Inferring translational efficiency from transcriptomes improves noncanonical neoantigen prioritization and cancer patient stratification

Abstract

Accurate assessment of protein translation is crucial for understanding disease variant functions, but mRNA-protein discrepancy limits transcriptomics-based clinical oncology. While ribosome profiling directly measures translation, its clinical application is constrained by cost and complexity. Deep learning models like Translatomer infer translation efficiency from RNA-seq, but whether in silico translatomes provide superior clinical utility over standard RNA-seq remains unexplored. Here, we present a multidimensional framework evaluating the translational inference strategy across 15 independent datasets. Inferred translational profiles outperform conventional RNA-seq proxies in recapitulating ribosome occupancy and uncover the “dark proteome” through lncRNA translational potential prediction. We integrate this strategy into a translation-aware neoantigen pipeline, identifying high-confidence noncanonical neoantigens neglected by expression-based filtering. Applying this framework to glioma stratification reveals distinct subtypes and corrects high-risk patient misclassification by expression-based methods, as validated by survival analysis. Our study establishes translational inference as a cost-effective enhancement for precision oncology, refining patient stratification and expanding immunotherapeutic targets.

Introduction

Gene expression is regulated at both the transcriptional and translational levels, yet the discrepancy between mRNA abundance and protein output remains a central challenge in functional genomics^1,2. Ribosome profiling (Ribo-seq) is the gold-standard experimental technique for directly monitoring translation at a transcriptome-wide scale³. However, its widespread application, particularly in clinical settings, is hampered by high costs and technical complexity. This limitation has spurred the development of computational methods to infer translational efficiency from more readily available RNA-seq data.

Subsequent deep learning models have advanced the field, yet many do not fully capture all aspects of translational regulation. For instance, RiboNN⁴ and Riboformer⁵ rely primarily on sequence features, Riboformer additionally requires control ribosome density data, limiting its applicability when such data are unavailable. DeepRibo⁶ combines Ribo‑seq signals with sequence motifs but is designed specifically for prokaryotic ORF annotation. Tools such as RiboTaper⁷, iXnos⁸, and Ribo‑MIMO⁹ leverage Ribo‑seq periodicity or codon usage without using transcript abundance as an input.

Recently, Translatomer was introduced as the first multimodal transformer framework that jointly models RNA-seq and nucleotide sequences for end-to-end translation efficiency prediction¹⁰. This design allows it to potentially capture the regulatory interplay between transcript abundance and cis-regulatory elements that earlier single-modality models could not. Although the technical accuracy of such models has been reported, whether integrating inferred translational data can improve clinical decision‑making beyond standard transcriptomic analysis remains unclear. Most current pipelines for neoantigen prioritization and molecular subtyping rely heavily on RNA expression thresholds, potentially leading to false positives (expressed but not translated) or false negatives (low expression but efficiently translated).

In this study, we conducted a comprehensive, three-dimensional evaluation of the translational inference strategy, using Translatomer as a representative model. We rigorously compared the model’s predictions against the standard RNA-seq baseline across 15 independent cohorts. Our study aims to: (1) quantify the gain in fidelity provided by translational inference across an extensive collection of independent datasets spanning multiple cancers and sample types; (2) evaluate its extensibility by systematically testing its ability to predict translation in long non-coding RNAs (lncRNAs), a capability not explored in the original study; and (3) demonstrate its practical utility by integrating its predictions into two key cancer genomics applications: translation-aware neoantigen prioritization and molecular subtyping (Fig. 1). Our analysis reveals that inferred translatomes capture biological nuances missed by RNA-seq, and that translation-aware analysis refines neoantigen selection and corrects prognosis-relevant molecular subtypes in glioma. This work provides an empirical foundation for adopting translation-inference tools to enhance the resolution of cancer genomics research.

Fig. 1. A systematic framework for evaluating translational inference in oncology.

a The Translatomer framework. RNA-seq data and one-hot-encoded gene sequences are used as inputs to predict ribosome-profiling coverage, capturing the relationships among transcript abundance, sequence context and translation efficiency. b The original Translatomer model is reproduced and then adapted for additional tasks, including benchmarking against conventional RNA-seq-based proxies, prediction in noncoding regions, integration into a translation-aware neoantigen optimization pipeline, and application to molecular subtyping. c Utility evaluation. Adaptability is further evaluated in personalized immunotherapy design through translation-aware neoantigen prioritization and translation-informed molecular classification of glioma.

Results

Inferred translatomes outperform RNA-seq proxies in recapitulating ribosome occupancy

To establish the baseline validity of the translational inference strategy, we first confirmed the reproducibility of the Translatomer model using publicly available checkpoints (Fig. 2a, b and Supplementary Data 1). Building on this, we extended the evaluation to a diverse collection of 15 independent datasets spanning multiple cancer types (e.g., prostate, epithelial, glioblastoma).

Fig. 2. Inferred translational profiles exhibit superior fidelity compared to RNA-seq baselines.

a, b Comparison between the original and reproduced results on the four benchmark datasets across two performance metrics, confirming the baseline integrity of model replication. PCC, Pearson correlation coefficient; MSE, mean squared error. c Model performance relative RNA-seq signal in the internal glioma cohort (CGGA, n = 24). The median, minimum, and maximum SCC values are indicated. d–m Comparative performance of Translatomer and RNA-seq signal across nine independent datasets encompassing distinct biological contexts. Each panel displays the SCC for individual samples, with median, minimum, and maximum values annotated. n, o Impact of RNA-seq preprocessing on model accuracy across all the validation datasets (n = 57). Each dot represents one sample; lines connect paired results between nonnormalized and normalized (RPKM-scaled) data. Comparisons between the two groups were conducted via the two-sided Wilcoxon rank-sum test. The centre line in the box represents the median among the 57 results, excluding outliers; the lower and upper bounds of the box represent the first (Q1) and third (Q3) quartiles, respectively.

We systematically compared Translatomer predictions with conventional RNA-seq-based representations, using experimentally measured Ribo-seq data as the ground truth. These datasets included an internal glioma cohort (n = 24) and nine public datasets, deliberately selected to test performance across different sample types (primary tissues vs. cell lines), disease states (normal vs. tumor), cancer origins (brain, ovarian, pancreatic, prostate), and temporal dynamics. Notably, these cancer types are often characterized by an immunosuppressive tumor microenvironment, allowing us to probe the model’s robustness under such challenging biological conditions.

Comprehensive analysis across all datasets revealed that Translatomer predictions showed a significantly stronger correlation with the experimental Ribo-seq ground truth (median SCC = 0.701, range 0.584–0.853; median MSE = 0.087 with low inter-dataset variability) compared to RNA-seq-based estimates (median SCC = 0.324, range 0.163–0.717; median MSE = 1.812). This performance gap was particularly pronounced in immunosuppressive cancers such as glioma and pancreatic cancer, and the advantage was maintained across primary tissues, cell lines, and dynamic time courses (Fig. 2c–m, Supplementary Fig. 1a–k and Supplementary Data 2). In primary tissues, Translatomer achieved median SCCs ranging from 0.664 to 0.767, substantially outperforming RNA‑seq (Fig. 2c–e). In cancer cell lines, the model performed consistently well across glioma (SCC 0.661–0.754), ovarian (0.853), prostate (0.584), and pancreatic (0.725–0.744) lines, again exceeding RNA‑seq in each case (Fig. 2f–k). Temporal stability was also high, with SCCs of 0.789–0.814 in U251 cells and 0.796–0.808 in U343 cells across a 24‑h time course (Fig. 2l, m).

A key finding of this assessment emerged when we investigated the impact of input preprocessing. Contrary to the expectation that normalization would enhance cross-dataset consistency, we observed systematic performance degradation. Across ten independent datasets (57 samples), the median SCC decreased from 0.70 with unnormalized inputs to 0.40 with normalized signals (Δ $\approx$ −0.30; Fig. 2n, o and Supplementary Data 2). This suggests that the pretrained representations of Translatomer are inherently calibrated to the native scale of RNA-seq coverage and that rescaling introduces a mild domain shift that disrupts this alignment, a practical insight for future users.

To further evaluate whether the predicted translational signals correspond to actual protein output, we leveraged a matched multi-omics dataset comprising RNA-seq, Ribo-seq, and mass spectrometry-based proteomics from four HeLa samples (GSE308892)¹¹. Translatomer achieved high predictive accuracy against Ribo-seq ground truth across all samples, with SCC values ranging from 0.851 to 0.855 and consistently low MSE from 0.057 to 0.071 (Supplementary Fig. 2a). Notably, when comparing Translatomer-derived translational activity, defined as the length-normalized sum of predicted Ribo-seq signals across CDS regions, to independently measured protein abundance, we observed significant positive correlations across all four samples (SCC = 0.41–0.42, all P < 0.001; Supplementary Fig. 2b, d, f, h). In contrast, conventional RNA-seq expression (log2TPM) showed substantially lower correlation with protein abundance in the same samples (SCC = 0.19–0.20, all P < 0.001; Supplementary Fig. 2c, e, g, i). These results provide direct evidence that Translatomer predictions capture not only ribosome occupancy but also the downstream protein output.

In summary, Translatomer demonstrates significantly superior and more stable performance in predicting translation efficiency across datasets and biological contexts compared to conventional RNA-seq-based approaches. This advantage is evident not only in correlation metrics but also in challenging scenarios such as immunosuppressive tumors and dynamic time-series data, underscoring its robustness and establishing its utility as a reliable tool for translational genomics.

Comparative evaluation against conventional expression-based proxies

To understand the source of the performance gain, we dissected the discrepancy between predicted translation and observed RNA-seq signal at the single- protein coding gene/open reading frame (ORF) level using the CGGA dataset. We first confirmed the model’s performance at the resolution of individual ORFs in our internal glioma cohort, where it achieved a median SCC of 0.701 (range: 0.604–0.774) and a median MSE of 0.143 (range: 0.079–0.208) (Supplementary Fig. 3 and Supplementary Data 2). To quantify the comparative advantage, we defined “effectively predicted” elements (protein-coding genes or ORFs) as those with a statistically significant correlation to the experimental Ribo-seq ground truth (SCC/PCC > = 0.3, P < 0.05). The ratio of elements effectively predicted by Translatomer versus those predicted by RNA-seq alone was consistently greater than 1 in all the samples analyzed (Fig. 3a, Supplementary Fig. 4a and Supplementary Data 3).

Fig. 3. Systematic benchmarking of the Translatomer against conventional RNA-seq abundance as a translation proxy.

a Ratio of elements effectively predicted by the Translatomer versus the RNA-seq proxy across 24 glioma samples. Each dot represents one sample. Left: protein-coding genes. Right: all ORFs. A ratio >1 (dashed line) indicates the Translatomer’s superior detection of translationally active elements. The PCC-based results are shown in Supplementary Fig. 4a. b Overlap of effectively predicted elements between methods. Radar charts display, for each sample, two ratios for protein-coding genes (left) and ORFs (right): intersection/RNA-seq (blue, fraction of RNA-seq-effective elements captured by Translatomer) and intersection/Translatomer (yellow, fraction of Translatomer’s effective predictions unique to it). The PCC-based results are shown in Supplementary Fig. 4b. c Distribution of each protein-coding gene performance differences. The ridge plot shows the density of the per-gene difference in the SCC (Translatomer-RNA-seq) for each sample (y-axis). The PCC-based results are shown in Supplementary Fig. 4c. d Each point represents a gene-sample pair. The x-axis shows the SCC between the predicted and observed Ribo-seq data; the y-axis shows the PCC. The point color corresponds to the observed Ribo-seq signal (log-scaled, summed signal across the CDS/the number of bins covering the CDS) for that protein-coding gene in that sample, as indicated by the adjacent color bar. Spearman’s correlation and P-values were used to quantify the positive correlation between the observed Ribo-seq signal and predictive accuracy (SCC and PCC). e Characterization of protein-coding genes whose Translatomer underperformed (ΔSCC < 0). Top, GSEA plot showing the enrichment of underperforming protein-coding genes with low translational activity. NES, normalized enrichment score. Bottom, distribution of the length-corrected observed Ribo-seq signals for these protein-coding genes. The pie chart inset shows that over 90% of these protein-coding genes fall within the lowest range of translational activity. PCC-based results and per-sample fGSEA summary are shown in Supplementary Fig. 4d, e. f Radar chart displaying the proportion of ORFs effectively predicted (PCC > = 0.3 and P < 0.05) across eight ORF categories. The SCC-based results are shown in Supplementary Fig. 4f. g–i Translation prediction for six glioma-associated tumor antigens. Scatter plots showing the prediction performance (SCC and PCC) across 24 patients. The quadrants are defined as SCC = 0.3 and PCC = 0.3. The proportion of samples in each quadrant is annotated. The point size reflects the statistical significance (−log10) of the PCC, and the point color represents the SCC significance.

A pivotal finding of this benchmark is the unique biological insight uncovered by the model. While there was a high overlap between the sets, indicating that Translatomer successfully recapitulates known transcription-translation relationships, it also identified a substantial, unique subset of elements with significant translational activity that were entirely missed by the RNA-seq proxy (intersection/Translatomer ratio < 0.5 in over 87.5% of samples; Fig. 3b, Supplementary Fig. 4b and Supplementary Data 3). These results demonstrate that the model provides access to a layer of biology that is inaccessible through transcriptomic analysis alone.

We further dissected the performance by calculating, for each protein-coding gene, the difference in correlation (SCC/PCC) between the two methods. Only a minimal fraction of protein-coding genes showed a marginally better correlation with RNA-seq (Δ < 0; Fig. 3c, Supplementary Fig. 4c and Supplementary Data 3). We hypothesized that this minor performance gap could be attributed to the inherent challenge of predicting translation for protein-coding genes with very low translational activity. Indeed, we observed a strong positive correlation between the prediction accuracy (SCC/PCC) and the protein-coding gene’s actual translation level (Fig. 3d). Protein-coding genes where the model underperformed relative to RNA-seq data were significantly enriched in the lowest range of length-corrected translational activity (over 90% residing in the 0–1.5 range; Fig. 3e and Supplementary Fig. 4d). Gene set enrichment analysis (GSEA) validated this trend (SCC-based: NES = −1.073, P = 9.99e-06; PCC-based: NES = −1.173, P = 9.99e-06). This pattern remained consistent across individual samples, as the vast majority (24/24) presented a negative enrichment score (Supplementary Fig. 4e and Supplementary Data 4).

We next analyzed the performance of Translatomer for different ORF types. We found that the Translatomer performed robustly on both canonical and noncanonical ORFs, with a tendency toward greater accuracy in specific noncanonical categories, such as out-of-frame and 5’ overlap uORFs (Fig. 3f and Supplementary Fig. 4f, g), suggesting a nuanced understanding of translation initiation and elongation. Notably, ORFs where the model underperformed relative to RNA-seq data were significantly enriched in the lowest range of length-corrected translational activity, with over 80% residing in the 0–1.5 range (Supplementary Fig. 5a, b). This observation is consistent with the results obtained from protein-coding genes. To further ensure that this performance is not driven by non-expressed regions, we additionally evaluated model performance in regions with observed translation signals (Ribo-seq > 0), where Translatomer consistently outperformed RNA-based estimates in both AUPRC and enrichment analyses (Supplementary Fig. 8 and Supplementary Fig. 11). Finally, to illustrate its potential for practical application, we evaluated the utility of the Translatomer in predicting the translation of a panel of six established glioma tumor antigens (TAs)-a task in which an accurate translation proxy is clinically relevant. The model successfully predicted the translation efficiency for the vast majority of these TAs across patients (ranging from 15/24 to 21/24 samples across different TAs; Fig. 3g–l). This high success rate highlights the model’s potential as a cost-effective computational tool for the initial screening of patients most likely to express specific TAs, thereby informing selection for targeted therapies.

Expanding the scope of translatomer to noncanonical translation

To evaluate the extensibility of the translatomer beyond its primary training context for protein-coding genes, we next performed a genome-wide assessment of its predictive ability. The genome was partitioned into non-overlapping 65,536 bp bins, and prediction accuracy was quantified in the CGGA_1282 sample by correlating the predicted and observed Ribo-seq signals. The bins were categorized into HIGH, MEDIAN, and LOW groups on the basis of correlation strength and statistical significance (Methods and Supplementary Fig. 6a). Analysis of the genomic composition of effectively predicted bins (HIGH and MEDIAN) revealed that, in addition to coding sequences (CDs), a substantial fraction of well-predicted regions corresponded to noncoding elements, including lncRNAs and untranslated regions (UTRs) (Fig. 4a and Supplementary Fig. 6b). Notably, 91.3% of all effectively predicted bins contained non-coding sequences, suggesting that the learned representations of Translatomer capture translation signals extending beyond annotated coding regions.

Fig. 4. Prediction of translational activity in noncoding regions.

a Distribution of noncoding sequence content within effectively predicted genomic bins. The pie chart shows the proportion of effectively predicted 65,536 bp bins (groups HIGH and MEDIAN, as defined in Methods) categorized by the percentage of their length annotated as noncoding sequences (e.g., lncRNAs, UTRs). The majority of bins contain a substantial fraction of noncoding bases. b–l Generalizability of the Translatomer for predicting lncRNA translational potential across diverse datasets. The performance is shown for the internal glioma cohort (b), primary tissues (c–e), cancer cell lines (f–j), and temporal dynamics (k, l). Each panel displays the SCC and MSE between the predicted and observed Ribo-seq signals for the lncRNAs, with the median, minimum, and maximum values annotated.

We next systematically evaluated the performance of the Translatomer in predicting putative translational signals on lncRNAs, a task not explored in the original study. In the internal glioma dataset, Translatomer achieved high predictive accuracy for lncRNA translational potential, with SCC values ranging from 0.738–0.824 (median = 0.801) and a median MSE of 0.044 (Fig. 4b and Supplementary Data 2). To verify that this capability was not context-specific, we further assessed performance across multiple independent external datasets. In brain tissues, the Translatomer maintained robust accuracy (normal: median SCC = 0.735, median MSE = 0.064; glioma: median SCC = 0.729–0.791, median MSE = 0.042–0.055) (Fig. 4c, d).

The model also demonstrated consistent performance across diverse cancer types, including glioma cell lines (SCC = 0.745–0.797, MSE = 0.093–0.153) (Fig. 4e, f), ovarian cancer (SCC = 0.908, MSE = 0.029) (Fig. 4g), prostate cancer (SCC = 0.666, MSE = 0.125) (Fig. 4h), and pancreatic ductal carcinoma (SCC = 0.794–0.809, MSE = 0.045–0.053) (Fig. 4i, j). Temporal evaluation of glioma cell lines (U251 and U343) across a time series (T0, T1, T24 h) revealed stable predictive accuracy across all time points (SCC = 0.820–0.843, MSE = 0.052–0.065) (Fig. 4k, l), underscoring the reliability of the predictions of the Translatomer for lncRNAs under dynamic conditions.

Collectively, our analyses demonstrate that Translatomer, which was originally trained and benchmarked on protein-coding genes, can be effectively extended to predict translation events in noncoding regions, particularly lncRNAs. The consistent performance across internal and public datasets underscores the broad extensibility of Translatomer and its potential as a valuable computational framework for probing the translational landscape of the noncoding genome.

Within each 65,536 bp genomic bin, over 90% of positions show zero Ribo-seq signals. Given the sparsity of Ribo-seq signals, we further evaluated whether performance metrics such as correlation may be influenced by shared zero values. To address this, we performed analyses separating regions with and without observed translation signals, and complemented correlation-based evaluation with AUPRC and enrichment metrics. In regions without observed translation (Ribo-seq = 0), predicted values remain consistently low, indicating a low false-positive rate, whereas RNA-seq signals show broader distributions, reflecting background noise (Supplementary Fig. 7-9). In regions with active translation (Ribo-seq > 0), Translatomer consistently outperforms RNA-based estimates across multiple thresholds in both AUPRC and enrichment analyses (Supplementary Figs. 10–12). These results demonstrate that model performance is not driven by sparsity, but reflects meaningful recovery of translational signals across protein-coding genes, ORFs, and lncRNAs.

A translation-centric framework for prioritizing neoantigens in glioma

Conventional neoantigen pipelines rely primarily on transcript abundance, which may not accurately reflect protein production. To address this limitation, we developed a benchmark framework to compare a translation-centric prioritization strategy with the conventional expression-based approach. The step-by-step filtering processes of the two pipelines are illustrated in Supplementary Fig. 13, which highlights the distinct criteria applied at each stage and the number of candidates retained. The translation-centric strategy was built around a “translation effect score” (TE), which quantifies mutation-induced alterations in predicted ribosome occupancy within genomic regions harboring potential epitopes (Fig. 5a). Applying a threshold derived from the score distribution (Fig. 5b) to the glioma cohort yielded a refined set of 4021 distinct neoantigens (TE score > 0.1). Benchmarking against the conventional expression-based method demonstrated that the translation filter was substantially more stringent, retaining only 8.01% of the initial candidates prior to HLA-binding assessment compared with the 54.82% retained by the expression-based filter (Fig. 5c and Supplementary Data 5). This initial stringency resulted in a more selective final neoantigen pool, with only 0.4% of candidates surviving HLA-binding filtering for the translation-based method versus 3% for the conventional approach, indicating that translation-based prediction more closely aligns with HLA-mediated selective pressure.

Fig. 5. A translation-centric framework for neoantigen prioritization in glioma.

a Schematic of the translation-aware neoantigen discovery pipeline. The Translatomer model takes RNA-seq data and sequences as inputs to predict Ribo-seq signals (left). A “translation effect score” is computed for each candidate neoantigen by summing the predicted ribosome occupancy within the genomic bins, comparing the mutant and reference states (middle). The candidates are filtered on the basis of this score (right). b Distribution of the “translation effect score” across all candidate epitopes. The curve depicts the number of retained epitopes at different thresholds. c Neoantigen filtering cascade. The proportion of initial candidates remaining after each sequential filter is shown for the translation-based (Translatomer, solid line) and expression-based (Non-Translatomer, dashed line) pipelines. The composition of the candidate pool at each step is broken down by neoantigen type. The labels show the percentage of the original pool remaining. d, e Immunogenic potential of the final neoantigens. d Comparison of predicted immunogenicity scores (−log10) between pipelines. The bottom labels denote the proportion of epitopes meeting this criterion. e Comparison of pMTnet scores, which predict the TCR-binding specificities of HLAps. The bottom text indicates the percentage of neoantigens with a rank < 0.01. f Neoantigen repertoire diversity. Ternary plot illustrating the per-sample composition of canonical, noncanonical and mixed neoantigens (n = 166). g HLA-I binding preference. The heatmap shows the binding proportion of each neoantigen type to HLA-I alleles, revealing distinct preferences across HLA-I alleles. h The upper annotation bars indicate key clinical and molecular features for each sample. The lower heatmap displays the Shannon diversity index of pHLA-TCR interactions per sample. i Scatter plots comparing long-term (>1.5 years) and short-term survivors. The y axis shows the fold-change in the sample-specific neoantigen load normalized to the cohort median; side density plots depict the distribution of values within each group.

To compare the neoantigen prioritization performance between the two pipelines, we evaluated predicted immunogenicity scores as a relative ranking metric. Neoantigens identified by the translation-based approach exhibited a modest yet consistent advantage, showing a greater proportion with homology to known immunogenic epitopes (40.09% vs. 38.80%; Fig. 5d) and more favorable pMTnet scores for pHLA-TCR complex formation (16.24% vs. 15.63%; Fig. 5e), suggesting enrichment for high-quality candidates. A distinct advantage of the Translatomer-based pipeline is its ability to uncover a more diverse neoantigen repertoire. Specifically, it identified a significantly greater proportion of neoantigens derived from noncanonical ORFs (40.2% vs. 27.7%; Supplementary Fig. 14a, b), a class often overlooked by expression-based methods. Moreover, the translation filter displayed notable mutational bias, markedly reducing the retention of SNV-derived candidates (to 4.36% of the initial set) while exerting a more moderate effect on frameshift and inframe variants (Supplementary Fig. 14c). This selectivity likely reflects the sequence-aware architecture of Translatomer, which quantitatively captures mutation-induced perturbations in translational dynamics. The stronger translational impact of frameshift or inframe mutations may thus lead to higher translation effect scores and greater retention. To further assess the biological relevance of the prioritized neoantigens, we compared the predicted epitopes with publicly available immunogenic peptide databases, including IEDB, TESLA¹², and known TCR-binding peptides^12,13 using sequence similarity analysis (BLASTP; mismatches = 0 and percent identity = 100%). About 60% of predicted candidates showed sequence similarity to known immunogenic peptides.

To further validate the biological relevance of our approach, we examined the RNA-Ribo correlation for genes generating neoantigens. Genes shared between both pipelines presented significantly greater RNA-Ribo correlations, which is consistent with a “high-expression, high-translation” phenotype, whereas genes uniquely identified by either method presented weaker correlations (Supplementary Fig. 14d). This finding supports our hypothesis that the translation-based pipeline captures antigens from genes with discordant expression and translation.

We next characterized the properties of the translation-predicted neoantigens. Neoantigens were classified as canonical, noncanonical, or mixed on the basis of the underlying nonsynonymous mutation (Methods). In glioma patients, canonical and mixed neoantigens predominated (Fig. 5f and Supplementary Fig. 15a–e). Interestingly, noncanonical and mixed neoantigens were enriched for HLA-C alleles, whereas canonical neoantigens preferentially bound HLA-B alleles (Fig. 5g). Analysis of T-cell clonal diversity via the Shannon diversity index of predicted pHLA-TCR triplets revealed more uniform neoantigen distributions in patients with high-grade tumors, recurrent disease, and advanced age (Fig. 5h), suggesting a distinct immunogenic topology in aggressive gliomas. Clinically, high neoantigen burden was significantly associated with long-term survival (threshold: 1.5 years; Fig. 5i).

To further assess the robustness of these findings, we performed sensitivity analyses under multiple parameter configurations by varying translation/expression thresholds and HLA-binding affinity cutoffs. The original control settings (TE score > 0.1; gene ≥ 1 TPM, ORF ≥ 0.1 TPM; HLA IC50 < 500 nM) were compared with a relaxed translation filter (TE score > 0.05; gene ≥ 0.5 TPM, ORF ≥ 0.05 TPM), a stringent translation filter (TE score > 0.2; gene ≥ 5 TPM, ORF ≥ 0.5 TPM), a moderately stringent HLA filter (IC50 < 150 nM), and a stringent HLA filter (IC50 < 50 nM). Across all tested settings, the main conclusions remained consistent, the Translatomer-based pipeline consistently yielded a higher proportion of noncanonical neoantigens and higher predicted immunogenicity score compared to the expression-based approach, while maintaining comparable pMTnet-based TCR-binding potential (Supplementary Fig. 16). These results indicate that the observed advantages are robust to parameter selection.

To evaluate the generalizability of the translation-aware neoantigen prioritization pipeline, we applied the framework to an independent glioma cohort¹⁴ with matched RNA-seq and whole-exome sequencing data (PRJNA1051047, n = 100). Consistent with the discovery cohort, the translation-based pipeline retained a higher proportion of noncanonical neoantigens compared to the expression-based approach (Supplementary Fig. 17a). In addition, immunogenicity scores were consistently higher in the Translatomer-based pipeline, with a greater fraction of epitopes exceeding commonly used thresholds (score ≥ 1e-16; Supplementary Fig. 17b), supporting improved prioritization performance.

To provide a global overview of translational activity, we integrated RNA-seq, observed Ribo-seq, and Translatomer-predicted Ribo-seq data across all protein-coding genes (Fig. 6). The resulting genome-wide correlation analysis demonstrated the superior predictive power of the Translatomer in capturing translational activity. Finally, to explore the link between translational regulation and neoantigen generation in glioma, we defined two gene sets: genes with dysregulated translation (tumor vs. normal) and genes with divergent neoantigen production capacities. The substantial overlap between these sets suggested a coordinated yet complex relationship between translational efficiency and immunogenic potential. To further elucidate this relationship, we performed an in-depth analysis of PLOD3, a gene implicated in glioma invasion through collagen remodeling within the tumor microenvironment¹⁵ which was identified here as a high neoantigen burden gene. We detected a C → CG mutation (chr7:101212645) that generates neoantigens with high translation effect scores and strong predicted HLA-I binding affinity. Notably, one of these peptides (LLLDYPPDRVTL) was independently confirmed in the glioma immunopeptidome^16–19. This case supports the mechanistic link between translational dysregulation and neoantigen generation, providing functional validation of the Translatomer as an effective tool for translation-aware neoantigen identification.

Fig. 6. Case-based illustration of translation-informed neoantigen prediction and functional integration in glioma.

Top, Multi-omics profiling of the PLOD3 locus and its properties in glioma. Tracks 1–3 display the observed Ribo-seq, Translatomer-predicted Ribo-seq, and observed RNA-seq signals, respectively, in sample CGGA_P116. Tracks 4–7 show each protein-coding gene correlation coefficients (PCCs and SCCs) between these data tracks, demonstrating the model’s accuracy. Tracks 8–9 show the mean translation efficiency of each gene in glioma versus normal tissue, and Track 10 indicates the neoantigen burden per gene. Genes with a neoantigen burden ≥ 9 are labeled. This integrated view identifies PLOD3 as a gene with high neoantigen burden and dysregulated translation in glioma. Bottom, detailed analysis of a patient-specific neoantigen derived from PLOD3. A zoomed-in view of the genomic region surrounding a C > CG mutation (chr7:101212645) shows the observed and predicted translation signals, the mutation site, and the derived mutant peptides. The translation effect score and HLA-binding affinity were calculated for the reference and mutant 8–12mer peptides. The mutant peptide LLLDYPPDRVTL was independently verified in the glioma immunopeptidome, functionally validating the Translatomer-based pipeline.

Translation-based molecular stratification corrects transcriptomic misclassification in glioma

The observed overlap between translational dysregulation and neoantigen generation suggested a coordinated relationship between these two regulatory layers in glioma. To explore whether the Translatomer-derived features could capture this structure, we leveraged its predictions for unsupervised molecular stratification. We first identified 2506 genes whose translation was significantly altered in glioma tissue compared with normal tissue. From these, we derived two opposing translational programs: a neural translation program (N-TP) enriched for neuronal functions and an oncogenic translation program (Onco-TP) associated with proliferation and oncogenic signaling (Supplementary Fig. 18a–d and Supplementary Data 6). To generate a compact feature set for clustering, we distilled these programs into core signatures—the neural translation signature (NTS) comprising 11 genes with high neoantigen burden and low translation, and the oncogenic translation signature (OTS) comprising 63 genes with low neoantigen burden and high translation.

Unsupervised clustering of glioma patients on the basis of the translational patterns of these 74 signature genes robustly stratified samples into four distinct subtypes (C1–C4) (Fig. 7a and Supplementary Data 7). These subtypes exhibited reciprocal patterns of signature activity: NTS expression was highest in C4 and lowest in C3, whereas OTS expression displayed the opposite trend. Functional characterization revealed distinct biological themes across subtypes (Supplementary Data 8). C1 (proliferative) was enriched for DNA replication, chromosome segregation, and mitotic pathways, indicating a highly proliferative phenotype. C2 (metabolic) was dominated by the mitochondrial electron transport and oxidative phosphorylation pathways, suggesting a hypermetabolic state. C3 (fibrotic/immunosuppressive) showed enrichment in extracellular matrix (ECM) receptor interaction and degradation pathways, which is consistent with a fibrotic, immunosuppressive, and invasive phenotype. C4 (neural) was enriched for neuronal signaling, synaptic transmission, and vesicle exocytosis, reflecting a neural-like differentiation state.

Fig. 7. Translation-based molecular stratification reveals clinically relevant subtypes in glioma.

a Unsupervised clustering of 166 gliomas based on the neural translation signature (NTS) and oncogenic translation signature (OTS) identified four subtypes (C1–C4). Upper: Normalized translation efficiency of the NTS and OTS genes. Lower: NESs of biological pathways specifically enriched in each subtype. The top annotation bars show the distributions of the Verhaak subtypes and clinical features. b Principal component analysis of pathway enrichment scores. Each point represents one signaling axis, including the ten canonical oncogenic pathways from the TCGA and NTS/OTS gene sets. c Kaplan–Meier survival analysis for the four translation-defined glioma subtypes (C1–C4). Survival differences were assessed by the log-rank test (two-sided). d Kaplan–Meier survival analysis between expression- and translation-based subtyping. Both-High (n = 6): patients classified as high-risk by both methods; Both-Low (n = 27): patients classified as low-risk by both methods; Trans-High (n = 21): patients classified as high-risk by translation-based subtyping but as low-risk by expression-based subtyping. e, f Pseudotemporal trajectory analysis of glioma progression. e Cell fate plot colored according to subtype and normal samples. f Points colored according to inferred pseudotemporal order. g Clinical characterization of subtypes. The left panel shows stacked bar plots of detailed clinical categories across subtypes. The right panel displays pie charts summarizing the distribution of key clinical features within each subtype. h Independent validation in the CPTAC proteomic cohort confirmed the poor prognosis of the C3 subtype compared with that of the C4 subtype.

Principal component analysis demonstrated a strong association of the OTS signature with established oncogenic pathways from TCGA, underscoring its protumorigenic nature (Fig. 7b). Importantly, these molecular distinctions translated into significant survival differences. Patients in the C4 subtype exhibited the most favorable outcomes, followed by those in the C2 and C1 subtypes, whereas the C3 subtype was associated with the poorest prognosis (Fig. 7c). UMAP visualization of the translational landscape revealed clear separation of the C3 and C4 subtypes (Supplementary Fig. 19a), and trajectory inference analysis suggested potential progression from a normal state through intermediate subtypes (C4, C2, C1) towards the most aggressive C3 subtype, recapitulating the malignant evolution of glioma (Fig. 7e, f). Consistent with this continuous trajectory, GSVA-derived enrichment scores for NTS and OTS revealed a strong negative correlation, defining a major axis of variation along which glioma samples are continuously distributed (Supplementary Fig. 20a). To assess the robustness of these subtype assignments, we performed a resampling-based co-clustering analysis with 100 iterations of 80% sample subsampling, which demonstrated that the majority of samples exhibited stability scores above 0.5, indicating moderate to high robustness (Supplementary Fig. 20b, c). To assess whether the translation-based subtypes could be recapitulated by existing clinical or molecular biomarkers, we systematically compared the four subtypes across multiple established markers, including tumor mutational burden (TMB), immune checkpoint genes (PDCD1, CD274, CTLA4), immune effector genes (GZMA, PRF1, CD8A), and the antigen presentation-related gene HLA-A. TMB showed distinct but partially overlapping distributions across subtypes, with no single threshold clearly separating all four groups (Supplementary Fig. 19l). Similarly, while the expression levels of immune-related markers such as PDCD1, CTLA4, and GZMA exhibited statistically significant differences across subtypes, substantial overlap in their value distributions was observed, indicating that no individual marker can fully distinguish the four translational subtypes (Supplementary Fig. 19e–k). Our subtypes showed strong concordance with established Verhaak classifications: C1 and C2 aligned with the proneural subtype, C3 with the mesenchymal, and C4 with the neural subtype, although the mapping was not complete (Fig. 7g and Supplementary Data 7). Clinically, C1, C2, and C4 share favorable features including younger age, lower grade, MGMT promoter methylation, and IDH mutation, which collectively distinguished them from the aggressive C3 subtype. Taken together, these results demonstrate that existing biomarkers are insufficient to recapitulate the translation-defined subtype structure, highlighting the added value of translation-based stratification.

To evaluate the added value of translation efficiency over transcript abundance, we performed molecular subtyping on the same patient cohort using the 74 signature genes based on their gene expression profiles. Notably, the expression-based classification failed to reproduce clear prognostic subtypes (Supplementary Fig. 19b, c). The survival curves of C1, C2, and C4 were largely overlapping, showing no significant stratification. This stands in stark contrast to the translation efficiency-based classification, which yielded highly significant prognostic separation with coherent biological interpretation (Fig. 7a, c).

To further assess whether the translation-based typing corrects misclassification by the expression-based approach, we identified patients with discordant assignments between the two methods (n = 54, 32.53% of the cohort). Among them, all patients assigned to the high-risk group (C3) by translation-based typing exhibited poor survival, including a subset of patients (n = 21, Trans-High) who had been misclassified as low-risk by expression-based typing (Fig. 7d). These results demonstrate that translation-based subtyping can identify truly high-risk patients who are overlooked by expression-based profiling and rectify its erroneous risk estimation. After adjusting for established clinical covariates for translation-based subtypes, including age, WHO grade, and IDH mutation status, multivariable Cox regression analysis further strengthened the subtype-associated survival differences (Supplementary Fig. 19d).

To independently validate the robustness and biological relevance of this classification, we applied the subtyping framework to an external glioma proteomics cohort and transcriptome cohort (GSE184941). The subtypes recapitulated the expected patterns of signature protein abundance (Supplementary Fig. 21a–c), preserved their distinct functional annotations (Supplementary Fig. 21a), and maintained consistent prognostic stratification, with C3 again showing the worst outcome (Supplementary Fig. 21d and Fig. 7h). To further assess the generalizability of the translation-based stratification framework, we applied the same subtyping approach to an independent glioma transcriptome cohort²⁰ (GSE184941, n = 99) after predicting translation efficiency via Translatomer. Subtype assignment using the discovery cohort centroids recapitulated the characteristic NTS/OTS translational activity patterns across C1–C4 (Supplementary Fig. 22a). Survival analysis revealed a consistent prognostic hierarchy, with C4 showing the most favorable outcomes and C3 the poorest (P = 0.066; C4 vs. C3, P = 0.049; Supplementary Fig. 22b). Multivariable Cox regression adjusting for IDH mutation status demonstrated that the prognostic distinction remained robust, with C1, C2, and C3 all showing worse survival compared to C4 (Supplementary Fig. 22c). Functional validation using GSVA confirmed that subtype-specific pathways, including MITOTIC_METAPHASE_AND_ANAPHASE (C1), NEGATIVE_REGULATION_OF_ALPHA_BETA_T_CELL_DIFFERENTIATION (C3), and NEURONAL_SYSTEM (C4), were consistently enriched in the corresponding subtypes in this independent cohort (Supplementary Fig. 22d–f). These results further support the robustness and generalizability of translation-based glioma stratification.

Collectively, these analyses demonstrate that the Translatomer-derived translational features, when integrated with immunogenic information, can be used to robustly stratify gliomas into biologically coherent and clinically meaningful subtypes. The consistency of this classification across independent datasets and molecular modalities underscores the central role of translation in shaping glioma heterogeneity.

Discussion

This study presents a comprehensive, multidimensional benchmark that evaluates a translational inference strategy with direct implications for precision oncology. While genomic studies have long relied on RNA-seq as a proxy for protein expression, our findings quantify the limitations of this assumption in clinical oncology. By systematically comparing in silico translatomes against the entrenched standard of RNA-seq signal across diverse biological contexts, we demonstrate that translational information provides a critical, missing layer of biological and clinical insight. Our key finding, that translation-aware analysis corrects consequential errors in patient risk stratification, argues for its integration into standard oncogenomic workflows.

Current benchmarks for computational translation models predominantly focus on horizontal performance comparisons (i.e., outperforming peers on curated datasets) or narrow technical validations. While these establish baseline accuracy, they leave more consequential, practical questions unanswered. How reliably these models perform across the diverse, real-world biological contexts encountered in research; whether their predictions can be extended to novel genomic domains; and, most critically, whether they provide tangible utility in driving downstream scientific discovery. To address these gaps, we implemented a comprehensive, three-dimensional evaluation of Translatomer, a state-of-the-art multimodal model. Our framework assesses: (1) the superior performance of translational inference over RNA-seq abundance; (2) method extensibility to noncoding RNA translation; and (3) practical utility in two translational oncology pipelines, i.e., refining neoantigen selection and correcting prognosis-relevant molecular subtypes in glioma via translation-aware analysis.

The performance benchmark across 15 independent datasets demonstrates that translational inference consistently outperforms RNA-seq signal, maintaining accuracy across sample types (primary tissues/cell lines), disease states (normal/tumor), cancer origins (brain, ovarian, pancreatic, and prostate), and temporal dynamics.

Our extensibility benchmark establishes, for the first time, that Translatomer can reliably predict translational potential in lncRNAs, opening new avenues for exploring the functional potential of the noncoding genome.

The Translatomer-based subtyping of glioma reveals biology that is invisible to transcriptomics. The identified neural and oncogenic translational programs, and the resulting subtypes (C1–C4), are not simply mirrors of existing transcriptomic classifications. The aggressive C3 subtype, characterized by a hyper-translated oncogenic program and a suppressed neural program, is defined by translational output, not mRNA levels. Its clear association with poor prognosis, mesenchymal features, and immunosuppression underscores that the translational landscape is a fundamental organizer of tumor biology and patient outcome. The validation of this stratification in an independent proteomic cohort, despite the technical challenges of proteomics, further strengthens its biological relevance.

Our evaluation also delineates the model’s current boundaries. The predictive accuracy decreases for genes with extremely low translational activity, reflecting the inherent challenge in capturing subtle biological signals within high-noise regions. This limitation highlights a frontier for methodological refinement. Future iterations could integrate additional layers of regulatory information, such as methylation, RNA secondary structure, or ribonucleoprotein interactions, to improve resolution for low-abundance targets. Additionally, we acknowledge that inferred ribosome occupancy does not equate to functional protein production. Experimental approaches such as mass spectrometry-based immunopeptidomics or reporter assays are necessary to validate these events.

Looking forward, the strategy outlined here is not confined to a single model. It establishes a framework for incorporating translation efficiency as a standard dimension in the analysis of routine RNA-seq data. Future directions should include: integrating translational profiles with other modalities (e.g., epigenomics, proteomics) for unified biomarker discovery; applying translation-aware neoantigen prediction to immunotherapy cohorts to predict response; and establishing inferred translational layers as a standard dimension in oncogenomic workflows, moving beyond the transcriptome-centric view. In conclusion, by rigorously benchmarking a translational inference strategy, this work provides both a methodological roadmap and empirical evidence that looking beyond the transcriptome to the translatome is essential for realizing the full promise of precision oncology.

Methods

Reproducibility of the original results

We conducted replication experiments via the public code of the Translatomer model from its GitHub repository (https://github.com/xiongxslab/Translatomer). The code package includes the model’s checkpoint (ckpt) files, environment configuration instructions, operational guidelines, and running examples. The data used for replication consisted of four public datasets from the Gene Expression Omnibus (GEO) for which detailed performance results were provided in the original paper: prostate (GSE130465²¹), epithelial (GSE200097²¹), hTERT-RPE1 (GSE138533²²), and RD (GSE103308²³) datasets. The code was executed via Python 3.9.16 on a computer equipped with an AMD EPYC 9654 96-Core Processor, running the Ubuntu 22.04.4 LTS operating system.

Glioma and validation datasets

The internal glioma dataset comprised 24 samples with paired RNA-seq and Ribo-seq data. For independent validation of model generalizability, we collected nine additional publicly available GEO-derived datasets, including brain (PRJNA525736²⁴, GSE51424²⁵, GSE229673²⁶, GSE229866²⁷, and GSE141013²⁸), ovarian cancer (GSE245504²⁹), pancreatic cancer (GSE235337³⁰), prostate (GSE130465²¹), and pancreatic ductal carcinoma (GSE235337³⁰) datasets. To further evaluate the relationship between predicted translational activity and protein output, we additionally analyzed a matched multi-omics dataset from HeLa cell lines (GSE308892¹¹) containing paired RNA-seq, Ribo-seq, and mass spectrometry-based proteomics across four samples. RNA-seq expression levels were quantified as log2-transformed transcripts per million (log2TPM). Spearman correlation coefficients were calculated between log2TPM values and protein abundance, as well as between length-normalized Translatomer-derived translational activity and protein abundance, for each of the four replicate samples. All public data were processed in accordance with the original data processing workflow of the original paper.

Transcriptome and genome data for glioma patients

We obtained the transcriptome and genome raw data of 166 patients with gliomas from the Chinese Glioma Genome Atlas (CGGA)³¹, along with the transcriptome data of 20 normal samples. Tumor/normal paired whole-exome sequencing (WES) FASTQs were aligned to the human genome (Gencode V38) via Burrows-Wheeler Aligner mem (version 0.7.17)³². All the alignments were postprocessed with the GATK 4.0.5.1 workflow³³, which included GATK Picard Mark Duplicates and Base Recalibrator. Somatic variants were called on the basis of tumor and normal WES data via Mutect2 (from GATK version 4.0.5.1) with default parameters. Tumor RNA-Seq FASTQs were aligned to the human genome (Gencode V38) via STAR (version 2.7.11b)³⁴, and the expression levels of ORFs and genes were quantified via featureCount (version v2.0.6)³⁵.

Ribo-seq of glioma patients

The Ribo-seq data from 24 glioma patients used in this study are the same as those used in one of our previous work³⁶. In brief, patient specimens were collected and flash-frozen in liquid nitrogen within 5 min of resection. The total lysate was digested with RNase I, and monosomes were isolated via size-exclusion chromatography. Following rRNA depletion and PAGE purification of short RPFs (20–38 nt), libraries were constructed and sequenced on the Illumina platform via the PE150 strategy.

Identification of Translated ORFs

To identify translated ORFs in glioma, the Ribo-seq data were processed as follows: ribosomal RNA, tRNA, and mtRNA were removed via Bowtie³⁷. The filtered reads were aligned to the HG38 genome and transcriptome using STAR. The RibORF (version 2.0)³⁸ pipeline was used for RPF offset correction and to predict ORFs with a minimum length of 8 amino acids and a RibORF score > 0.7. ORFs were reannotated on the basis of the metadata column in the GENCODE GTF. To ensure high confidence, we retained only those ORFs verified by at least one external public Ribo-seq dataset (GSE78959³⁹, GSE78960³⁹, GSE51424²⁵, GSE75947⁴⁰, and Translnc⁴¹), resulting in a final set of 82,068 ORFs.

Identification of HLA-I-bound peptides

We used OptiType software (version 1.3.5)⁴² to determine the four-digit HLA-I of patients. Next, netMHCpan 4.1⁴³ was used to predict the binding affinity of peptides and HLA-I. We derived all possible 8–12 mer peptides from ORFs and selected those with an IC 50 threshold of 500 nM (a concentration that inhibits 50% binding of the fluorescein-labeled reference peptide) as strong HLA-I binders.

Neoantigen prediction and analysis pipeline

Somatic nonsynonymous mutations were called from whole-exome sequencing data via a standardized pipeline. For each mutation, we identified nonsynonymous mutations occurring in ORFs, including single nucleotide variants (SNVs), in-frame insertions, frameshift insertions, in-frame deletions, and frameshift deletions. Next, the mutated sequences were translated to proteins according to the standard codon table. These proteins were then fragmented into short peptides containing at least one non-self-residue (8–12 mer) by sliding windows. The Translatomer-predicted ribosome profiling signal was summed for the bins from which the neoantigens were derived. The translation effect score (TE) was calculated as the difference (alternative-reference). The representation is expressed as follows:

$$\mathit{TE}=\sum _{b_i\in C_e}(f_{\mathit{Translatomer}}{(x}_{\exp },x_{\mathit{alt}})-f_{\mathit{Translatomer}}{(x}_{\exp },x_{\mathit{ref}}\left)\right)$$ 1

where $C_e$ represents the bin set from which epitopes are derived, $x_{\exp }$ represents the RNA-seq signals, $x_{\mathit{alt}}$ represents the modified sequence with the alternative allele and $x_{\mathit{ref}}$ represents the reference genome sequence. A threshold of 0.1 was applied to filter the candidates. For the conventional baseline, filtering was based on the expression levels of the genes and ORFs that derive neoantigens, using a threshold set to 1 TPM for genes and 0.1 TPM for ORFs. The probabilities of peptide binding to HLA-I molecules were then predicted via netMHCpan 4.1 (IC50 < 500 nM). The conventional expression-based workflow was adapted from the MINER pipeline³⁶.

Assessment of neoantigen immunogenic potential

The immunogenic potential of each predicted neoantigen was assessed by quantifying its homology to known immunogenic peptides from the Immune Epitope Database & Tools (IEDB, https://www.iedb.org/home_v3.php) and individual publications^13,44 regardless of peptide source to calculate the probability of HLAp being recognized by the TCR (score ≥ 1e-16)⁴⁵.

Sensitivity analysis of neoantigen prioritization

To evaluate the robustness of the neoantigen prioritization pipeline to parameter choices, we performed sensitivity analyses across five threshold combinations.

Baseline group (original settings): TE score threshold = 0.1; expression-based gene threshold = 1 TPM, ORF threshold = 0.1 TPM; HLA-binding IC50 threshold = 500 nM.

Group 1: TE score threshold reduced to 0.05; expression thresholds lowered to gene ≥ 0.5 TPM, ORF ≥ 0.05 TPM; HLA threshold unchanged.

Group 2: TE score threshold increased to 0.2; expression thresholds raised to gene >= 5 TPM, ORF ≥ 0.5 TPM; HLA threshold unchanged.

Group 3: translation and expression thresholds same as control; HLA IC50 threshold tightened to 150 nM.

Group 4: translation and expression thresholds same as control; HLA IC50 threshold tightened to 50 nM.

For each threshold combination, we evaluated (i) the composition of final neoantigen sets (canonical, noncanonical, mixed), (ii) predicted immunogenicity scores (≥1e-16 threshold and mean scores), and (iii) pMTnet-based TCR binding predictions (rank < 0.01). Results were compared between the translation-based and expression-based pipelines across all five settings.

Independent validation of neoantigen prioritization

To assess the generalizability of the translation-aware neoantigen prioritization framework, we applied the pipeline to an independent glioma cohort¹⁴ (PRJNA1051047, n = 100) with matched RNA-seq and whole-exome sequencing data. The same filtering thresholds and HLA-binding prediction parameters were applied as in the discovery cohort. The resulting neoantigen sets were compared between the translation-based and expression-based pipelines, with evaluation focusing on noncanonical neoantigen proportion and predicted immunogenicity scores, as described in the discovery analysis.

Identification of TCR clonotypes

The raw RNA-Seq data of 166 glioma patients were analyzed via MiXCR (version 3.0.13)⁴⁶. The parameters were modified according to the manual to assess the TCR-β chain repertoire via RNA-seq. Analysis of the TCR clonotypes was performed by immunarch via R for statistical computing. pMTnet (version 1, https://dbai.biohpc.swmed.edu/pmtnet/index.php) was employed to identify the relationships among HLA-I, neoantigens, and TCR CDR3β (rank < 0.01).

Definitions of canonical, noncanonical and mixed neoantigens

Neoantigens are classified according to the type of nonsynonymous mutation they derive from, with this classification determined by the nature of amino acid changes induced across multiple ORFs⁴⁷. Specifically, a nonsynonymous mutation is categorized as “canonical” when all affected ORFs are canonical, “noncanonical” when all affected ORFs are noncanonical, and “mixed” when it impacts both canonical and noncanonical ORFs.

Benchmarking metrics and baseline definition

To evaluate the predictive performance of the Translatomer in modeling translation efficiency, the PCC and SCC were calculated between the predicted and observed Ribo-seq signals at both the protein-coding gene and ORF levels. For protein-coding genes, genomic coordinates were obtained from GENCODE v43, whereas ORF regions were defined via ribosome profiling-guided ORF calling from the in-house glioma dataset (n = 24).

Model performance was compared with a transcriptome-based baseline using RNA-seq signals as a proxy for translation. For each sample, we calculated the proportion of genes or ORFs with effective predictions via a widely adopted threshold (correlation ≥ 0.3 and P < 0.05). This comparison quantified the extent to which Translatomer improves translation prediction over RNA abundance-based estimates.

Genome-wide performance evaluation

To assess the performance of the Translatomer across diverse genomic contexts, we conducted a systematic genome-wide analysis. The human reference genome (GRCh38) was partitioned into consecutive, non-overlapping 65,536 bp windows. For each window, we extracted the predicted Ribo-seq signal, the experimentally observed Ribo-seq signal, and the prediction accuracy was quantified via the PCC and SCC correlation coefficients. On the basis of the correlation strength and statistical significance, each window was assigned to one of three categories: low (P ≥ 0.05 or both PCC and SCC < 0.3), high (P < 0.05 and PCC, SCC ≥ 0.5), or median (all others).

Each 65,536 bp window was annotated on the basis of its overlap with genomic features defined in the GENCODE (v.43) comprehensive annotation. For each genomic window, we calculated the fraction of bases overlapping coding sequences (“CDS”), 5’ UTRs (“five_prime_UTRs”), 3’ UTRs (“three_prime_UTRs”), and long noncoding RNA (“lncRNA”) regions. The lncRNA regions were specifically defined by the gene_type or transcript_type attribute labeled “lncRNA”, and any regions overlapping with protein-coding exons were excluded to prevent misclassification. These values were further normalized by the total genomic length of each feature type to account for differences in their genome-wide abundance.

Sparsity analysis

To account for sparsity in Ribo-seq signals, we performed additional evaluations by stratifying genomic regions based on observed translation activity. Regions were divided into those with no observed translation (Ribo-seq signal = 0) and those with active translation (Ribo-seq signal > 0) in each 65,536 bp genomic bin.

To ensure the robustness of our analyses, we designed multiple analytical settings, including all regions and regions with zero translation signals removed. We also applied distinct thresholds for defining highly translationally active positions, as well as a series of k values (1, 5, 10, 20, 50) for the top-k analysis.

For regions with active translation, high-translation positions were defined using percentile-based thresholds (top 50%, 75%, and 90% of non-zero Ribo-seq signals). Model performance was evaluated using the area under the precision-recall curve (AUPRC), comparing Translatomer-predicted signals and RNA-seq-based estimates against these high-translation labels.

In addition, top-k enrichment analysis was performed by ranking genomic positions based on predicted signals and calculating the proportion of true high-translation positions within the top k% predictions, normalized by the global proportion of high-translation positions. All analyses were conducted across protein-coding genes, ORFs, and lncRNAs.

Construction of the glioma translatome atlas

A comprehensive glioma translatome atlas was constructed for 166 samples via the Translatomer to predict the translation efficiency of both protein-coding genes and lncRNAs. For protein-coding genes, translation efficiency was quantified as the sum of the predicted Ribo-seq signals across all 65,536 bp bins overlapping their CDS regions. For lncRNAs that lack annotated CDSs, we quantified translation from their predicted ORFs. This generated a unified translation efficiency matrix encompassing both gene types.

To ensure robust signal detection, genes were retained only if (i) translation efficiency was non-zero in at least 5% of the samples and (ii) the mean efficiency across detected samples exceeded the 25th percentile (2.46) of the global non-zero distribution. After filtering, 14,180 genes were retained for downstream analyses.

Normalization was performed using a CPM-like approach to account for sequencing depth³. This process involves scaling each gene’s value by the total signal in the sample and then multiplying by a constant factor (1 × 10⁶).

Identification of differentially translated genes

Differential translation analysis was performed between glioma samples (n = 166) and normal tissue samples (n = 20) via the Wilcoxon rank-sum test. The fold change (FC) in translation efficiency for each gene was calculated as the ratio of the mean efficiency in glioma samples to the mean in control samples. To handle zero values, a pseudocount of 1 × 10⁻⁶ was added to each mean prior to FC calculation.

Genes were classified as significantly upregulated if they met the dual thresholds of a p-value < 0.05 and FC > 1.5, and as downregulated if they met the p-value < 0.05 and FC < 1/1.5. This analysis revealed 1183 significantly upregulated genes and 1323 significantly downregulated genes in glioma.

Definition of translation-based gene signatures

To define molecular signatures that integrate neoantigen immunogenicity and translational regulation, we implemented a multistep gene classification strategy based on two orthogonal layers of information: sample-specific neoantigen burden and alterations in translation efficiency. For classification by neoantigen burden (NB), a gene‒sample-neoantigen count matrix was constructed. The mean neoantigen count across all gene-sample pairs was calculated as a global threshold. A gene was classified as “high NB” if, in every sample where it generated neoantigens, its neoantigen count was greater than or equal to the global mean. Conversely, a gene was classified as “low NB” if, in every sample where it generated neoantigens, its count was below the mean. Based on prior differential analysis, genes were also labeled as “trans-up” (upregulated in tumor) or “trans-down” (downregulated in tumor). Using the classifications above, we defined two composite gene programs via set operations: Oncogenic Translation Program (Onco-TP) as (Low NB ∪ Trans-Up)\(High NB ∪ Trans-Down) and Neural Translation Program (N-TP) as (High NB ∪ Trans-Down)\(Low NB ∪ Trans-Up). Furthermore, we extracted core signatures: the Oncogenic Translation Signature (OTS) as Low NB ∩ Trans-Up and the Neural Translation Signature (NTS) as High NB ∩ Trans-Down.

Identification of translation-based subtypes

Molecular subtypes were derived by performing k-means clustering (k = 4) on the glioma translation efficiency matrix via the oncogenic translation signature (OTS, n = 63 genes) and neural translation signature (NTS, n = 11 genes). The resulting subtypes (C1–C4) exhibited distinct patterns of OTS and NTS activity: C1, low-low; C2, high-high; C3, high OTS/low NTS; and C4, low OTS/high NTS.

The choice of k = 4 was guided by the biological continuum of the data, GSVA-derived enrichment scores for NTS and OTS revealed a strong negative correlation defining a major axis of variation, and trajectory analysis identified four distinct states along a continuous progression from normal brain tissue to the most aggressive subtype (C3). These four states exhibited distinct functional enrichment patterns and significant survival differences, supporting k = 4 as the most interpretable discretization of the underlying biological spectrum.

To evaluate subtype robustness, we performed a resampling-based co-clustering stability analysis. In each of 100 iterations, 80% of samples (n = 132) were randomly subsampled from the NTS/OTS translation efficiency matrix without replacement, and k-means clustering (k = 4) was applied using the same strategy as in the main analysis. For each sample pair (i, j), the co-clustering frequency was calculated as the proportion of iterations in which both samples were included and assigned to the same cluster. For each sample, a stability score was defined as the mean co-clustering frequency with other samples assigned to the same original subtype. Stability scores were summarized across samples within each subtype.

For comparison, glioma samples were also categorized according to the Verhaak molecular subtypes (proneural, neural, classical and mesenchymal) via established subtype-specific signatures⁴⁸. The translation efficiency matrix was z-score normalized, which is consistent with the normalization approach used in the original Verhaak analysis. For each sample, correlation scores were computed between the scaled gene translation values and a binary indicator vector (1 for genes in the subtype signature, 0 otherwise) for each of the four subtypes (proneural, neural, classical, mesenchymal). Each sample was assigned to the subtype with the highest correlation score, resulting in predicted subtype labels for all samples.

Identification of expression-based subtypes

For the expression profile-based molecular stratification of glioma samples, we employed the same set of 74 translation-based gene signatures and calculated Euclidean distances between each sample’s expression and the subtype centroids to assign patients to the C1–C4 subtypes.

Gene set enrichment analysis

To elucidate the biological functions and pathways driving each subtype, we conducted gene set enrichment analysis (GSEA) via the clusterProfiler R package⁴⁹. For each subtype, we performed a comparison against all other subtypes combined. Genes were ranked on the basis of their fold change in translation efficiency in the subtype of interest versus the others. This ranked list was then tested for enrichment against a comprehensive collection of gene sets from the KEGG⁵⁰, Gene Ontology (GO), and Reactome databases. A pathway was considered significantly and specifically enriched in a given subtype if it met the following criteria: (1) a normalized enrichment score (NES) > 0, indicating bias toward the top of the ranked list; (2) a statistically significant false discovery rate (FDR) < 0.05; and 3 enrichment exclusively in the target subtype.

Validation of subtype classification in external cohorts

To validate the robustness of the identified subtypes, an external validation was performed using proteomic data from the CPTAC⁵¹ (Clinical Proteomic Tumor Analysis Consortium, PDC000446) glioblastoma multiforme cohort and transcriptomic data from an independent glioma cohort²⁰ (GSE184941, n = 99). Subtype-specific centroids (C1–C4) were generated by computing the mean expression of signature genes for each subtype in the in-house cohort.

Proteomic data from CPTAC were filtered to retain tumor samples and proteins with <50% missing values across samples. Missing values were imputed via a K-nearest neighbor (KNN) algorithm after quantile normalization (via normalizeBetweenArrays from the limma package).

To predict subtypes for CPTAC samples on the basis of their similarity to the in-house cohort centroids, the Euclidean distance was calculated between each sample’s scaled protein expression and the subtype centroids, and each sample was assigned to the closest centroid.

For the GSE184941 cohort, RNA-seq data were processed through Translatomer to obtain predicted translation activity, and samples were assigned to subtypes by calculating Euclidean distance to the same centroids derived from the discovery cohort.

Statistical analysis

Correlation with survival outcomes was performed with the R packages survival (version 3.7-0) and survminer (version 0.5.0). Multivariable Cox proportional hazards models were constructed with established clinical covariates. In the discovery cohort (CGGA), age, WHO grade, and IDH mutation status were included, while additional variables were omitted due to substantial missingness that would have markedly reduced sample size. In the external validation cohort (GSE184941), IDH mutation status was included as the key available clinical covariate. To characterize the properties of OTS and NTS, we investigated oncogenic signaling pathways (n = 10). GSVA enrichment analysis was performed on these 12 gene sets via the in-house glioma translation profile. The resulting enrichment matrix was subjected to PCA for dimensionality reduction, and the first two principal components (PC1 and PC2) were selected for visualization. Trajectory analysis was performed via Monocle on the basis of the in-house glioma translation profile. Comparisons between the two groups were performed with the Wilcoxon rank-sum test. All P values are two-sided unless otherwise stated.

Author contributions

Conceptualization, J.X., Y.L., Q.J., and K.W.; design, Y.M., C.G., K.X., and Y.C.; investigation, D.L, K.W., C.G., J.M., Y.M., and S.L.; analysis and visualization, Y.M., C.G., Y.C., K.X., and S.L.; writing, Y.M., Y.L., J.X., K.W., and C.G. All authors contributed to the editing and review of the manuscript.

Data availability

The original four benchmark datasets for the Translatomer replication are available from GEO under accession codes GSE130465, GSE200097, GSE138533, and GSE103308. The independent datasets used for the generalizability assessment in this study are available from GEO under accession codes PRJNA525736, GSE51424, GSE229673, GSE229866, GSE141013, GSE245504 and GSE235337. The matched HeLa multi-omics dataset¹¹ (RNA-seq, Ribo-seq and mass spectrometry-based proteomics) used for orthogonal experimental validation is available under accession code GSE308892. The external glioma cohort used for independent validation of neoantigen prioritization is available under BioProject accession PRJNA1051047¹⁴. The external glioma transcriptome cohort used for independent validation of patient stratification is available under accession code GSE184941²⁰. The internal glioma cohort was obtained from the CGGA portal (http://www.cgga.org.cn/). The dataset comprised RNA-seq data from 166 glioma and 20 normal samples, WES data from the same 166 glioma patients, and Ribo-seq data from a subset of 24 of these glioma patients. These data were used for model application, neoantigen analysis, and subtyping. Access requires registration and completion of a data access agreement as stipulated by the data providers. The proteomic data from the CPTAC used for independent validation of the molecular subtypes are available from the CPTAC Data Portal under the PDC000446 project. Glioma immunopeptidomics data were retrieved from the PRIDE database (https://www.ebi.ac.uk/pride/), with the following accession numbers: PXD003790¹⁹, PXD008127¹⁸, PXD008984¹⁷, and PXD020186¹⁶. All other data supporting the findings of this study are available within the article, its Supplementary Information files, or from the corresponding authors upon reasonable request.

Code availability

The original code of the Translatomer is available at https://github.com/xiongxslab/Translatomer. The codes with detailed instructions to run the reproducibility results and analyze the benchmarking are available via GitHub at https://github.com/ComputationalEpigeneticsLab/Translatomer_Reusability_Study.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s41698-026-01567-y.