where the mORF remained unchanged in sym-AD while the dORF showed significant fold change, each supported by moderate-to-strong PROSIT grades (Fig. 4j–l).

We next tested whether this decorrelation extended to another RNA extraction, sequencing workflow and brain region by re-analyzing parahippocampal gyrus expression data from the Mount Sinai Brain Bank (MSBB)²⁵. The ROSMAP bimodal architecture of smORF-mORF RNA expression was concordant in MSBB (r = 0.81; Fig. 4m). For example, a dORF encoded within NCOA1 is supported by high-confidence spectral evidence, partially decoupled from the mORF in both cohorts (ROSMAP pooled r = 0.38; MSBB pooled r = 0.32) and significantly upregulated in sym-AD while the mORF remained stable (Fig. 4n–p).

Posttranscriptional splicing motif enrichment for genes encoding MPs

Expression divergence for genes that express both smORF-containing and mORF-containing isoforms raises a mechanistic question. Do these variants arise from distinct transcription start sites (TSS) or a shared primary transcript via alternative splicing? Aberrant splicing is a hallmark of neurodegeneration and may drive selective isoform dysregulation in AD^26,27. Even if an MP lacks a direct biological function, its detection reveals previously unrecognized regulation of its encoding gene.

Alternative transcription initiation does not appear to be the major explanation for genes whose transcripts containing smORFs and mORFs decorrelate, as approximately 75% of smORF promoters shared motif instances with their paired mORF, per HOMER analysis across 3,631 matched gene pairs (Extended Data Fig. 5 and Supplementary Table 8). While a smaller subset with unique promoters may warrant follow-up, independent initiation for transcripts carrying smORFs did not appear to be the predominant explanation for unique transcript expression.

Extended Data Fig. 5 — Analysis of 3,631 matched genes containing both a MS-detectable smORF-encoded microprotein and its mORF (that is, parent locus) (HOMER v4.11, hg38; TSS window −1000 to +100 bp). Unit of study = individual promoter (a) or smORF–parent pair (b). (a) Proportion of promoters overlapping CpG islands (UCSC cpgIslandExt): smORF promoters 2,777/4,344 (63.9%) vs. mORF promoters 18,785/62,559 (30.0%). Statistics: two-sided Fisher’s exact test, OR = 4.13, P < 1 × 10⁻³⁰⁰. (b) Pie chart of transcription-factor motif sharing across the 3,631 matched pairs: 71.2% shared, 13.6% unique to smORF, 15.2% unique to parent (mean 118 motifs per smORF promoter, 120 per parent; 1,897 smORFs [52%] have ≥1 unique motif). Values are descriptive proportions; no statistical test applied.

Source data

Posttranscriptional splicing emerged as an explanation for smORF-mORF expression decorrelation. RBP motif enrichment at splice junctions in smORF-harboring transcripts differed from transcripts lacking MP evidence by smORF class (Fig. 5a,b and Supplementary Table 9). For example, dORFs showed a distinct 3′ splice site/polyadenylation signature (for example, U2AF2), implicating different posttranscriptional programs depending on smORF location.

At the MKKS locus, the smORF-mORF correlation (r = 0.15; Fig. 5c,d) falls below the 10th percentile of the decorrelation distribution. Of seven annotated MKKS transcript isoforms (Ensembl), three lack the 570-aa mORF and exclusively encode the 63-aa MP²⁸. At the shared second intron splice junction, dementia-associated RBP motifs were identified (Open Targets Platform GraphQL API; Fig. 5c), offering a candidate mechanism that could shift the balance toward full-length over smORF-containing isoforms in AD. The MKKS mORF-carrying transcripts are unchanged in sym-AD (FDR = 0.60), while smORF-containing transcripts are significantly downregulated (log₂(fold change) = −0.20, FDR = 8.54 × 10⁻³; Fig. 5e,f).

At the MKKS locus, an MP is the predominant translation product, not the mORF, and coexpresses with a mitochondrial gene network

Because MKKS was among the most decoupled smORF-mORF pairs in the atlas, we examined its translational output closely. We found that the predominant detectable protein product in human DLPFC is not the 570-aa Swiss-Prot entry (Q9NPJ1), which has been linked to the rare disease McKusick–Kaufman syndrome^4,29; instead, the predominant detectable protein is a 63-aa smORF-encoded MP cataloged in unreviewed TrEMBL (Q9HB66), which we designated micro-MKKS63.

Ribosome profiling in human DLPFC revealed localized coverage at the micro-MKKS63 smORF but sparse signal across the main MKKS ORF (Fig. 6a). No peptides from the canonical 570-aa MKKS sequence were detected across the ROSMAP TMT-MS samples, while micro-MKKS63 yielded over 1,000 PSMs, including peptides with strong-grade and moderate-grade coverage (Fig. 6b). A complementary search of the PepCentric meta-database (66,700 public MS runs)³⁰ confirmed this imbalance, with only 14 PSMs assignable to canonical MKKS versus thousands for micro-MKKS63 (Extended Data Fig. 6). TMT quantitation showed significantly decreased micro-MKKS63 abundance in sym-AD versus non-AD (Fig. 6c).

Extended Data Fig. 6 — MS and ribosome profiling analysis of *MKKS* using PepCentric, a proteogenomic search engine applied to 2.3 billion spectra from 66,700 public MS runs. Canonical MKKS (Q9NPJ1, teal) yields 14 unique PSMs, accounting for <0.1% of all PSMs from the *MKKS* locus when considering micro-MKKS63 (orange). Micro-MKKS63 produces thousands of PSMs, representing the predominant *MKKS*-derived protein across datasets.

Source data

To gain functional insight, we built coexpression networks anchored to ROSMAP RNA-seq reads mapping exclusively to the micro-MKKS63 smORF CDS. This network was strongly enriched for mitochondrial biology across all three Gene Ontology (GO) domains (Fig. 6d–f). However, the network anchored to the canonical Q9NPJ1 mORF CDS showed no significant GO enrichment in either condition (Fig. 6g and Supplementary Tables 10–13). As such, micro-MKKS63 expression—not the main MKKS ORF—correlates with processes strongly related to mitochondrial biology.

Micro-MKKS63 localizes to the mitochondria and CRISPR–Cas9-mediated KO of its smORF lowers microglial oxidative phosphorylation

The mitochondrial coexpression signature of micro-MKKS63 motivated further functional experiments. Given established roles for mitochondria and microglia in AD^31,32, we tested for endogenous micro-MKKS63 expression in HMC3 human microglial cells. MS on HMC3 protein extracts identified robust micro-MKKS63 detection (PXD071500), with no peptides recovered from the canonical 570-aa MKKS protein, mirroring the pattern seen across ROSMAP brain tissue.

We overexpressed a C-terminal FLAG-tagged micro-MKKS63 construct in HMC3 cells and co-immunostained for the mitochondrial marker IMMT, observing signal concentrated at peripheral mitochondrial regions (Fig. 7a–e). To validate this at the endogenous level, we re-analyzed a published BioID proximity-labeling dataset targeting the cytosolic face of the outer mitochondrial membrane (PXD060912)³³, which confirmed that endogenous micro-MKKS63 ranked among the most highly enriched proteins in that fraction (Fig. 7f and Supplementary Table 14).

Fig. 7 — a–d, Immunofluorescence (IF) of HMC3 microglia overexpressing C-terminal FLAG-tagged micro-MKKS63: DAPI (blue) (a); IMMT (red) (b); micro-MKKS63 (green) (c); merge, partial colocalization (d). n = 3 independent transfections; representative images are shown. e, Enlarged boxed region from d. Manders coefficients: M1, micro-MKKS63 overlapping IMMT; M2, IMMT overlapping micro-MKKS63. f, Re-analysis of fractionated mitochondrial proteomics (PXD060912), comparing cells expressing mitochondrially targeted full-length BirA (mito-FL) versus untransduced controls lacking the biotin ligase (n = 3 per group); Benjamini–Hochberg-adjusted two-sided t-test. Micro-MKKS63 was enriched in the mitochondrial fraction (log₂(fold change) = 2.36, 95% CI 1.61–3.12; FDR = 1.41 × 10⁻³). g–k, Seahorse XF Mito Stress Test, oxygen consumption rate (OCR) in a polyclonal nontargeting control (NTC) cell line versus two independent micro-MKKS63 KO lines, sgRNA1 and sgRNA2. n = 23 (NTC), 24 (micro-MKKS63-sgRNA1), 24 (micro-MKKS63-sgRNA2). g, OCR time course: the line represents the mean; the shaded bands represent the mean ± s.e.m. h–k, Basal OCR (h), maximal respiration (i), ATP-linked OCR (j) and reserve capacity (k): box plots (center line represents the median; the boundaries represent the 25th–75th percentiles; and the whiskers represent the most extreme points within 1.5× the IQR). Both lines independently reduced OCR relative to NTC. Two-sided one-way analysis of variance (d.f. = 2, 68), Dunnett’s post hoc test versus NTC; mean difference (95% CI) P. h, F = 8.56, P = 5 × 10⁻⁴; sgRNA1 = 41.68 (14.94–68.41), P = 0.0015 (**); sgRNA2 = 43.44 (16.71–70.18), P = 0.0009 (***). i, F = 8.96, P = 4 × 10⁻⁴; sgRNA1 = 77.73 (36.17–119.3), P = 0.0001 (***); sgRNA2 = 44.59 (3.03–86.14), P = 0.0337 (*). j, F = 14.12, P < 0.0001; sgRNA1 = 20.70 (7.40–33.99), P = 0.0015 (**); sgRNA2 = 28.89 (16.39–41.39), P < 0.0001 (****). k, F = 8.84, P = 4 × 10⁻⁴; sgRNA1 = 57.69 (26.49–88.89), P = 0.0002 (***); sgRNA2 = 23.44 (−7.76 to 54.65), NS (P = 0.165). *P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001.

Source data

To test whether micro-MKKS63 is functionally required for mitochondrial activity, we generated two independent polyclonal CRISPR–Cas9 KO HMC3 lines using single-guide RNAs (sgRNAs) disrupting the micro-MKKS63 smORF while preserving the mORF. Seahorse Mito Stress Test analysis showed significant reductions in basal oxygen consumption rate (OCR), ATP-linked OCR and maximal respiration across KO lines relative to nontargeting controls (Fig. 7g–k), with consistent effects across two independent sgRNA lines, supporting a role for micro-MKKS63 in microglial oxidative phosphorylation.

Ribosome footprints and MS coverage provide orthogonal evidence for brain-expressed pseudogenes

The MP detection framework used to build this atlas highlights how technical limitations can influence our understanding of genome function. Similar challenges may apply to other genomic elements, such as pseudogenes, which are often classified as nonfunctional because of mutations, including premature stop codons, frameshifts or disrupted splice sites³⁴. However, some pseudogenes may be misclassified because they are difficult to detect with conventional approaches^35,36.

To investigate pseudogene RNA expression, we revisited native transcriptomics data from six AD and six non-AD DLPFC samples, the same datasets used to build our MS proteogenomic search space, as short-read RNA-seq cannot reliably assign multi-mapping reads to their originating loci. Actin pseudogene transcripts were broadly downregulated in AD at nominal significance (Extended Data Fig. 7 and Supplementary Tables 15 and 16), motivating us to test whether ribosome footprint coverage, ShortStop functional predictions and unique tryptic peptide detection using MS could together support MP identification from these loci.

Extended Data Fig. 7 — Long-read RNA sequencing was used to quantify transcript-level changes between AD (n = 6) and control (n = 6) brains (Aguzzoli Heberle et al.). Differentially expressed psORFs were then mapped to their ancestral protein-coding paralogs to identify cytoskeletal gene families. (a) Volcano plot of differential RNA expression of psORF-coding transcript families, AD vs. control. Statistics: two-sided DESeq2 Wald test (negative binomial GLM, design ~ diagnosis + sex), Benjamini-Hochberg-adjusted FDR; n = 6 AD / 6 control biologically independent brains (distinct individuals). Transcripts were called significant at FDR < 0.05 and |log2 fold change|> 1 (colored points). The bar plot (right) summarizes the proportion of significantly altered psORF-coding transcripts by functional class; actin-related and other cytoskeletal pseudogenes show the strongest repression. (b) Box plots of representative actin-related psORFs (ACTB/ACTG1-derived: ACTG1P20, ACTBP9, ACTG1P10, ACTBP11, ACTG1P4) with nominally lower RNA expression in AD vs. control. Plotted values are CPM (shown as CPM + 1). Box plots: center line, median; box bounds 25th and 75th percentiles; whiskers extend to the most extreme data points within 1.5× the interquartile range (defining the plotted minima and maxima); points beyond the whiskers are not shown. Statistics: significance is from per-transcript two-sided DESeq2 Wald tests (negative binomial GLM, design ~ diagnosis + sex) computed on rounded integer ESPRESSO counts (not on the CPM values shown); effect size = log2 fold change. All reach nominal (unadjusted) P < 0.05 but do not survive Benjamini-Hochberg correction (FDR ≥ 0.05): ACTG1P20 log2FC = −1.67, P = 0.022; ACTBP9 −1.93, P = 0.025; ACTG1P10 −1.87, P = 0.028; ACTBP11 −1.69, P = 0.033; ACTG1P4 − 1.84, P = 0.046.

Source data

For ACTBP7, ribosome profiling across 43 non-AD DLPFC samples⁵ revealed unique footprint coverage at the psORF (Fig. 8a), although without typical periodic ribosome occupancy. Periodicity cannot be assessed at ACTBP7 because the window of unique sequence is too short (Fig. 8b). We also mapped two unique tryptic peptides exclusive to ACTBP7 and distinguishable from its ancestral ACTB sequence; both received moderate-grade PROSIT spectra (Fig. 8c,d and Extended Data Fig. 8).

Extended Data Fig. 8 — Alignment of two ACTBP7-smORF peptides against the ACTB reference protein (UniProt P60709) and 342 ACTB missense variants (dbSNP). (a) Top: peptide ESRGIHETTFNSIMK (ACTB positions 268–284); the aligned Arg corresponds to Cys-272 in ACTB. Bottom: peptide DPYANTVLSGGTTRYPGIADR (ACTB positions 293–313); the aligned Arg corresponds to Met-305. Orange, residues unique to the smORF peptide; grey, residues matching ACTB; peach, missense substitutions matching neither. Alignment of the ACTBP11-smORF peptide DDDIAALVIDNSSGMCK to the ACTB N-terminal window (positions 1–18) across the same ACTB variants. (b) No variant matches either smORF peptide, confirming both derive from their smORF loci, not ACTB variation. (c) Representative immunofluorescence, HMC3 cells transfected with empty vector (EV, top) or ACTBP11-smORF (bottom): DAPI (nuclei, blue), anti-ALFA (ACTBP11-smORF, orange), phalloidin-Cy5 (F-actin, grey). Transfections performed 3 times. Transfections performed alongside ACTBP7-psORF presented in Fig. 8 and share same EV comparison. (**d–e**) Effect of ACTBP11-smORF vs. EV on F-actin distribution, HMC3 microglia. Unit of study = individual cell. N = 3 independent transfections (biological replicates), 13–27 EV and 12–14 ACTBP11-smORF cells per transfection (total EV n = 55, ACTBP11-smORF n = 38). (d) Radial F-actin intensity profile, edge to center: line, mean; shaded band, 95% CI. (e) Box plots of whole-cell, nuclear, and nuclear-to-cell F-actin intensity: center line, median; bounds, 25th–75th percentiles; whiskers, most extreme points within 1.5× IQR; outliers not shown; points, individual cells. Statistics (e): linear mixed-effects model, transfection as random intercept for biological replicate structure; two-sided Wald test, ACTBP11-smORF vs. EV; effect size = adjusted mean difference (95% CI, raw a.u.). Whole-cell: −365.7 a.u. (−519.1 to −212.3, 955.1 vs. 1248.3), −29.3%, P = 2.98 × 10⁻⁶ (***). Nuclear: −345.4 a.u. (−496.0 to −194.7), −38.4%, P = 6.99 × 10⁻⁶ (***). Nuclear-to-cell ratio: −0.097 (−0.183 to −0.010), −13.1%, P = 0.028 (*). Asterisks denote * P < 0.05; ** P < 0.01; *** P < 0.001.

Source data

As canonical ACTB is fundamental to cytoskeletal biology, we tested whether ACTBP7-psORF overexpression affects cytoskeletal organization. Its endogenous function is hard to resolve directly because pseudogene loci often are not amenable to CRISPR perturbation and may act only under disease or stress conditions³⁷. Therefore, we overexpressed ACTBP7-psORF in HMC3 microglial cells and assessed F-actin dynamics; after 24 h, IF showed a significant reduction in whole-cell F-actin intensity relative to control (Fig. 8e–i), which is consistent with its ancestral actin-related function. We additionally showed a similar effect for an MP encoded by ACTBP11 (Extended Data Fig. 8). This illustrates how orthogonal evidence can motivate functional studies at loci otherwise difficult to interrogate.

Discussion

Ribosome profiling has established that thousands of smORFs are actively translated. However, extending these catalogs to direct protein-level evidence in aged human tissue and neurodegeneration has remained largely unaddressed. The atlas presented in this article integrates ribosome profiling, MS and a four-tier spectral grading with genomic annotation in the aged human DLPFC. This atlas provides an evidence-graded resource that can directly inform hypothesis-driven experiments on brain MPs in aging and AD.

Our proteogenomic framework offers a scalable model for future MP discovery. Unlike prior studies that often pair transcriptomic and proteomic data from unrelated tissues and donors, our analyses are strictly matched to the same brain region. Searching for MS evidence across more than 600 DLPFC samples using a reference database built from 12 native DLPFC long-read transcriptomes should reduce the risk of unexpressed transcripts inflating the FDR penalty. Verifying RNA expression of MP-coding smORFs in the same donor samples further restricted the search space. This pipeline identified over 3,000 MPs not reviewed in UniProt/Swiss-Prot (Fig. 1), over 1,000 of which strongly match PROSIT-predicted spectra (Fig. 2).

Future workflows can leverage the PROSIT-based spectral grading framework established in this study. False positives remain a major challenge in large-scale proteomics, yet manually inspecting individual PSMs is impractical at scale. Our pipeline instead grades thousands of PSMs empirically, across nearly 2,000 raw files from ROSMAP TMT-MS data. This is especially valuable for MPs, which often lack multiple tryptic peptides. Conventional guidelines require two unique peptides per protein, but MPs rarely meet this threshold given their short length (Fig. 2). By identifying over 800 single-peptide strong-grade unreviewed MPs, our framework improves detection confidence without sacrificing sensitivity. Ribosome footprint quantitation adds orthogonal evidence of translation. This matters most at challenging loci, where read multi-mapping and peptide non-uniqueness complicate detection, as shown for pseudogenes through convergent long-read transcriptomic, ribosome profiling and MS evidence.

Most MPs in this atlas are encoded by loci with existing protein-coding annotations, which is consistent with prior observations across organisms^8,20,38,39, yet these additional products can reveal regulatory complexity missed by gene-centric analyses. The NYAP1 locus, encoding an adapter protein in PI3K/Akt signaling and neuron projection morphogenesis, also produces a distinct uoORF-encoded MP with strong spectral evidence (Fig. 2) that carries a predicted TDP-43 (TARDBP) splicing motif. Given TDP-43’s central role in frontotemporal dementia and amyotrophic lateral sclerosis (ALS)^26,40, age-related changes in its activity may alter isoform usage at neurodegeneration-relevant loci in ways standard analyses miss. Similarly, the BCL3 iORF is downregulated in sym-AD at the exploratory tier while canonical BCL3 is unchanged (Fig. 3), showing MPs can be independently regulated and serve as biomarkers distinct from their parent genes, consistent with iORFs representing an understudied source of bioactive peptides⁴¹.

A reproducible feature of this atlas is that a gene’s smORF-containing transcripts can be expressed independently of the transcripts carrying its main ORF. This pattern held across brain regions, cohorts, and workflows when analyzing ROSMAP and MSBB transcriptomics data (Fig. 4). Alternative promoters explain independent expression for some uORFs separate from their mORF⁴²; however, in our studies, nearly three-quarters of MPs share promoters with their main ORFs (Extended Data Fig. 5), pointing instead to posttranscriptional mechanisms. Accordingly, polyadenylation and 3’ splice site factors were enriched at dORF junctions, and GC-rich splicing regulators at short isoform junctions. This suggests isoform-selective RBP activity shapes which transcripts retain smORFs (Fig. 5)^43–45.

Two interpretive paradigms apply to MPs with evidence of expression. First, an MP may reflect altered transcription or splicing of its encoding gene, serving as a marker of regulatory disruption rather than a driver of disease processes^46–48. Second, it may function as an independent bioactive polypeptide with activities distinct from its gene’s main canonical protein product. These possibilities are not mutually exclusive and will require endogenous perturbation approaches to distinguish between them. Regardless of the independent functions of MPs, MP-producing loci reveal an additional layer of genomic output that is obscured by conventional gene-level analyses.

The MKKS locus illustrates the second paradigm in which the MP itself has independent biology. In human DLPFC, the canonical 570-aa MKKS protein shows minimal ribosome profiling support (Fig. 6) and is largely undetected in both ROSMAP TMT-MS data and the PepCentric meta-database of 66,700 public MS runs (Extended Data Fig. 6). Instead, micro-MKKS63 is the locus’s dominant proteomic product, is downregulated in symptomatic AD, localizes to mitochondria and is required for microglial oxidative phosphorylation (Fig. 7). Although the main MKKS protein product is clinically associated with McKusick–Kaufman syndrome^29,49, a developmental ciliopathy, its mechanisms may not reflect those of the aged brain. This highlights a broader annotation challenge, whereby the most biologically relevant protein at a locus can vary according to tissue and differ from the canonical annotated product. How often this occurs, and whether some disease-associated loci are shaped by co-encoded MP biology, remains unknown.

Confidence in MP identity scales with the convergence of independent evidence rather than any single criterion. Ribosome profiling establishes translational engagement, MS provides direct peptide-level detection and spectral grading separates reliable identifications from unreliable ones. No single readout is sufficient on its own. Peptide-count thresholds alone would have excluded most bona fide MPs (Fig. 2). Gene-level transcript counts alone would have missed the isoform-specific regulation uncovered for BCL3, RAB3C and MKKS (Figs. 3–5). Pseudogene-derived MPs may also be overlooked (Fig. 8). Integrating these methods enables confident calls where any single one would fail; we expect this convergence-based framework to generalize to MP discovery beyond the brain and beyond AD.

This atlas provides the foundation to test the role of MP-encoding genes as an additional layer of biological regulation relevant to neurodegeneration. The expression profiles, spectral grading outputs and processing pipelines described in this study are publicly available as an evidence-graded reference and template for future MP discovery.

Technical considerations and limitations

The following technical limitations should be considered for the MP detection workflow used here. First, the absence of MS detection does not imply absence of expression. For instance, over one-third of reviewed MPs lacked MS support despite evidence of RNA expression and translation. Moreover, our MS search database that was built from 12 native transcriptomes may have missed some brain-expressed MPs from unassembled transcript isoforms. Future work should resolve which transcript isoform each detectable MP originates from and examine whether the sequences reported arise from alternative transcripts and/or ribosome slippage, which may present isoforms that our models cannot capture. In addition, our smORF-mORF analyses relied on short-read RNA-seq, which cannot fully resolve isoform-level expression at complex loci; scaling native transcriptomes in future work can help address this. MP nomenclature also remains unstandardized, reflecting each discovery’s historical context rather than any intrinsic difference between proteins. For example, somatostatin meets the size criterion but predates the label by decades, while ELABELA’s inclusion reflects its host transcript’s initial noncoding annotation⁵⁰. Positional labels such as dORF or uORF assume that a canonical ORF exists at the locus, which is not always the case, as at SLC35A4 (ref. ⁵¹). We retain the MP designation per convention, although terminology may change as standards mature.

Broader implications

The brain proteome is more complex than current gene models reflect. Analyses anchored to canonical annotations risk overlooking MPs by collapsing transcriptomic reads from the mORF and smORF into a single gene body, as illustrated by MKKS. Micro-MKKS63 raises a broader question: if an MP, rather than the canonical protein, is a locus’s main translational output in a given tissue, then functional studies based on canonical annotation may target the less abundant, and potentially less disease-relevant, protein. This framework, integrating ribosome profiling, MS and spectral grading, is not specific to the DLPFC or AD and could support MP discovery elsewhere.

Methods

Human postmortem brain cohorts

Human postmortem brain tissue and associated data were collected under institutional review board-approved protocols at the institution of origin. Four autopsy collections were included: the MSBB (AMP-AD Knowledge Portal, Synapse ID: syn3159438), the ROSMAP (AMP-AD Knowledge Portal, Synapse ID: syn3219045), the Sanders-Brown Center on Aging at the University of Kentucky (long-read sequencing data deposited under Synapse ID: syn52047893) and the UC San Diego (UCSD) Shiley-Marcos Alzheimer’s Disease Research Center (DIA-MS data deposited at the PRIDE Archive under accession no. PXD071500). Long-read transcriptomic data were generated from DLPFC (Brodmann area 9/46) samples from the Sanders-Brown cohort using Oxford Nanopore sequencing (n = 12; six control, six AD; 50% female; mean age 84.4 ± 6.2 years). These data were used to characterize transcript structures and smORF-containing isoforms. The primary TMT-MS proteomics cohort consisted of ROSMAP DLPFC samples. A total of 610 samples were processed through the proteogenomic pipeline and 608 survived quality thresholds (round 1, n = 400; round 2, n = 208; 69.1% female; mean age 87.8 ± 4.4 years). Based on neuropathological diagnosis and principal component analysis-based outlier removal, 480 samples were retained for differential peptide analysis (111 non-AD, 182 asym-AD, 187 sym-AD; 69.8% female; mean age 87.9 ± 4.3 years). These samples were used to generate residualized peptide abundance values for statistical analysis. A matched ROSMAP bulk RNA-seq cohort (n = 373; 90 non-AD, 141 asym-AD, 142 sym-AD; 66.2% female; mean age 88.0 ± 4.4 years) was used for the transcriptomic analyses corresponding to the TMT-MS cohort. An independent replication RNA-seq cohort was obtained from MSBB parahippocampal gyrus tissue (Brodmann area 36; n = 258; 57 non-AD, 22 asym-AD, 179 sym-AD; 64.3% female; mean age 83.8 ± 8.0 years). Fresh-frozen frontal cortex tissue from the UCSD cohort was used for DIA proteomics (n = 4; mean age 82.5 years; two with AD, one with amnestic mild cognitive impairment, one cognitively normal). In addition, publicly available ribosome profiling data from Duffy et al.⁵ were re-analyzed to provide an independent translation dataset (n = 43 adult brain samples; 41.9% female; mean age 39.6 ± 14.7 years). Neuropathological evaluation used Consortium to Establish a Registry for Alzheimer’s Disease (CERAD) criteria to assess neuritic plaque burden and Braak staging to evaluate neurofibrillary tangle progression. Additional neuropathological diagnoses were assigned according to established criteria and guidelines^52–55. ROSMAP and MSBB cases were classified using a cross-cohort harmonization framework²³: non-AD, CERAD 0–1 and Braak 0–3 without dementia at last evaluation (with CERAD required to equal 0 when Braak = 3); asym-AD, CERAD 1–3 and Braak 3–6 without dementia; and symptomatic AD, CERAD 2–3 and Braak 3–6 with dementia. Dementia was defined as a Mini-Mental State Examination < 24 or Clinical Dementia Rating ≥ 1. Per-case demographic and clinical variables are provided in Supplementary Table 17.

Long-read transcriptome assembly and putative ORF database construction

To support proteomic discovery of MPs, we built a custom putative ORF database by appending unannotated protein sequences absent from the reference human proteome (UniProtKB UP000005640). These sequences were identified using Oxford Nanopore long-read RNA-seq data from dataset syn52047893, generated by the Sanders-Brown Center on Aging at the University of Kentucky¹⁶. The dataset includes RNA from 12 postmortem brain samples (Brodmann area 9/46). Raw reads were aligned to GRCh38 (with External RNA Controls Consortium spike-ins) using minimap2 (v.2.26-r1175, -t 64 -ax splice -ub -k14 -w 4 --secondary=no), compressed with samtools (v.1.18) and sorted with sambamba (v.1.0.1). ESPRESSO (v.1.0) assembled transcript models and quantified isoforms via its S, C and Q steps against Gencode v.43, using --sort_buffer_size 450 G, --read_num_cutoff 2 and --read_ratio_cutoff 0 for the S and Q steps. The resulting Gene Transfer Format (GTF) and isoform count table (via expectation maximization) were sorted with GFF3sort for frame-preserving translation. We retained isoforms exceeding 0.5 CPM, three-frame translated them using our GTFtoFASTA tool⁸ and appended the resulting peptides to the canonical proteome (UniProtKB UP000005640, 105,486 Swiss-Prot/TrEMBL entries). This yielded 2,174,307 putative ORFs for downstream MS searches.

TMT-based proteomics and protein identification analysis

ROSMAP TMT-MS data included TMT 10-plex and TMTpro 16-plex labeling, high-pH fractionation and liquid chromatography (LC)–MS/MS on Orbitrap-class instruments. Raw spectral data were analyzed using FragPipe (v.22.0), Philosopher (v.5.0) and IonQuant (v.1.8.5), with searches run in TMT mode using MS² or MS³ quantification depending on the batch acquisition method. PSMs were identified via a two-pass search and rescoring workflow. The first pass searched the full proteogenomic database with a precursor and fragment mass tolerance of ±20 ppm, trypsin digestion allowing two missed cleavages and a minimum peptide length of seven amino acids. Carbamidomethylation of cysteine was fixed, while methionine oxidation and protein N-terminal acetylation were variable modifications. TMT reporter ion quantification used manufacturer-provided lot correction factors. Identifications were rescored with Percolator via FragPipe’s MSBooster module and filtered at FDR <0.01 at both PSM and protein levels using Philosopher’s decoy-based approach. The second pass re-searched proteins identified at FDR < 0.01 using the same parameters to improve sensitivity while limiting false positives. Only identifications supported in both passes at 1% FDR were retained.

smORF classification according to genomic context and isoform homology

We classified smORF-encoded MPs using Ensembl transcript biotype annotation and BLASTP protein homology. Strand-specific overlaps between predicted smORFs and Ensembl models were identified using bedtools (v.2.30.0), with a custom parser assigning annotations according to feature type and biotype. Full smORF definitions are listed in Supplementary Table 18. Isoform-derived smORFs resembling canonical proteins were identified using BLASTP alignment against UniProtKB UP000005640 (E-value 1 × 10⁻⁴), with bit scores of 40 or greater classified as protein isoforms.

Ribosome profiling re-analysis using Rp3 to assess translation of MS-identified MPs

To assess translational evidence for MS-detected MPs, we re-analyzed publicly available Ribo-seq data from 43 adult human DLPFC samples (Duffy et al.⁵; dbGaP accession no. phs002489.v1.p1) using our Rp3 framework (v.1.2.2)³⁹, which combines proteogenomics searches with Ribo-seq read re-mapping. While RiboCode identifies translated smORFs via 3-nt periodicity and ribosome occupancy, Rp3’s ribocov mode re-maps reads with featureCounts under less stringent, multi-parameter conditions allowing multi-mapping and ambiguous mapping, recovering pseudogenes and multi-copy smORFs that RiboCode’s uniquely mapped read approach would discard. Ribocov trims reads with fastp (adapter/poly(A) removal), removes rRNA and tRNA contamination via STAR (--outSAMstrandField intronMotif, --outReadsUnmapped Fastx), then aligns unmapped reads with STAR (--outSAMstrandField intronMotif, --outFilterMismatchNmax 2, --outFilterMultimapNmax 9999, --chimScoreSeparation 10, --chimScoreMin 20, --chimSegmentMin 15, --outSAMattributes All, --clip5pNbases 3; library-prep specific), the 9,999 cap retaining reads at pseudogenes and complex regions. featureCounts was run four times (default, -M, -O, and -M and -O combined), with coverage plotted as heatmaps via pheatmap. To score the smORF coding potential, the ESPRESSO-generated GTF was input to TransDecoder to annotate CDS features and predict the longest ORF per transcript; a custom STAR (v.2.7.4a) index was built via --sjdbGTFfile alongside GRCh38. Rp3-trimmed reads were aligned to the transcriptome (as required by RiboCode) with STAR (--quantMode TranscriptomeSAM GeneCounts, --outFilterType bySJout, --alignEndsType EndToEnd). We inferred 3-nt periodicity via RiboCode’s prepare_transcripts, metaplots and RiboCode commands (default parameters, 1% FDR cutoff).

Machine-learning-based classification of translated smORFs using ShortStop

To identify which smORFs detected as translated by RiboCode encode MPs with features of bioactivity, we used our previously reported ShortStop classifier¹². ShortStop classifies smORFs based on physicochemical and sequence-derived features. The model was trained on two biologically grounded reference sets, with Swiss-Prot annotated proteins shorter than or equal to 150 aa serving as positive examples and physicochemically resembling in silico MPs serving as negative examples. Physicochemically resembling in silico MPs are synthetic smORFs that match the length, codon bias and amino acid composition of actively translated smORFs with stochastic generative sequence order.

DIA MS of human DLPFC tissue

Human frontal cortex tissue (four 125-mg samples, UCSD Shiley-Marcos Alzheimer’s Disease Research Center) was lysed in 8 M urea in 100 mM sodium phosphate (pH 7.5) with Halt protease/phosphatase inhibitor cocktail (Thermo Fisher Scientific) using a Precellys Cryolys Evolution homogenizer (Bertin Technologies) in 2-ml CK14 tubes with 1.4-mm zirconium beads (3 × 20 s, 6,800 rpm, 30-s rests), then centrifuged (17,000g, 30 min, 4 °C). Supernatants were sonicated (3 × 5 s, 30% amplitude, 15 s on ice between cycles), centrifuged again identically and processed for MS; protein concentration was measured using a bicinchoninic acid assay. Extracts were reduced (dithiothreitol), alkylated (iodoacetamide), digested overnight (trypsin/Lys-C), quenched with formic acid, dried, desalted on C18 columns (Nest Group BioPureSPN Mini, cat. no. HUM S18V), dried again and spiked with PRTC standard (cat. no. 88320, Pierce) for normalization. Peptides were separated on a Vanquish Neo (Thermo Fisher Scientific) with an EASY-Spray PepMap Neo column (75 µm × 150 mm, ES75150PN; 75 µm × 500 mm, ES75500PN for 45–60-min runs; Buffer A, 0.1% formic acid and water; Buffer B, 80% acetonitrile and 0.1% formic acid), with MS/MS on an Orbitrap Astral (Thermo Fisher Scientific). All analysis used DIA-NN (v.2.2.0), with a spectral library from a custom FASTA (DDA-identified, RiboCode-identified and ShortStop-identified sequences appended to the human UniProt reference proteome). Search settings were: tryptic digestion (K/R cleavage, not before proline, one missed cleavage); peptide length 6–45 aa; precursor charge +2 to +4, precursor m/z 395–1,005, fragment m/z 200–1,800; fixed cysteine carbamidomethylation; variable methionine oxidation/N-terminal acetylation; methionine excision enabled; MS/MS tolerance 10 ppm; MS¹ tolerance 4 ppm; and protein group level 1 reporting. All mzML files were searched simultaneously (parquet output), then filtered in R at q ≤ 0.01 and Global.PG q ≤ 0.01, requiring at least one proteotypic peptide.

Spectral validation of MP PSMs using PROSIT

To validate MP PSMs, we compared observed and predicted fragment spectra according to SA. Fragment-ion intensities for all unique tryptic MP peptides were predicted with the Prosit_2020_intensity_TMT model via the Koina API (koina.wilhelmlab.org; NCE 38, HCD)¹⁹; observed spectra were retrieved from mzML files across 64 TMT batches using pyteomics. For each peptide, up to five candidate PSMs were collected, filtered according to spectral purity (≥ 0.5) and restricted to the rank-1 hit’s charge state and modified sequence; the representative spectrum was chosen by most matched fragment ions, then evidence tier, then lowest SA. Spectral similarity was the normalized SA = 1 − (2/π) × arccos(cos θ), where cos θ is the cosine similarity between matched predicted and observed intensity vectors at 20 ppm tolerance (SA bounded 0–1). Match coverage was the fraction of PROSIT-predicted fragments (normalized intensity > 1 × 10⁻⁴) with a matched observed peak within tolerance. Each peptide received an evidence label (strong, moderate, weak, insufficient) combining SA rating (high ≤ 20 degrees, moderate 20–45, poor 45–70, terrible > 70) and coverage category (majority > 50%, partial 25–50%, sparse < 25%), according to Wacholder et al.²⁰.

Normalization, covariate adjustment and differential expression analysis of TMT proteomics data

TMT reporter ion intensities were extracted using IonQuant and processed in R via a custom pipeline to integrate, normalize and model proteomic data across both acquisition rounds. Quantified FragPipe protein group tables (protein, peptide and PSM FDR 0.01) were merged across all TMT batches. Normalization used TMT Analysis via Median Polish on Repeated standards, adapted from Johnson et al.²³ and Dammer et al.⁵⁶, leveraging repeated global internal standards (GIS) to control batch effects while preserving the biological signal. Intensities were normalized to the pooled GIS channel by subtracting log₂-transformed GIS intensity per sample, then harmonized across batches via median polish, removing batch-specific offsets. This was processed separately for rounds 1 (TMT 10-plex) and 2 (TMTpro 16-plex), then combined. Principal component analysis on the GIS-normalized matrix identified outliers (z-score > 3 along the first two principal components). Bootstrap-based linear modeling fitted age, sex, postmortem interval and diagnosis, taking median coefficients. Expression was adjusted by regressing out age, sex and postmortem interval, retaining diagnosis-associated variance for differential analysis, using the model:

\begin{matrix} {Protein}_{ij} = β_{0} + β_{1} \times {Age}_{j} + β_{2} \times {Sex}_{j} + β_{3} \times {PMI}_{j} \\ + β_{4} \times {Batch}_{j} + β_{5} \times {Round}_{j} + ϵ_{ij} \end{matrix}

The residual $ϵ_{ij}$ was used for t-tests, and P values were corrected using Benjamini–Hochberg correction to control the FDR. Analyses used R with custom scripts. This pipeline reproduced previously reported fold changes for UniProt/Swiss-Prot proteins on the same dataset²³.

Bulk RNA-seq profiling and differential expression analysis

Transcriptomic profiling was performed on DLPFC tissue from ROSMAP tissues, matched according to individual ID and tissue block to TMT proteomics samples for multi-omic integration; only samples with high-quality RNA-seq and matching proteomics were retained. One subset used 150-bp paired-end reads aligned via STAR (v.2.7.10a) with a splice-aware index and --alignIntronMax 1000000, --outFilterMultimapNmax 2, --outFilterMismatchNoverReadLmax 0.04, producing coordinate-sorted BAMs; a second subset used pre-aligned BAMs directly. Gene expression was quantified with featureCounts (subread v.2.0.1, -O flag for multi-CDS assignment) against Gencode v.43 and a custom GTF appending smORFs of interest, merged into a unified matrix in R. Differential expression used DESeq2 on 373 matched samples (AD n = 142, asym-AD n = 141, non-AD n = 90), with median-of-ratio normalization, empirical Bayes dispersion shrinkage and negative binomial generalized linear models (GLMs) with Wald tests under a multi-factor design. The following model was used.

Expression \sim RIN + sequencing Batch + Sex + Age + PMI + Diagnosis

To test transcriptional coupling between MPs and host genes, we computed Pearson correlations between each smORF and parent gene separately in non-AD and sym-AD (VST reads; minimum three observations per group, single Ensembl identifier required). Correlation shift Δ-r = r(sym-AD) − r(non-AD) classified pairs as coupled (|Δ-r| ≤ 0.1) or non-coupled (|Δ-r| > 0.1), tested using Fisher’s z-test. Three nested models (null: smORF ~ mORF; additive: + diagnosis; interaction: × diagnosis) used likelihood ratio tests to assess independent AD effects and slope differences between groups. As a replication cohort, MSBB RNA-seq²⁵ was processed identically using pre-aligned BAMs against hg19. Genes expressing a smORF and an mORF were assessed for coexpression by constructing Pearson’s correlation values on VST matrices. Cross-cohort concordance used Pearson correlation between ROSMAP and MSBB pair-wise r estimates. Analyses used R with DESeq2, limma, dplyr, ggplot2, psych and patchwork.

snRNA-seq integration

To assess cell-type-specific associations between MP-coding genes and AD pathology, we mapped MS-detected MPs onto previously reported snRNA-seq differential expression data from ROSMAP DLPFC participants (2.3 million nuclei, 54 cell types)²⁴. Four candidate genes (SST, MDK, BCL3, RAB3C) were examined in depth: log(fold change) and log(CPM) were extracted per cell type and visualized as a heatmap, with cell types grouped into broad classes (excitatory and inhibitory neurons, astrocytes, oligodendrocytes, microglia), marking genes reaching FDR significance (P_adj < 0.05) per cell type. Analyses used R with dplyr, ggplot2 and ggh4x.

Transcriptional regulatory and RBP motif analyses

Promoter regions were defined per gene using a custom annotation combining Gencode v.43 protein-coding entries and MP loci, with windows spanning −1,000 to +100 bp relative to each TSS (1,100 bp per gene). Transcription factor motif enrichment used HOMER (v.4.11) findMotifsGenome.pl against a random genomic background, combining de novo discovery with enrichment testing against the HOMER-curated transcription factor database. Combined known and de novo motif models were scanned per gene with HOMER annotatePeaks.pl, applied separately to MP and canonical TSS windows to record motif counts and positions. For each smORF with an annotated parent gene, transcription factor motifs were compared between smORF and parent TSS windows and classified as unique to the smORF, unique to the parent or shared. RBP motif occurrences at splice-junction windows were identified with FIMO (MEME Suite v.5.5.9) against 98 position weight matrices from the CISBP-RNA human RBP database, using a uniform nucleotide background and P < 1 × 10⁻⁴. Each gene was assigned to one of four positional classes (downstream, upstream, internal or isoform) by its predominant smORF type. Statistical testing was two-stage: a pooled Fisher’s exact test (Benjamini–Hochberg FDR correction) compared all smORF-associated junctions against the no-smORF background, retaining RBPs at FDR < 0.05, followed by separate Fisher’s exact tests (independent Benjamini–Hochberg correction) for each smORF subtype against the same background. Top RBPs for visualization were selected according to cross-subtype variance in log₂(odds ratio). RBP disease associations were queried via the Open Targets Platform GraphQL API (v.4), resolving each gene symbol to an Ensembl ID and retrieving up to 10,000 diseases per gene. Associations were filtered with 12 dementia-related and neurodegeneration-related keywords (dementia, Alzheimer, ALS, frontotemporal, neurodegeneration, Parkinson, Lewy body, Huntington, prion, tauopath, inclusion body, motor neuron disease), retaining the maximum Open Targets overall association score (0–1, integrating genetics, transcriptomics, animal models, literature mining and drug target evidence) across matched diseases per RBP.

Differential expression analysis of long-read RNA-seq data

After ESPRESSO isoform assembly (v.1.0), expression values were extracted from the resulting .esp file (long-read RNA-seq, 12 DLPFC samples: six with AD, six controls), which reports unnormalized fractional isoform counts. Counts were rounded to integers and modeled in DESeq2 via negative binomial GLMs (Wald tests) on diagnosis and sex. psORF-encoded MPs were grouped according to parental gene family: actin-cytoskeleton (for example, ACTB, ACTG1), non-actin-cytoskeleton (for example, TUBA1A/TUBA1B, SEPTIN7, BIN2) and other, comparing fold change, expression trends and differential expression proportions across groups. Visualization used ggplot2 and patchwork.

Cell culture, transient transfection and immunocytochemistry

HMC3 human microglial cells (cat. no. CRL-3304, ATCC) were transfected with pcDNA3.1 constructs bearing C-terminal FLAG or ALFA tags separated from the CDS by a GGSG linker. Constructs were ordered from GenScript and sequence-verified. Cells were transfected using the Neon Transfection System (Thermo Fisher Scientific): 1 × 10⁶ cells in 130 µl Neon Buffer R with 2 µg plasmid DNA, electroporated at 1,200 V, 20 ms, two pulses, then incubated 24 h before lysis or fixation. For immunocytochemistry, cells were seeded in µ-Slide chambers (25,000–75,000 per well, Ibidi), fixed in 4% paraformaldehyde (10 min), washed (2× Tris, 1× PBS), permeabilized (0.1% Triton X-100, 10 min) and blocked (2% BSA, 60 min). Primary antibodies included mouse anti-FLAG M2 (1:1,000 dilution, cat. no. F1804, Sigma-Aldrich), rabbit anti-Mitofilin (1:1,000 dilution, cat. no. 10179-1-AP, Proteintech), rabbit anti-ALFA polyclonal (1:500 dilution, cat. no. N1581, NanoTag) and FluoTag-X2 anti-ALFA 568 (1:500 dilution, cat. no. N1502, NanoTag), with phalloidin (1:400 dilution, Invitrogen) for F-actin. Cells were stained overnight (4 °C) with anti-FLAG M2 and anti-Mitofilin, or for 60 min with FluoTag-X2 anti-ALFA 568 and phalloidin. Secondary detection used anti-mouse Alexa Fluor 488 (cat. no. 115-545-003, Jackson ImmunoResearch) and anti-rabbit Alexa Fluor 594 (cat. no. 111-585-003, Jackson ImmunoResearch). Nuclei were stained with DAPI and fixed-cell preparations were mounted in Ibidi antifade mounting medium (cat. no. 50011). Imaging used a Nikon A1R confocal (×40–60 oil immersion, 488/561-nm, GaAsP PMT, 1.0 AU pinhole, NIS-Elements v.AR.6.10.02), processed in Fiji. Micro-MKKS63/IMMT colocalization was quantified using Otsu-thresholded Manders coefficients (M¹, M²) and Pearson’s correlation within mitochondrial regions of interest, plus a line intensity profile at maximal overlap, using Python (NumPy, SciPy, scikit-image). F-actin distribution was quantified from ND2 images via an automated Python pipeline (GitHub: brain-microprotein-atlas). Nuclei were segmented from DAPI (Gaussian smoothing, Otsu thresholding, watershed, size/shape/circularity filtering); cell bodies were delineated from F-actin (Cy5) via seeded watershed matched to nuclei and area-ratio-filtered. Intensity was profiled along the minor axis in ten bins (nuclear centroid to cortex), with whole-cell and nuclear means extracted separately and assessed using a linear mixed-effects model.

Acid-based extraction and LC–MS/MS analysis of HMC3 microglial cells

HMC3 microglial cells were lysed using acid-based extraction optimized for low-molecular-weight recovery (50 mM HCl, 0.1% β-mercaptoethanol, 0.05% Triton X-100). Centrifuged supernatants underwent solid-phase extraction on BondElut C8 cartridges (~100 mg C8 per 10 mg lysate protein, Agilent Technologies), conditioned with methanol and equilibrated with triethylammonium formate (TEAF) (pH 3.0). After loading and a TEAF wash, MP-enriched fractions were eluted in 75% acetonitrile/25% TEAF (pH 3.0) and lyophilized. Eluates were digested overnight at 37 °C in 2 M urea in 100 mM tetraethylammonium bicarbonate (pH 8.5) with sequencing-grade trypsin (Promega Corporation), quenched with 5% formic acid and desalted on C18 StageTips. Peptides were analyzed on an Orbitrap Fusion Lumos Tribrid (Thermo Fisher Scientific) with an Easy-nLC 1000, on a 25 cm × 100 µm in-house column (BEH 1.7 µm C18, Waters). Separation ran at 400 nl min⁻¹ over 140 min (Buffer B, 0.1% formic acid/90% acetonitrile: 1–25% over 100 min, 25–40% per 20 min, 40–90% per 10 min, 90% per 10 min; Buffer A, 0.1% formic acid/water; 2.5 kV spray). The Lumos ran in data-dependent acquisition mode: full MS at 120,000 resolution (400–1,500 m/z, AGC 4 × 10⁵), with a top-speed 3-s cycle selecting precursors for CID in the ion trap. MS/MS used 1.6 m/z isolation, AGC 2 × 10⁴, a minimum intensity of 5,000 and injection times of 50 ms (MS¹) and 35 ms (MS²), with monoisotopic precursor selection and 5-s dynamic exclusion. Raw spectra were searched with the standardized FragPipe pipeline and combined FASTA database used for the ROSMAP TMT-MS analysis.

Coexpression and GO enrichment analysis

To investigate the transcriptional context of micro-MKKS63 relative to the canonical MKKS transcript, we performed coexpression analysis using RNA-seq data from DLPFC samples in the ROSMAP cohort, with count matrices generated as described for the manuscript’s differential expression analyses. VST-normalized expression values were used to compute Pearson correlations between each target transcript’s CDS (micro-MKKS63 at chr20:10420549–10420737 and canonical MKKS ENSG00000125863.20) and all other expressed genes (over 61,000 tested), separately for sym-AD, non-AD and combined groups. Genes with |r| > 0.7 were considered strongly coexpressed; network overlap between sym-AD and non-AD was quantified using Jaccard indices. GO enrichment analysis was performed using clusterProfiler with enrichGO() and the org.Hs.eg.db database, assessing the Biological Process, Cellular Component and Molecular Function domains, with terms considered significant at a Benjamini–Hochberg P_adj< 0.05.

Re-analysis of BioID proximity-labeling MS data

Raw MS data were processed in FragPipe (v.22.0). MSFragger searched the human proteome (UniProt UP000005640, with contaminants and decoys) using trypsin (two missed cleavages), 7–50 residue peptides, 20-ppm tolerances, fixed cysteine carbamidomethylation and variable methionine oxidation, and N-terminal acetylation. Results were filtered to 1% FDR (PSM/peptide/protein) using Percolator, ProteinProphet, and Philosopher, and quantified using IonQuant (MaxLFQ, normalized). Differential enrichment used the IonQuant MSstats output. For proteins quantified in both mito-FL and untransduced conditions (≥ 2 replicates per group, non-zero variance), Welch’s t-tests on log₂ intensities gave Benjamini–Hochberg-corrected log₂(fold changes) (FDR < 0.05 significant), shown as a volcano plot highlighting micro-MKKS63. Analyses used R (dplyr, broom, ggplot2 packages).

CRISPR–Cas9 editing of micro-MKKS63 and mitochondrial stress testing

Lentivirus was produced in HEK 293FT cells (cat. no. R70007, Invitrogen) by transfecting pMD2.G (cat. no. 12259, Addgene), pMDLg/pRRE (cat. no. 12251), pRSV-Rev (cat. no. 12253) and either lentiGuide-Puro (cat. no. 52963, Addgene) or lentiCas9-Blast (cat. no. 52962) with PEI-Max (cat. no. 24765, Polysciences); supernatant was collected at 48 h, 0.45-µm-filtered and stored at −80 °C. Stable Cas9-expressing HMC3 cells were generated using transduction with lentiCas9-Blast (multiplicity of infection = 0.5) and blasticidin selection (10 µg ml⁻¹), and used as the parental line for all sgRNA work.

Two sgRNAs targeting micro-MKKS63 were designed with CCTop⁵⁷ (--pam NGG --fwdOverhang CACCG --revOverhang AAAC --maxOT 5): sgMicro-MKKS63(1), 5′-ATTCCTGAGCAAGATAGTCT-3′; and sgMicro-MKKS63(2), 5′-AGTCCCAGACTATCTTGCTC-3′. A nontargeting control from the Brie library⁵⁸ (5′-AAAAAGCTTCCGCCTGATGG-3′) served as the negative control. Complementary sgRNA oligonucleotides (Eton Bioscience) containing BsmBI-compatible or Esp3I-compatible overhangs were annealed, phosphorylated with T4 PNK and cloned using Golden Gate assembly into an in-house lentiviral sgRNA backbone derived from lentiGuide-Puro. The puromycin-resistance cassette was replaced with BFP for nontargeting controls or mCherry for micro-MKKS63-targeting sgRNAs. Assemblies were transformed into Stbl3 Escherichia coli, selected on 100 µg ml⁻¹ ampicillin; single colonies were expanded and sequence-verified. Cas9-expressing HMC3 cells were transduced (multiplicity of infection = 0.3), and polyclonal marker-positive populations were recovered using fluorescence-activated cell sorting against untransduced HMC3. Editing was confirmed using PCR across the cut site (forward, 5′-GTAATGGTAAGGAGTCAGGGGCT-3′; reverse, 5′-TGTTGCCAAAAAGTCAGTATTGC-3′; ~700 bp) followed by Sanger sequencing and TIDE analysis⁵⁹. Constructs generated in this study are available from the corresponding author upon request. Cas9/sgRNA HMC3 cells were seeded at 40,000 per well in Seahorse XF96 plates and assayed with the XF Mito Stress Test (oligomycin 2 µM, CCCP 3 µM, rotenone and antimycin A 1 µM each) on a Seahorse XF96 Analyzer. The experiment was performed in two polyclonal micro-MKKS63 cell lines using two sgRNAs across three independent biological replicates per day with similar results. OCR was normalized to total protein per well (bicinchoninic acid assay) and expressed as OCR per mg per experiment.

Statistics and reproducibility

Sample sizes were not statistically predetermined but are consistent with previous publications; our TMT-MS workflow reproduced the direction and magnitude of previously reported effects on the same ROSMAP datasets²³. Cohort sizes reflect donors with high-quality data (12 long-read RNA-seq, 610 TMT proteomics, 373 matched bulk RNA-seq); in vitro sizes followed standard practice. For TMT proteomics, technical outliers (|z-score| > 3 along the first two GIS-normalized principal components) were excluded, yielding 480 individuals; no other data were excluded. Data collection and analysis were not blinded; cohort studies were observational, with diagnoses by established criteria (CERAD, Braak, Mini-Mental State Examination); cell and CRISPR experiments compared genotypes and treatments without randomization. Statistical tests, detailed in the Methods, spanned two-sided t-tests, negative binomial GLM Wald tests (DESeq2), Fisher’s exact tests, Fisher z and likelihood ratio tests, Wilcoxon rank-sum tests and one-way analysis of variance with post hoc testing. Normality was assumed but not formally tested; parametric analysis figures show box plots, violin plots and individual points. Multiple testing was Benjamini–Hochberg-corrected (FDR/P_adj< 0.05 unless otherwise stated). Analyses used R (v.4.4.1) and, where noted, Python (NumPy v.2.0.2, SciPy v.1.13.1, scikit-image v.0.24.0). Analyses used R with DESeq2 (v.1.44.0), limma (v.3.60.6), dplyr (v.1.1.4), ggplot2 (v.4.0.0), psych (v.2.6.1), patchwork (v.1.3.2), ggh4x (v.0.3.1), broom (v.1.0.10) and clusterProfiler (v.4.12.6), and, where noted, Python (NumPy v.2.0.2, SciPy v.1.13.1, scikit-image v.0.24.0).

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Supplementary information

Supplementary Information^{(316.6KB, pdf)}

1. All data and code generated in this study are publicly available across four repositories. 2. Source data for all figures are provided.

Reporting Summary^{(3MB, pdf)}

Peer Review file^{(1.2MB, pdf)}

Supplementary Tables 1–18.^{(27.9MB, xlsx)}

Source data

Source Data Fig. 1^{(1.4MB, xlsx)}

Statistical source data.

Source Data Fig. 2^{(551.9KB, xlsx)}

Statistical source data.

Source Data Fig. 3^{(409KB, xlsx)}

Statistical source data.

Source Data Fig. 4^{(747.2KB, xlsx)}

Statistical source data.

Source Data Fig. 5^{(38.8KB, xlsx)}

Statistical source data.

Source Data Fig. 6^{(2.8MB, xlsx)}

Statistical source data.

Source Data Fig. 7^{(122.8KB, xlsx)}

Statistical source data.

Source Data Fig. 8^{(38.4KB, xlsx)}

Statistical source data.

Source Data Extended Data Fig. 1^{(7.1KB, xlsx)}

Statistical source data.

Source Data Extended Data Fig. 2^{(601.1KB, xlsx)}

Statistical source data.

Source Data Extended Data Fig. 3^{(311.7KB, xlsx)}

Statistical source data.

Source Data Extended Data Fig. 4^{(370.2KB, xlsx)}

Statistical source data.

Source Data Extended Data Fig. 5^{(403.7KB, xlsx)}

Statistical source data.

Source Data Extended Data Fig. 6^{(6.8KB, xlsx)}

Statistical source data.

Source Data Extended Data Fig. 7^{(62.3KB, xlsx)}

Statistical source data.

Source Data Extended Data Fig. 8^{(39.1KB, xlsx)}

Statistical source data.

Acknowledgements

The data available in the AD Knowledge Portal would not be possible without the participation of research volunteers and the contribution of data by collaborating researchers. We thank the UCSD Shiley-Marcos Alzheimer’s Disease Research Center for providing tissues for the proteomics studies; we acknowledge the contributions of the Sanders-Brown Center on Aging at the University of Kentucky for providing sequencing data from patient samples. We further acknowledge Duffy et al.⁵ for the publicly available ribosome profiling dataset (dbGaP accession no. phs002489.v1.p1) used for translational re-analysis in this study.

Extended data

Author contributions

B.M. conceived the study, constructed the putative ORF database, performed the smORF genomic classification, ran the ShortStop classifier and PROSIT spectral grading pipeline, performed the TMT proteomics statistical analyses, the bulk RNA-seq differential expression and the smORF-parent gene coexpression analyses and the sn-RNA-seq integration and RBP motif analyses, generated the cell lines, including the micro-MKKS63 CRISPR KO lines, carried out the overexpression experiments and performed the IF and F-actin quantification studies. C.L. conducted the ESPRESSO long-read transcriptome assembly. E.V.S. performed the RiboCode and Rp3 analyses. J.M.V. carried out the proteomics experiments in the HMC3 cells. V.J.P. assisted with the CRISPR sgRNA design. S.G. and C.C.O. provided technical assistance with confocal microscopy imaging and aided interpretation. A.R. assisted with the Seahorse assays. J.K.D. performed the MS. D.A.B. facilitated data retrieval. A.S. obtained funding and approvals for use of the postmortem cohort data and supervised the project. B.M. and A.S. wrote the manuscript with input from all authors.

Peer review

Peer review information

Nature Aging thanks Julian Langer and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. Peer reviewer reports are available.

Funding

A.S. acknowledges financial support from the National Institutes of Health (NIH) (grant nos. P30CA014195 and R01GM102491) and from the Clayton Medical Research Foundation. E.V.S. acknowledges the NOMIS Center and the NOMIS Foundation for their support. Work involving the UCSD Shiley-Marcos Alzheimer’s Disease Research Center tissues was supported by the National Institute on Aging (NIA) (grant no. AG062429). The MSBB study was supported by NIA grant nos. U01AG046170, RF1AG057440, RF1AG054014 and R01AG057907, with brain tissue collection and characterization supported by NIH HHSN271201300031C. ROSMAP data collection was supported by NIA grant nos. P30AG10161 (ROS), R01AG15819 (ROSMAP; genomics and RNA-seq), R01AG17917 (MAP), R01AG30146, R01AG36042 (5hC methylation, assay for transposase-accessible chromatin using sequencing), RC2AG036547 (H3K9Ac), R01AG36836 (RNA-seq), R01AG48015 (monocyte RNA-seq), RF1AG57473 (sc-RNA-seq), U01AG32984 (genomic and whole-exome sequencing), U01AG46152 (ROSMAP AMP-AD, targeted proteomics), U01AG46161 (TMT proteomics), U01AG61356 (whole-genome sequencing, targeted proteomics, ROSMAP AMP-AD), P30AG072975, the Illinois Department of Public Health (ROSMAP) and the Translational Genomics Research Institute (genomic). The TMT proteomics data were provided through the Accelerating Medicines Partnership (U01AG046161 and U01AG061357); associated data collection was supported by NIA grant nos. P30AG10161, R01AG15819, R01AG17917, R01AG30146, R01AG36836, U01AG32984 and U01AG46152, the Illinois Department of Public Health and the Translational Genomics Research Institute.

Data availability

The data used in this study were in part obtained through the AD Knowledge Portal (https://doi.org/10.7303/9618238). Transcriptomics and associative data were provided by the Rush Alzheimer’s Disease Center, Rush University Medical Center; additional phenotypic data can be requested from www.radc.rush.edu. Transcriptomics data were also generated from postmortem brain tissue collected through the Mount Sinai VA Medical Center Brain Bank and provided by E. Schadt from Mount Sinai School of Medicine. The TMT proteomics data were based on samples provided by the Rush Alzheimer’s Disease Center, Rush University Medical Center. The raw and processed data used in this study are available through multiple public repositories. Data from the AMP-AD Knowledge Portal (https://adknowledgeportal.synapse.org), including the raw and processed datasets from ROSMAP (Synapse ID: syn3219045) and MSBB (Synapse ID: syn3159438), and methods developed by the AMP-AD Target Discovery Program (supported by the NIA), are shared without publication embargo and are available for secondary analysis; access requires registration and adherence to data use and attribution guidelines (https://adknowledgeportal.synapse.org/#/DataAccess/Instructions). ROSMAP resources can also be requested directly through the Rush Alzheimer’s Disease Center (www.radc.rush.edu). All raw long-read sequencing data generated from 12 postmortem human brain samples are available through Synapse (ID: syn52047893) under restricted access, requiring a free Synapse account. The Ribo-seq data from Duffy et al.⁵ are available through dbGaP under accession no. phs002489.v1.p1. The UniProtKB human proteome database (Swiss-Prot and TrEMBL reference sequences, UP000005640, downloaded on 20 January 2025) used for the spectral searches is available at https://ftp.uniprot.org/pub/databases/uniprot/previous_releases/. Raw MS data for the HMC3 studies and the UCSD DIA-MS cohort have been deposited at the PRIDE Archive under accession no. PXD071500.

Code availability

Code is available on GitHub at https://github.com/brendan-miller-salk/brain-microprotein-atlas. The repository includes analysis scripts, custom pipelines, results, BED/GTF/FASTA files and tools for MP annotation and visualization. The code and data can be further visualized and browsed at https://huggingface.co/spaces/brmiller/brain-microprotein-atlas-app.

Competing interests

B.M., E.V.S. and A.S. have filed a patent related to ShortStop. A.S. is a shareholder and cofounder of Exo Therapeutics. The other authors declare no competing interests.

Footnotes

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Contributor Information

Brendan Miller, Email: brmiller@salk.edu.

Alan Saghatelian, Email: asaghatelian@salk.edu.

Extended data

is available for this paper at https://doi.org/10.1038/s43587-026-01207-x.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s43587-026-01207-x.

References

1.Prensner, J. R. et al. What can Ribo-Seq, immunopeptidomics, and proteomics tell us about the noncanonical proteome? Mol. Cell Proteomics22, 100631 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
2.Chothani, S., Ho, L., Schafer, S. & Rackham, O. Discovering microproteins: making the most of ribosome profiling data. RNA Biol.20, 943–954 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
3.Hassel, K. R., Brito-Estrada, O. & Makarewich, C. A. Microproteins: overlooked regulators of physiology and disease. iScience26, 106781 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
4.Bateman, A. et al. UniProt: the universal protein knowledgebase in 2025. Nucleic Acids Res.53, D609–D617 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
5.Duffy, E. E. et al. Developmental dynamics of RNA translation in the human brain. Nat. Neurosci.25, 1353–1365 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
6.Deutsch, E. W. et al. Expanding the human proteome with microproteins and peptideins. Nature654, 813–825 (2026). [DOI] [PMC free article] [PubMed] [Google Scholar]
7.Mudge, J. M. et al. Standardized annotation of translated open reading frames. Nat. Biotechnol.40, 994–999 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
8.Martinez, T. F. et al. Accurate annotation of human protein-coding small open reading frames. Nat. Chem. Biol.16, 458–468 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
9.Ma, J. et al. Improved identification and analysis of small open reading frame encoded polypeptides. Anal. Chem.88, 3967–3975 (2016). [DOI] [PMC free article] [PubMed] [Google Scholar]
10.Saghatelian, A. & Couso, J. P. Discovery and characterization of smORF-encoded bioactive polypeptides. Nat. Chem. Biol.11, 909–916 (2015). [DOI] [PMC free article] [PubMed] [Google Scholar]
11.Slavoff, S. A. et al. Peptidomic discovery of short open reading frame-encoded peptides in human cells. Nat. Chem. Biol.9, 59–64 (2013). [DOI] [PMC free article] [PubMed] [Google Scholar]
12.Miller, B. et al. ShortStop: a machine learning framework for microprotein discovery. BMC Methods2, 16 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
13.Camargo, A. P., Sourkov, V., Pereira, G. A. G. & Carazzolle, M. F. RNAsamba: neural network-based assessment of the protein-coding potential of RNA sequences. NAR Genom. Bioinform.2, lqz024 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
14.Zhu, M. & Gribskov, M. MiPepid: MicroPeptide identification tool using machine learning. BMC Bioinformatics20, 559 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
15.Du, Z. et al. Microproteins in human physiology and pathology. Biochemistry65, 812–834 (2026). [DOI] [PMC free article] [PubMed] [Google Scholar]
16.Aguzzoli Heberle, B. et al. Mapping medically relevant RNA isoform diversity in the aged human frontal cortex with deep long-read RNA-seq. Nat. Biotechnol.43, 635–646 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
17.Ping, L. et al. Global quantitative analysis of the human brain proteome and phosphoproteome in Alzheimer’s disease. Sci. Data7, 315 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
18.Aggarwal, S., Raj, A., Kumar, D., Dash, D. & Yadav, A. K. False discovery rate: the Achilles’ heel of proteogenomics. Brief. Bioinform.23, bbac163 (2022). [DOI] [PubMed] [Google Scholar]
19.Gabriel, W. et al. Prosit-TMT: deep learning boosts identification of TMT-labeled peptides. Anal. Chem.94, 7181–7190 (2022). [DOI] [PubMed] [Google Scholar]
20.Wacholder, A. et al. Community benchmarking and evaluation of human unannotated microprotein detection by mass spectrometry based proteomics. Nat. Commun.17, 1241 (2026). [DOI] [PMC free article] [PubMed] [Google Scholar]
21.Gupta, N. & Pevzner, P. A. False discovery rates of protein identifications: a strike against the two-peptide rule. J. Proteome Res.8, 4173–4181 (2009). [DOI] [PMC free article] [PubMed] [Google Scholar]
22.Sandmann, C.-L. et al. Evolutionary origins and interactomes of human, young microproteins and small peptides translated from short open reading frames. Mol. Cell83, 994–1011 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
23.Johnson, E. C. B. et al. Large-scale deep multi-layer analysis of Alzheimer’s disease brain reveals strong proteomic disease-related changes not observed at the RNA level. Nat. Neurosci.25, 213–225 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
24.Mathys, H. et al. Single-cell atlas reveals correlates of high cognitive function, dementia, and resilience to Alzheimer’s disease pathology. Cell186, 4365–4385 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
25.Coleman, C. et al. Multi-omic atlas of the parahippocampal gyrus in Alzheimer’s disease. Sci. Data10, 602 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
26.Chang, K. et al. Loss of TDP-43 splicing repression occurs early in the aging population and is associated with Alzheimer’s disease neuropathologic changes and cognitive decline. Acta Neuropathol.147, 4 (2023). [DOI] [PubMed] [Google Scholar]
27.Raj, T. et al. Integrative transcriptome analyses of the aging brain implicate altered splicing in Alzheimer’s disease susceptibility. Nat. Genet.50, 1584–1592 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]
28.Dyer, S. C. et al. Ensembl 2025. Nucleic Acids Res.53, D948–D957 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
29.Stone, D. L. et al. Genetic and physical mapping of the McKusick–Kaufman syndrome. Hum. Mol. Genet.7, 475–481 (1998). [DOI] [PubMed] [Google Scholar]
30.Yu, F. et al. Pepcentric enables fast repository-scale proteogenomics searches. Preprint at bioRxiv 10.1101/2025.02.24.639867 (2025). [DOI]
31.Garland, E. F. et al. The microglial translocator protein (TSPO) in Alzheimer’s disease reflects a phagocytic phenotype. Acta Neuropathol.148, 62 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
32.Verma, A. et al. Targeting the overexpressed mitochondrial protein VDAC1 in a mouse model of Alzheimer’s disease protects against mitochondrial dysfunction and mitigates brain pathology. Transl. Neurodegener.11, 58 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
33.Bezawork-Geleta, A. et al. Proximity proteomics reveals a mechanism of fatty acid transfer at lipid droplet-mitochondria- endoplasmic reticulum contact sites. Nat. Commun.16, 2135 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
34.International Human Genome Sequencing Consortium Finishing the euchromatic sequence of the human genome. Nature431, 931–945 (2004). [DOI] [PubMed] [Google Scholar]
35.Pink, R. C. et al. Pseudogenes: pseudo-functional or key regulators in health and disease? RNA17, 792–798 (2011). [DOI] [PMC free article] [PubMed] [Google Scholar]
36.Dubois, M.-L. et al. UBB pseudogene 4 encodes functional ubiquitin variants. Nat. Commun.11, 1306 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
37.Sun, M. et al. Systematic functional interrogation of human pseudogenes using CRISPRi. Genome Biol.22, 240 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
38.Tong, G., Hah, N. & Martinez, T. F. Comparison of software packages for detecting unannotated translated small open reading frames by Ribo-seq. Brief. Bioinform.25, bbae268 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
39.Vieira de Souza, E. et al. Rp3: ribosome profiling-assisted proteogenomics improves coverage and confidence during microprotein discovery. Nat. Commun.15, 6839 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
40.Estades Ayuso, V. et al. TDP-43-regulated cryptic RNAs accumulate in Alzheimer’s disease brains. Mol. Neurodegener.18, 57 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
41.Su, H., Katz, S. G. & Slavoff, S. A. Nested small open reading frames are translated from alternative transcripts. Preprint at bioRxiv 10.1101/2024.10.22.619581 (2025). [DOI]
42.Weber, R. et al. Monitoring the 5′UTR landscape reveals isoform switches to drive translational efficiencies in cancer. Oncogene42, 638–650 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
43.Devaux, A. et al. Identification of microprotein-coding intronic polyadenylation isoforms and function in genotoxic anticancer drug response. Genome Biol.26, 366 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
44.Haberman, N. et al. Widespread 3′UTR capped RNAs derive from G-rich regions in proximity to AGO2 binding sites. BMC Biol.22, 254 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
45.Malka, Y. et al. Post-transcriptional 3′-UTR cleavage of mRNA transcripts generates thousands of stable uncapped autonomous RNA fragments. Nat. Commun.8, 2029 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]
46.Ruiz Cuevas, M. V. et al. Most non-canonical proteins uniquely populate the proteome or immunopeptidome. Cell Rep.34, 108815 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
47.Russell, P. J. et al. Translation reinitiation after uORFs does not fully protect mRNAs from nonsense-mediated decay. RNA29, 735–744 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
48.Zhang, H. et al. Determinants of genome-wide distribution and evolution of uORFs in eukaryotes. Nat. Commun.12, 1076 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
49.Akimoto, C. et al. Translational repression of the McKusick–Kaufman syndrome transcript by unique upstream open reading frames encoding mitochondrial proteins with alternative polyadenylation sites. Biochim. Biophys. Acta1830, 2728–2738 (2013). [DOI] [PubMed] [Google Scholar]
50.Chng, S. C., Ho, L., Tian, J. & Reversade, B. ELABELA: a hormone essential for heart development signals via the apelin receptor. Dev. Cell27, 672–680 (2013). [DOI] [PubMed] [Google Scholar]
51.Rocha, A. L. et al. Abnormal mitochondrial structure and function in brown adipose tissue of SLC35A4-MP knockout mice. Sci. Adv.11, eads7381 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]
52.Bennett, D. A., Schneider, J. A., Bienias, J. L., Evans, D. A. & Wilson, R. S. Mild cognitive impairment is related to Alzheimer disease pathology and cerebral infarctions. Neurology64, 834–841 (2005). [DOI] [PubMed] [Google Scholar]
53.Bennett, D. A. et al. Neuropathology of older persons without cognitive impairment from two community-based studies. Neurology66, 1837–1844 (2006). [DOI] [PubMed] [Google Scholar]
54.Bennett, D. A. et al. Natural history of mild cognitive impairment in older persons. Neurology59, 198–205 (2002). [DOI] [PubMed] [Google Scholar]
55.Bennett, D. A. et al. Decision rules guiding the clinical diagnosis of Alzheimer’s disease in two community-based cohort studies compared to standard practice in a clinic-based cohort study. Neuroepidemiology27, 169–176 (2006). [DOI] [PubMed] [Google Scholar]
56.Dammer, E. B., Seyfried, N. T. & Johnson, E. C. B. Batch correction and harmonization of -omics datasets with a tunable median polish of ratio. Front. Syst. Biol.3, 1092341 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
57.Stemmer, M. et al. CCTop: an intuitive, flexible and reliable CRISPR/Cas9 target prediction tool. PLOS ONE10, e0124633 (2015). [DOI] [PMC free article] [PubMed] [Google Scholar]
58.Doench, J. G. et al. Optimized sgRNA design to maximize activity and minimize off-target effects of CRISPR–Cas9. Nat. Biotechnol.34, 184–191 (2016). [DOI] [PMC free article] [PubMed] [Google Scholar]
59.Brinkman, E. K. et al. Easy quantitative assessment of genome editing by sequence trace decomposition. Nucleic Acids Res.42, e168 (2014). [DOI] [PMC free article] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

Supplementary Information^{(316.6KB, pdf)}

1. All data and code generated in this study are publicly available across four repositories. 2. Source data for all figures are provided.

Reporting Summary^{(3MB, pdf)}

Peer Review file^{(1.2MB, pdf)}

Supplementary Tables 1–18.^{(27.9MB, xlsx)}

Source Data Fig. 1^{(1.4MB, xlsx)}

Statistical source data.

Source Data Fig. 2^{(551.9KB, xlsx)}

Statistical source data.

Source Data Fig. 3^{(409KB, xlsx)}

Statistical source data.

Source Data Fig. 4^{(747.2KB, xlsx)}

Statistical source data.

Source Data Fig. 5^{(38.8KB, xlsx)}

Statistical source data.

Source Data Fig. 6^{(2.8MB, xlsx)}

Statistical source data.

Source Data Fig. 7^{(122.8KB, xlsx)}

Statistical source data.

Source Data Fig. 8^{(38.4KB, xlsx)}

Statistical source data.

Source Data Extended Data Fig. 1^{(7.1KB, xlsx)}

Statistical source data.

Source Data Extended Data Fig. 2^{(601.1KB, xlsx)}

Statistical source data.

Source Data Extended Data Fig. 3^{(311.7KB, xlsx)}

Statistical source data.

Source Data Extended Data Fig. 4^{(370.2KB, xlsx)}

Statistical source data.

Source Data Extended Data Fig. 5^{(403.7KB, xlsx)}

Statistical source data.

Source Data Extended Data Fig. 6^{(6.8KB, xlsx)}

Statistical source data.

Source Data Extended Data Fig. 7^{(62.3KB, xlsx)}

Statistical source data.

Source Data Extended Data Fig. 8^{(39.1KB, xlsx)}

Statistical source data.

Data Availability Statement

[CR1] 1.Prensner, J. R. et al. What can Ribo-Seq, immunopeptidomics, and proteomics tell us about the noncanonical proteome? Mol. Cell Proteomics22, 100631 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR2] 2.Chothani, S., Ho, L., Schafer, S. & Rackham, O. Discovering microproteins: making the most of ribosome profiling data. RNA Biol.20, 943–954 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR3] 3.Hassel, K. R., Brito-Estrada, O. & Makarewich, C. A. Microproteins: overlooked regulators of physiology and disease. iScience26, 106781 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR4] 4.Bateman, A. et al. UniProt: the universal protein knowledgebase in 2025. Nucleic Acids Res.53, D609–D617 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR5] 5.Duffy, E. E. et al. Developmental dynamics of RNA translation in the human brain. Nat. Neurosci.25, 1353–1365 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR6] 6.Deutsch, E. W. et al. Expanding the human proteome with microproteins and peptideins. Nature654, 813–825 (2026). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR7] 7.Mudge, J. M. et al. Standardized annotation of translated open reading frames. Nat. Biotechnol.40, 994–999 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR8] 8.Martinez, T. F. et al. Accurate annotation of human protein-coding small open reading frames. Nat. Chem. Biol.16, 458–468 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR9] 9.Ma, J. et al. Improved identification and analysis of small open reading frame encoded polypeptides. Anal. Chem.88, 3967–3975 (2016). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR10] 10.Saghatelian, A. & Couso, J. P. Discovery and characterization of smORF-encoded bioactive polypeptides. Nat. Chem. Biol.11, 909–916 (2015). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR11] 11.Slavoff, S. A. et al. Peptidomic discovery of short open reading frame-encoded peptides in human cells. Nat. Chem. Biol.9, 59–64 (2013). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR12] 12.Miller, B. et al. ShortStop: a machine learning framework for microprotein discovery. BMC Methods2, 16 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR13] 13.Camargo, A. P., Sourkov, V., Pereira, G. A. G. & Carazzolle, M. F. RNAsamba: neural network-based assessment of the protein-coding potential of RNA sequences. NAR Genom. Bioinform.2, lqz024 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR14] 14.Zhu, M. & Gribskov, M. MiPepid: MicroPeptide identification tool using machine learning. BMC Bioinformatics20, 559 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR15] 15.Du, Z. et al. Microproteins in human physiology and pathology. Biochemistry65, 812–834 (2026). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR16] 16.Aguzzoli Heberle, B. et al. Mapping medically relevant RNA isoform diversity in the aged human frontal cortex with deep long-read RNA-seq. Nat. Biotechnol.43, 635–646 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR17] 17.Ping, L. et al. Global quantitative analysis of the human brain proteome and phosphoproteome in Alzheimer’s disease. Sci. Data7, 315 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR18] 18.Aggarwal, S., Raj, A., Kumar, D., Dash, D. & Yadav, A. K. False discovery rate: the Achilles’ heel of proteogenomics. Brief. Bioinform.23, bbac163 (2022). [DOI] [PubMed] [Google Scholar]

[CR19] 19.Gabriel, W. et al. Prosit-TMT: deep learning boosts identification of TMT-labeled peptides. Anal. Chem.94, 7181–7190 (2022). [DOI] [PubMed] [Google Scholar]

[CR20] 20.Wacholder, A. et al. Community benchmarking and evaluation of human unannotated microprotein detection by mass spectrometry based proteomics. Nat. Commun.17, 1241 (2026). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR21] 21.Gupta, N. & Pevzner, P. A. False discovery rates of protein identifications: a strike against the two-peptide rule. J. Proteome Res.8, 4173–4181 (2009). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR22] 22.Sandmann, C.-L. et al. Evolutionary origins and interactomes of human, young microproteins and small peptides translated from short open reading frames. Mol. Cell83, 994–1011 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR23] 23.Johnson, E. C. B. et al. Large-scale deep multi-layer analysis of Alzheimer’s disease brain reveals strong proteomic disease-related changes not observed at the RNA level. Nat. Neurosci.25, 213–225 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR24] 24.Mathys, H. et al. Single-cell atlas reveals correlates of high cognitive function, dementia, and resilience to Alzheimer’s disease pathology. Cell186, 4365–4385 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR25] 25.Coleman, C. et al. Multi-omic atlas of the parahippocampal gyrus in Alzheimer’s disease. Sci. Data10, 602 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR26] 26.Chang, K. et al. Loss of TDP-43 splicing repression occurs early in the aging population and is associated with Alzheimer’s disease neuropathologic changes and cognitive decline. Acta Neuropathol.147, 4 (2023). [DOI] [PubMed] [Google Scholar]

[CR27] 27.Raj, T. et al. Integrative transcriptome analyses of the aging brain implicate altered splicing in Alzheimer’s disease susceptibility. Nat. Genet.50, 1584–1592 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR28] 28.Dyer, S. C. et al. Ensembl 2025. Nucleic Acids Res.53, D948–D957 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR29] 29.Stone, D. L. et al. Genetic and physical mapping of the McKusick–Kaufman syndrome. Hum. Mol. Genet.7, 475–481 (1998). [DOI] [PubMed] [Google Scholar]

[CR30] 30.Yu, F. et al. Pepcentric enables fast repository-scale proteogenomics searches. Preprint at bioRxiv 10.1101/2025.02.24.639867 (2025). [DOI]

[CR31] 31.Garland, E. F. et al. The microglial translocator protein (TSPO) in Alzheimer’s disease reflects a phagocytic phenotype. Acta Neuropathol.148, 62 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR32] 32.Verma, A. et al. Targeting the overexpressed mitochondrial protein VDAC1 in a mouse model of Alzheimer’s disease protects against mitochondrial dysfunction and mitigates brain pathology. Transl. Neurodegener.11, 58 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR33] 33.Bezawork-Geleta, A. et al. Proximity proteomics reveals a mechanism of fatty acid transfer at lipid droplet-mitochondria- endoplasmic reticulum contact sites. Nat. Commun.16, 2135 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR34] 34.International Human Genome Sequencing Consortium Finishing the euchromatic sequence of the human genome. Nature431, 931–945 (2004). [DOI] [PubMed] [Google Scholar]

[CR35] 35.Pink, R. C. et al. Pseudogenes: pseudo-functional or key regulators in health and disease? RNA17, 792–798 (2011). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR36] 36.Dubois, M.-L. et al. UBB pseudogene 4 encodes functional ubiquitin variants. Nat. Commun.11, 1306 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR37] 37.Sun, M. et al. Systematic functional interrogation of human pseudogenes using CRISPRi. Genome Biol.22, 240 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR38] 38.Tong, G., Hah, N. & Martinez, T. F. Comparison of software packages for detecting unannotated translated small open reading frames by Ribo-seq. Brief. Bioinform.25, bbae268 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR39] 39.Vieira de Souza, E. et al. Rp3: ribosome profiling-assisted proteogenomics improves coverage and confidence during microprotein discovery. Nat. Commun.15, 6839 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR40] 40.Estades Ayuso, V. et al. TDP-43-regulated cryptic RNAs accumulate in Alzheimer’s disease brains. Mol. Neurodegener.18, 57 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR41] 41.Su, H., Katz, S. G. & Slavoff, S. A. Nested small open reading frames are translated from alternative transcripts. Preprint at bioRxiv 10.1101/2024.10.22.619581 (2025). [DOI]

[CR42] 42.Weber, R. et al. Monitoring the 5′UTR landscape reveals isoform switches to drive translational efficiencies in cancer. Oncogene42, 638–650 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR43] 43.Devaux, A. et al. Identification of microprotein-coding intronic polyadenylation isoforms and function in genotoxic anticancer drug response. Genome Biol.26, 366 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR44] 44.Haberman, N. et al. Widespread 3′UTR capped RNAs derive from G-rich regions in proximity to AGO2 binding sites. BMC Biol.22, 254 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR45] 45.Malka, Y. et al. Post-transcriptional 3′-UTR cleavage of mRNA transcripts generates thousands of stable uncapped autonomous RNA fragments. Nat. Commun.8, 2029 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR46] 46.Ruiz Cuevas, M. V. et al. Most non-canonical proteins uniquely populate the proteome or immunopeptidome. Cell Rep.34, 108815 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR47] 47.Russell, P. J. et al. Translation reinitiation after uORFs does not fully protect mRNAs from nonsense-mediated decay. RNA29, 735–744 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR48] 48.Zhang, H. et al. Determinants of genome-wide distribution and evolution of uORFs in eukaryotes. Nat. Commun.12, 1076 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR49] 49.Akimoto, C. et al. Translational repression of the McKusick–Kaufman syndrome transcript by unique upstream open reading frames encoding mitochondrial proteins with alternative polyadenylation sites. Biochim. Biophys. Acta1830, 2728–2738 (2013). [DOI] [PubMed] [Google Scholar]

[CR50] 50.Chng, S. C., Ho, L., Tian, J. & Reversade, B. ELABELA: a hormone essential for heart development signals via the apelin receptor. Dev. Cell27, 672–680 (2013). [DOI] [PubMed] [Google Scholar]

[CR51] 51.Rocha, A. L. et al. Abnormal mitochondrial structure and function in brown adipose tissue of SLC35A4-MP knockout mice. Sci. Adv.11, eads7381 (2025). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR52] 52.Bennett, D. A., Schneider, J. A., Bienias, J. L., Evans, D. A. & Wilson, R. S. Mild cognitive impairment is related to Alzheimer disease pathology and cerebral infarctions. Neurology64, 834–841 (2005). [DOI] [PubMed] [Google Scholar]

[CR53] 53.Bennett, D. A. et al. Neuropathology of older persons without cognitive impairment from two community-based studies. Neurology66, 1837–1844 (2006). [DOI] [PubMed] [Google Scholar]

[CR54] 54.Bennett, D. A. et al. Natural history of mild cognitive impairment in older persons. Neurology59, 198–205 (2002). [DOI] [PubMed] [Google Scholar]

[CR55] 55.Bennett, D. A. et al. Decision rules guiding the clinical diagnosis of Alzheimer’s disease in two community-based cohort studies compared to standard practice in a clinic-based cohort study. Neuroepidemiology27, 169–176 (2006). [DOI] [PubMed] [Google Scholar]

[CR56] 56.Dammer, E. B., Seyfried, N. T. & Johnson, E. C. B. Batch correction and harmonization of -omics datasets with a tunable median polish of ratio. Front. Syst. Biol.3, 1092341 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR57] 57.Stemmer, M. et al. CCTop: an intuitive, flexible and reliable CRISPR/Cas9 target prediction tool. PLOS ONE10, e0124633 (2015). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR58] 58.Doench, J. G. et al. Optimized sgRNA design to maximize activity and minimize off-target effects of CRISPR–Cas9. Nat. Biotechnol.34, 184–191 (2016). [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR59] 59.Brinkman, E. K. et al. Easy quantitative assessment of genome editing by sequence trace decomposition. Nucleic Acids Res.42, e168 (2014). [DOI] [PMC free article] [PubMed] [Google Scholar]

PERMALINK

A microprotein atlas of the human frontal cortex in Alzheimer’s disease

Brendan Miller

Eduardo Vieira de Souza

Calvin Lau

Joan M Vaughan

Victor J Pai

Servando Giraldez

Andréa Rocha

Jolene K Diedrich

Clodagh C O’Shea

David A Bennett

Alan Saghatelian

Abstract

Main

Results

Unique MS evidence of MPs in aged human brains

Fig. 1. MS evidence of MPs in the aged human frontal cortex.

Integrating Ribo-seq, MS and PROSIT provides an evidence framework for confidently identifying brain-expressed MPs

Fig. 2. MP MS grades derived from PROSIT SA and fragment-ion coverage.

Extended Data Fig. 1. Manual blind grading of predicted spectral quality correlates with PROSIT-based scores.

Extended Data Fig. 2. PhyloCSF scores of Reviewed (Swiss-Prot) versus Unreviewed (TrEMBL + Unannotated) microproteins.

Differential MP abundance in AD using TMT-MS quantification

Fig. 3. Differential abundance of MPs in AD using TMT quantitation.

Extended Data Fig. 3. Batch correction of ROSMAP TMT-MS data across two rounds.

Extended Data Fig. 4. Volcano plot of the UniProt/Swiss-Prot reviewed protein sequences in sym-AD versus non-AD brain samples from ROSMAP TMT-MS.

A subset of smORFs diverges in expression from the main ORF of the encoding gene

Fig. 4. Differential RNA expression of MPs in AD and correlation with their corresponding gene’s mORF.

Posttranscriptional splicing motif enrichment for genes encoding MPs

Extended Data Fig. 5. Promoter motif enrichment analyses comparing smORFs genomic coordinates vs. mORF genomic coordinates.

Fig. 5. RBP motif enrichment of genes encoding MPs.

At the MKKS locus, an MP is the predominant translation product, not the mORF, and coexpresses with a mitochondrial gene network

Fig. 6. Micro-MKKS63 is the dominant MKKS translation product and coexpresses with a mitochondrial network.

Extended Data Fig. 6. Micro-MKKS63 is the dominant MKKS-derived translation product across public proteomic datasets.

Micro-MKKS63 localizes to the mitochondria and CRISPR–Cas9-mediated KO of its smORF lowers microglial oxidative phosphorylation

Fig. 7. Micro-MKKS63 localizes to mitochondria and Cas9-mediated KO of its encoding smORF disrupts microglial oxidative phosphorylation.

Ribosome footprints and MS coverage provide orthogonal evidence for brain-expressed pseudogenes

Extended Data Fig. 7. Actin-related pseudogene transcripts are nominally lower in AD brains.

Fig. 8. Orthogonal evidence of actin-related pseudogene MP expression.

Extended Data Fig. 8. Pseudogene-encoded microproteins are sequence-unique relative to ACTB missense variation.

Discussion

Technical considerations and limitations

Broader implications

Methods

Human postmortem brain cohorts

Long-read transcriptome assembly and putative ORF database construction

TMT-based proteomics and protein identification analysis

smORF classification according to genomic context and isoform homology

Ribosome profiling re-analysis using Rp3 to assess translation of MS-identified MPs

Machine-learning-based classification of translated smORFs using ShortStop

DIA MS of human DLPFC tissue

Spectral validation of MP PSMs using PROSIT

Normalization, covariate adjustment and differential expression analysis of TMT proteomics data

Bulk RNA-seq profiling and differential expression analysis

snRNA-seq integration

Transcriptional regulatory and RBP motif analyses

Differential expression analysis of long-read RNA-seq data

Cell culture, transient transfection and immunocytochemistry

Acid-based extraction and LC–MS/MS analysis of HMC3 microglial cells

Coexpression and GO enrichment analysis

Re-analysis of BioID proximity-labeling MS data

CRISPR–Cas9 editing of micro-MKKS63 and mitochondrial stress testing

Statistics and reproducibility

Reporting summary

Supplementary information

Source data

Acknowledgements

Extended data

Author contributions

Peer review

Peer review information

Funding

Data availability

Code availability

Competing interests

Footnotes

Contributor Information

Extended data

Supplementary information

References