DDA-BERT: end-to-end training for data-dependent acquisition mass spectrometry-based proteomics

Abstract

Peptide-spectrum match (PSM) rescoring is critical for accurate peptide identification in data-dependent acquisition (DDA)-based proteomics. Existing rescoring frameworks typically combine search-engine scores with heuristic or learned auxiliary features to refine PSM ranking and confidence estimation. Although recent approaches incorporate deep learning-derived representations of spectra, retention time, or ion mobility, the final decision stage still commonly relies on separately trained shallow classifiers, constraining the expressive capacity of the overall scoring framework. Here, we introduce DDA-BERT, a transformer-based end-to-end deep learning model trained with ~271 million PSMs from 11 species. DDA-BERT consistently outperforms existing tools across species-specific benchmarks, achieving 2.24%–269.35%, 3.73%–141.46%, 5.53%–45.64%, and 3.68%–62.77% increases in peptide identifications on human, yeast, Drosophila, and Arabidopsis datasets, respectively. The model retains high sensitivity in trace-level proteomics samples. On HLA immunopeptidomics data, DDA-BERT further increases peptide identifications by 4.14%–87.47%. The main limitations of DDA-BERT include the requirement for GPU-based computing and the need for substantial, diverse training datasets to achieve optimal model performance. This study introduces an alternative DDA rescoring approach and establishes a methodological foundation for scalable, AI-driven peptide identification in DDA proteomics.

Peptide-spectrum match (PSM) rescoring is crucial for accurate peptide identification in DDA proteomics. Here, the authors introduce DDA-BERT, an end-to-end transformer-based deep learning model that improves peptide identification across multiple species and HLA immunopeptidomics.

Introduction

In data-dependent acquisition (DDA)-based proteomics, peptide identification typically relies on a protein sequence database search approach. This process begins with the in silico digestion of input protein sequences to generate theoretical peptide spectra, which are subsequently matched against experimental MS/MS spectra. To improve peptide identification accuracy, machine learning-based rescoring algorithms integrate the initial scores from database search engines with a set of features, such as peptide length, precursor charge state, and precursor mass. This integrative strategy enables a more robust and comprehensive assessment of peptide-spectrum match (PSM) confidence^1,2. To control false discovery rate (FDR), a target-decoy strategy is commonly used. This method utilizes a decoy database containing reversed or randomized sequences with similar compositional properties to the target database³. By ranking PSMs by score and applying an FDR threshold, final PSM identifications were obtained.

Over the past few decades, numerous PSM rescoring algorithms have been developed to improve peptide identification reliability. Commonly used tools include, but are not limited to, Percolator¹, PeptideProphet², Scavager⁴, and Nokoi⁵. PeptideProphet², a widely used post-processor, applies a mixture model to estimate posterior probabilities without external training data; however, its simplicity limits adaptability to complex or noisy spectra. Percolator employs a semi-supervised support vector machine framework with customizable features and is widely used for its flexibility¹. Scavager utilizes gradient boosting techniques to achieve robust, high-accuracy rescoring⁴. In contrast, Nokoi prioritizes interpretability and computational efficiency by employing L1-regularized logistic regression for sparse and scalable rescoring⁵. Despite their widespread use, these methods typically rely on shallow machine learning models and manually engineered features, which can limit their ability to model the complex, nonlinear relationships inherent in PSMs. Furthermore, rescoring models are often trained independently on individual MS files or projects, which constrains information sharing across datasets and preventing the capture of broader proteomic patterns. While project-level rescoring workflows are generally effective, particularly for large datasets acquired under homogeneous experimental conditions, dataset-specific model training can still lead to variable performance. This is especially relevant for studies involving heterogeneous samples, limited data volumes, or strong batch effects within a project.

Recent advances in deep learning have enabled highly accurate prediction of peptide properties that are critical for improving PSM rescoring. Deep neural networks, particularly those based on convolutional or transformer architectures, have been successfully applied to predict peptide retention time (RT)^6–8, fragmentation patterns (MS/MS spectra)^6,7,9, collisional cross section (CCS)^7,10, ion mobility (IM)^7,10, and binding affinity (BA)¹¹. These predicted properties provide orthogonal evidence that complements traditional search engine scores and are increasingly integrated into modern rescoring frameworks. Tools, such as Percolator and Scavenger, can incorporate predicted RT or spectral similarity as additional discriminative features to improve classification accuracy. MS²Rescore¹² leverages MS2PIP-predicted theoretical MS/MS spectra and DeepLC-predicted RTs to generate engineered features, which, when combined with raw search engine scores, are input into post-processing tools like Percolator or Mokapot¹³ to enhance PSM ranking. InSPIRE¹⁴ further extends this strategy by predicting the similarity between theoretical and experimental spectra, BA, and RT, and integrating them with search engine scores into a unified rescoring framework. MSBooster¹⁵, part of the FragPipe¹⁶ suite, supports spectral and RT prediction via DIA-NN¹⁰ or Koina¹⁷, and integrates these features into Percolator-based rescoring. AlphaPeptDeep adopts Transformer and LSTM architectures to predict peptide properties, such as MS/MS spectral intensities, RT, CCS, and other properties. These predictions are subsequently integrated with search engine outputs and incorporated into spectral quality assessment and FDR control workflows⁷.

Despite their promising performance, these tools are built on a fragmented architecture that decouples feature extraction from model training. In these frameworks, deep learning is used only to generate precomputed features for subsequent shallow rescoring, rather than to enable end-to-end learning. As such, the overall system is limited in its ability to benefit from joint optimization of representation learning and downstream task adaptation.

End-to-end deep learning methods offer significant advantages by learning directly from raw data without the need for explicit feature engineering or separate representation learning. This unified framework enables models to capture complex nonlinear relationships while facilitating holistic optimization across all stages of the pipeline. In de novo sequencing tasks, models like Casanovo¹⁸, π-PrimeNovo¹⁹, and GraphNovo²⁰ have successfully employed end-to-end approaches to determine peptide sequences from mass spectra without relying on handcrafted or precomputed features, achieving superior performance over traditional methods. This observation naturally raises the question: since end-to-end deep learning methods can achieve better performance on de novo sequencing tasks where the sequence space is unconstrained, could they also show improvements in rescoring tasks where the sequence space is constrained? Thus, we introduce DDA-BERT, a transformer model that uses an end-to-end approach to learn rescoring functions for PSMs directly from raw MS spectra and peptide sequences, without requiring any additional manual feature engineering or separate representation learning.

Results

DDA-BERT computational workflow

DDA-BERT is an end-to-end PSM rescoring framework designed to enhance peptide identification from DDA-MS data. The model was trained on a large-scale proteomics corpus comprising 12,285 LC–MS/MS runs from 110 independent projects, spanning 11 species and 10 mass spectrometer platforms, and totaling approximately 271 million PSMs (Fig. 1A and Supplementary Data 1). As shown in Fig. 1B, the DDA-BERT workflow adopts an iterative refinement paradigm, in which database searching, model-based rescoring, and hard negative mining are integrated into a closed-loop process to progressively improve discrimination between correct and incorrect peptide-spectrum matches. Raw DDA-MS files were first searched using multiple database search engines, including Sage, AlphaPept, and FragPipe (with MSBooster), to generate candidate PSMs across diverse datasets. High-confidence target PSMs identified at 1% FDR were then aggregated across search engines to form a robust set of positive training examples. Decoy PSMs were exclusively obtained from Sage and subjected to a series of custom filtering steps to generate a set of curated negatives. These curated positives and negatives were used to train an initial base model. During iterative hard-negative selection, the base model was used to rescore candidate PSMs, thereby identifying heterogeneous and challenging negative examples that were difficult to distinguish from true matches. These hard negatives were then incorporated into subsequent rounds of training, progressively refining the model and improving its robustness and discrimination. This procedure yielded the final DDA-BERT rescoring model. For analysis of experimental DDA-MS files, candidate PSMs are first generated by standard database search engines (e.g., Sage, AlphaPept, and FragPipe) and subsequently rescored by DDA-BERT, with final peptide and protein identifications obtained through downstream FDR control. Figure 1C illustrates the architecture of DDA-BERT, which integrates a spectrum encoder and a peptide precursor encoder. Model inputs include raw MS/MS spectra, precursor m/z, and charge state. To improve representation learning and robustness, we apply multimodal corruption during training: spectral inputs are perturbed via spectrum dropout, and peptide sequences are masked using contiguous-span masked language modeling (span-level MLM), in which one or more nonoverlapping sequence segments are masked. Model training further adopts a multi-task learning strategy that jointly optimizes masked peptide reconstruction and PSM quality estimation, supporting accurate peptide identification and improved generalization across diverse DDA-MS datasets.

Fig. 1. Training data composition, workflow, and model architecture of DDA-BERT.

A Composition of the DDA-BERT training corpus. The dataset comprises 12,285 DDA LC–MS/MS runs collected from 110 independent projects spanning 11 species and acquired on 10 mass spectrometer platforms, covering animals, plants, and microorganisms. The table summarizes the number of species, projects, LC–MS/MS runs, and PSMs included in the training set for each category; representative species and their corresponding PSM counts are shown on the right. B Overview of the DDA-BERT workflow. Raw DDA-MS files are searched using multiple database search engines (Sage, AlphaPept, and FragPipe with MSBooster) to generate candidate PSMs. High-confidence positives and curated negatives are used to train an initial model, which is further refined through iterative hard negative mining. In the application stage, candidate PSMs are rescored by DDA-BERT, followed by FDR filtering and downstream peptide and protein identification. C Architecture and training objectives of DDA-BERT. MS/MS spectra and peptide precursor information are processed by dedicated spectrum and peptide encoders. Training employs spectrum dropout and contiguous-span masking of peptide sequences, and jointly optimizes amino acid prediction and PSM scoring objectives. AP AlphaPept, FP FragPipe (MSBooster), DBSE database search engines.

Robust target-decoy score distributions from DDA-BERT

To evaluate whether DDA-BERT produces statistically appropriate score distributions under the target-decoy framework and supports reliable FDR control, we examined the score distributions of target and decoy PSMs. Following the evaluation framework established by Debrie et al.²¹, we generated percentile-percentile (P-P) plots by comparing the cumulative distribution of target scores against the theoretical uniform distribution of decoys, and visualized the separation between the two populations using score histograms. For comparison, Sage and FragPipe were analyzed in parallel using the same diagnostic plots. As shown in Fig. 2A, while the target-decoy score distribution produced by DDA-BERT on the human dataset differs from that of other tools, the target scores exhibit a clear bimodal structure. A large proportion of the target PSMs is concentrated in the high-scoring region, forming a sharp, compact peak, while the remaining target PSMs are distributed in the low-scoring region. Compared to other tools, DDA-BERT shows a clearer separation between high-scoring target PSMs and lower-scoring matches (including decoys), with the high-score target mode distinctly separated from the low-scoring background. The P-P plots (Fig. 2B) show that the DDA-BERT curve is continuous and monotonic, without discontinuities or anomalous clustering. At low score thresholds, the decoy cumulative fraction (Fd) increases rapidly, reflecting the predominance of decoy PSMs in the low-score region. As the score threshold increases, the target cumulative fraction (Ft) initially rises gradually and then increases more prominently in the high-score region. As Fd approaches 1, Ft rises more sharply, yet remains below. Sage and FragPipe exhibit similar overall trends, although the late-stage increase of Ft as Fd approaches 1 is more pronounced than that observed for DDA-BERT. Across nonhuman benchmark datasets (Supplementary Fig. 1), FragPipe and DDA-BERT exhibit patterns largely consistent with those observed in the human dataset. For Sage, distinct behaviors are observed across species: in D. melanogaster, the P-P curve is less smooth, with the score histogram indicating a small subset of higher-scoring decoy PSMs; in A. thaliana, the target score distribution is not strictly bimodal and instead displays two high-score peaks; whereas in S. cerevisiae, Sage shows a clearer bimodal target distribution and a smooth P-P trend, similar to the human dataset.

Fig. 2. Target-decoy score distributions and peptide-length dependence of DDA-BERT on the human dataset.

A Score distributions of target (teal) and decoy (purple) PSMs produced by Sage, FragPipe, and DDA-BERT. Histograms show the distribution of scores for target and decoy PSMs for each tool. DDA-BERT scores are mapped to the [0, 1] interval via a sigmoid transformation. B Percentile-percentile (P-P) plots comparing the empirical cumulative distribution functions (ECDFs) of target and decoy scores for each tool. The grey diagonal line (y = x) indicates identical target and decoy distributions; deviations below the diagonal reflect enrichment of higher scores among target PSMs compared to decoys. C Distribution of DDA-BERT scores as a function of peptide length for target (left) and decoy (right) PSMs. Box plots summarize the interquartile range (IQR), with the median indicated by the center line. The whiskers extend to the minima and maxima within 1.5 × IQR, and outliers are plotted separately as individual points. In total, 477,548 PSMs were analyzed by DDA-BERT, including 371,639 target PSMs and 105,909 decoy PSMs, derived from database search results generated by FragPipe, Sage, and AlphaPept. Unless otherwise specified, all references to FragPipe in the text refer to FragPipe with MSBooster enabled. Source data are provided as a Source data file.

To further assess score stability across peptides of different lengths, we examined DDA-BERT scores as a function of peptide length (Fig. 2C). Target PSMs consistently retained high scores with relatively low dispersion across a broad length range (7–35 amino acids), whereas decoy PSMs remained tightly clustered near zero with minimal variance. Notably, the clear separation between targets and decoys persisted as peptide length increased.

DDA-BERT achieves superior PSM recovery across species and FDR thresholds

To evaluate the overall sensitivity of DDA-BERT in PSM identification, we systematically benchmarked its performance against five state-of-the-art rescoring and search frameworks: AlphaPept, AlphaPeptDeep, MS²Rescore, Sage, and FragPipe (MSBooster), across four representative species: Homo sapiens, Drosophila melanogaster, Arabidopsis thaliana, and Saccharomyces cerevisiae, with PSM counts quantified at varying FDR thresholds ranging from 0.002 to 0.01. As shown in Fig. 3, DDA-BERT achieves higher PSM yields than the compared methods across the four benchmark datasets evaluated in this study over the evaluated FDR range. For the human dataset, comprising colorectal cancer (CRC) formalin-fixed paraffin-embedded (FFPE) tissue samples analyzed on an Orbitrap-based platform with a 120-min liquid chromatography gradient²², at a 1% FDR threshold, DDA-BERT identified over 63,556 PSMs. This corresponded to improvements of 192.84% over AlphaPept (21,704 PSMs), 82.98% over AlphaPeptDeep (34,734 PSMs), 44.09% over MS²Rescore (44,109 PSMs), 47.45% over Sage (43,105 PSMs), and 18.11% over FragPipe (53,809 PSMs). In the D. melanogaster dataset, acquired on a Q Exactive HF with a 50-min chromatographic gradient²³, at 1% FDR, DDA-BERT reported an average of 25,500 PSMs, outperforming AlphaPept (16,074 PSMs) by 58.65%, AlphaPeptDeep (21,539 PSMs) by 18.39%, MS²Rescore (18,821 PSMs) by 35.49%, Sage (18,282 PSMs) by 39.49%, and FragPipe (21,038 PSMs) by 21.21%. In the A. thaliana dataset obtained from tryptic digestion of seedling samples and analyzed on a Q Exactive HF mass spectrometer²⁴, DDA-BERT identified 50,757 PSMs. This corresponds to 26.38% more PSMs than those identified by MS²Rescore (40,161), DDA-BERT also outperformed several other tools, yielding 100.93% more PSMs than AlphaPept (25,261), 13.21% more than AlphaPeptDeep (44,834), 30.77% more than Sage (38,812), and 18.81% more than FragPipe (42,720). Finally, in the S. cerevisiae dataset, acquired on a Orbitrap Exploris 480 mass spectrometer with a 120-min LC gradient, at a 1% FDR threshold, DDA-BERT identified 48,901 PSMs. Compared to other tools, this corresponded to gains of 138.47% over AlphaPept (20,506 PSMs), 2.59% over AlphaPeptDeep (47,666 PSMs), 78.96% over MS²Rescore (27,326 PSMs), 59.93% over Sage (30,577 PSMs), and 33.68% over FragPipe (36,581 PSMs). These results highlight that DDA-BERT consistently outperforms existing rescoring and search tools across diverse datasets and application scenarios.

Fig. 3. Performance of different tools across varying FDR thresholds in multiple datasets.

Line plots showing the average number of PSMs identified by each tool at FDR thresholds ranging from 0.002 to 0.01 across four species datasets: A Homo sapiens, B Drosophila melanogaster, C Arabidopsis thaliana, and D Saccharomyces cerevisiae. Each curve corresponds to a different identification tool, including DDA-BERT, MS²Rescore, FragPipe (MSBooster), AlphaPeptDeep, AlphaPept, and Sage, reflecting their identification performance under different levels of FDR control. For each species, performance was evaluated on at least three independent DDA-MS files. Source data are provided as a Source data file.

Superior peptide-level sensitivity of DDA-BERT across diverse species

Under a fixed 1% FDR threshold, we further compared the number of unique peptides identified by DDA-BERT with those identified by five established rescoring and search frameworks: AlphaPept, AlphaPeptDeep, Sage, MS²Rescore, and FragPipe (MSBooster), across datasets from four species: Homo sapiens, Drosophila melanogaster, Arabidopsis thaliana, and Saccharomyces cerevisiae. As shown in Fig. 4, in the human dataset, DDA-BERT identified a total of 40,105 unique peptides, achieving the highest performance among all evaluated tools. The second-best tool, FragPipe, identified 39,227 peptides, representing a 2.24% fewer than DDA-BERT; MS²Rescore identified 32,379 peptides, 23.86% fewer than DDA-BERT. AlphaPeptDeep identified 15,182 peptides (164.15% fewer than DDA-BERT), while AlphaPept showed the lowest performance, identifying only 10,858 peptides—a 269.35% reduction compared to DDA-BERT. In the D. melanogaster dataset, DDA-BERT identified over 12,834 peptides, whereas other methods identified between 8,813 and 12,162 peptides, corresponding to relative gains of 5.53–45.64%. In the A. thaliana dataset, DDA-BERT again achieved the best performance, identifying 15,141 peptides, which is 3.68% higher than FragPipe and substantially higher than other tools, with improvements ranging from 10.13% to 62.77%. In the S. cerevisiae dataset, DDA-BERT identified 19,797 peptides, outperforming FragPipe (19,086 peptides) and yielding substantially higher peptide counts than AlphaPept, AlphaPeptDeep, MS²Rescore, and Sage, with increases of 141.46%, 26.56%, 37.88%, and 27.69%, respectively. In addition, we evaluated DDA-BERT at the peptide precursor and protein group levels (Supplementary Figs. 2 and 3). DDA-BERT consistently achieved the highest number of peptide precursor identifications across all datasets (Supplementary Fig. 2). At the protein group level (Supplementary Fig. 3), it remained highly competitive, achieving the highest counts in the Drosophila and Arabidopsis datasets, performing comparably to FragPipe in the human dataset, and ranking second in S. cerevisiae.

Fig. 4. Peptide identification performance across four species.

Bar plots comparing the number of peptides identified at 1% FDR by six computational tools across A Homo sapiens, B Drosophila melanogaster, C Arabidopsis thaliana, and D Saccharomyces cerevisiae. For each species, performance was evaluated on at least three independent DDA-MS files. Data are presented as mean ± SD. Individual data points from independent DDA-MS files are overlaid to show the distribution (n = 4 DDA-MS files for A–C; n = 5 for D). E The UpSet plot showing peptide identifications across six tools in the human dataset, including AlphaPept, AlphaPeptDeep, Sage, FragPipe (MSBooster), MS²Rescore, and DDA-BERT. Vertical bars represent the number of peptides shared by specific tool combinations, as indicated by the filled dots in the matrix below. Horizontal bars on the left show the total number of peptides identified by each tool. Source data are provided as a Source data file.

To examine protein-level evidence patterns, we compared protein group identifications at a 1% FDR threshold, stratified by the number of unique peptides supporting each protein group. Across all four species, DDA-BERT showed generally comparable or higher performance across most peptide support categories (Supplementary Fig. 4), without exhibiting systematic enrichment restricted to any single category. In the human dataset, DDA-BERT achieved protein group counts comparable to FragPipe (MSBooster) across all peptide support levels, while outperforming other methods in several categories. For D. melanogaster, DDA-BERT consistently identified the largest number of protein groups across most peptide support bins, including both low- and high-support categories. In A. thaliana, DDA-BERT yielded the highest protein group counts in the one- and two-peptide-supported categories and remained among the top-performing methods for higher-support protein groups. In S. cerevisiae, DDA-BERT and FragPipe (MSBooster) outperformed the other methods across all peptide-support categories except the single-peptide-supported protein groups, where AlphaPept yielded the highest counts.

Next, we compared peptide identifications obtained by DDA-BERT with those from five widely used, published tools using the human dataset (Fig. 4E). As expected, DDA-BERT yielded the highest number of peptide identifications among all evaluated methods, followed by FragPipe, whereas AlphaPept reported the fewest identifications. A substantial overlap was observed between DDA-BERT and other tools, with 14,080 peptides commonly identified across all methods. When restricting the comparison to FragPipe, Sage, and MS²Rescore, the number of peptides shared with DDA-BERT increased to 25,331. Beyond these shared identifications, at a 1% FDR threshold, DDA-BERT uniquely identified 3611 peptides that are not reported by any other tool. A similar trend was observed in the D. melanogaster, A. thaliana, and S. cerevisiae datasets (Supplementary Figs. 5, 6 and 7), where DDA-BERT consistently identified the largest number of peptides, with the majority of peptide identifications being supported by multiple tools.

Rescoring performance on single-cell proteomics

DDA-BERT was further evaluated on trace-amount HeLa cell proteomics data using two previously published datasets^25,26, with benchmarking against five widely used rescoring and search frameworks, including AlphaPept, AlphaPeptDeep, MS²Rescore, Sage, and FragPipe (MSBooster), at a 1% FDR threshold. In the first dataset²⁵, HeLa cell lysates were analyzed using solid-phase microextraction-assisted capillary zone electrophoresis-mass spectrometry, with injection volumes gradually increasing from 0.1 to 10 ng (approximately 0.4–40 HeLa cells). Across this dilution series, DDA-BERT generally achieved the highest identification yields at both the PSM level (Fig. 5A) and the peptide level (Fig. 5B), except at the highest input (~40 cells), where FragPipe slightly outperformed DDA-BERT. Notably, under ultra-low peptide input conditions (0.1 ng), DDA-BERT identified 432 PSMs, corresponding to a 10.93–1387.93% increase compared with the other evaluated tools. Consistent with this observation, we further evaluated DDA-BERT on an independent dataset derived from a single HeLa cell lysate, acquired on an Orbitrap Exploris 480 mass spectrometer using a 165-min chromatographic gradient²⁶. Across FDR thresholds ranging from 0.2 to 1%, DDA-BERT exhibited stable performance (Fig. 5C). At a 1% FDR threshold, the number of peptides identified by DDA-BERT was comparable to that obtained with FragPipe and exceeded those reported by MS²Rescore, Sage, AlphaPeptDeep, and AlphaPept by 2.78%, 10.43%, 147.23%, and 498.83%, respectively (Fig. 5D).

Fig. 5. Rescoring performance of different tools in trace proteomics using HeLa cell samples.

Bar plots showing the number of PSMs (A) and peptides (B) identified at 1% FDR by six tools: DDA-BERT, MS²Rescore, FragPipe (MSBooster), Sage, AlphaPept, and AlphaPeptDeep, across different sample inputs (0.4, 2, 4, and 40 HeLa cells). Bars represent the mean, and error bars indicate the standard deviation (SD). Individual data points from each benchmark file are overlaid to show the distribution (n = 2). C Line plot illustrating the average number of PSMs identified at varying FDR thresholds (0.002, 0.004, 0.006, 0.008, and 0.01) by each tool in the single-cell dataset. D Bar plot comparing the average number of peptides identified at 1% FDR by each tool in the single-cell dataset. Data are presented as mean ± SD, with individual data points overlaid across replicates (n = 4). Source data are provided as a Source data file.

Generalization of DDA-BERT to HLA immunopeptidomics

In addition to single-cell proteomics, we evaluated DDA-BERT on HLA immunopeptidomics data, a fundamentally distinct proteomics modality characterized by nonenzymatic peptide generation, variable peptide lengths, and the absence of predictable cleavage patterns. As shown in Supplementary Fig. 8, across multiple FDR thresholds, DDA-BERT generally identified more PSMs than all other tools, except for the most stringent 0.2% FDR setting. At the commonly used 1% FDR threshold, DDA-BERT increased PSM identifications by 7.63%, 36.74%, 11.50%, 62.76%, and 114.98% compared to FragPipe, Sage, MS²Rescore, AlphaPeptDeep, and AlphaPept, respectively. Under the same condition, DDA-BERT also yielded the highest numbers of peptide precursors (Supplementary Fig. 8B) and unique peptides (Supplementary Fig. 8C), with peptide-level gains of 4.14, 34.09, 7.45, 53.64, and 87.47% over the corresponding tools. These gains extended to the protein level, with DDA-BERT identifying the largest number of protein groups at 1% FDR (Supplementary Fig. 8D), indicating that increased sensitivity was accompanied by consistent protein-level support.

Assessing sequence memorization and potential overfitting to human data

Given that approximately 50% PSMs from the training dataset were derived from human samples, we performed a series of diagnostic analyses to assess whether DDA-BERT exhibits peptide sequence memorization or overfitting to human-specific data. First, we evaluated whether DDA-BERT introduces systematic scoring bias in the evaluation of peptides present in and absent from the training set. The original training samples were randomly divided into training (50%) and held-out (50%) sets, with overlapping peptide sequences removed to ensure complete sequence-level independence. DDA-BERT was then applied to score the human dataset. After excluding PSMs beyond 1.5× the interquartile range, target score distributions remained comparable between the two datasets, with only a minor reduction in the held-out set (Fig. 6A). When considering all PSMs without outlier exclusion, the mean target scores were 0.13 ± 0.32 and 0.16 ± 0.35 for the training and held-out sets, respectively, while decoy PSM scores were identical in both sets (0.003 ± 0.032). These results indicate no systematic score bias between the training and held-out sets. Second, we assessed whether target-decoy separation could emerge in the absence of informative spectral evidence. The m/z values in the spectra were randomly perturbed while preserving peptide sequence information. As shown in Fig. 6B, target and decoy score distributions became nearly indistinguishable, and no confident target identifications were obtained at a 1% FDR. These results indicate that effective discrimination by DDA-BERT critically depends on meaningful peptide-spectrum relationships rather than peptide sequence information alone.

Fig. 6. Diagnostic analyses of overfitting, sequence memorization, and FDP in DDA-BERT.

A Box plots comparing DDA-BERT scores for target and decoy PSMs between the training and held-out sets. For the target group, scores were derived from 656,891 PSMs in the training set and 657,484 in the held-out set. For the decoy group, 960,750 and 945,963 PSMs were utilized for the training and held-out sets, respectively. The data split was performed at the stripped peptide sequence level to ensure that the training and held-out sets share no common sequences. Target and decoy groups were defined using a concatenated database search, where decoy sequences were generated by reversing all target sequences prior to the search to ensure a balanced and pre-defined null distribution. Box plots show the median (center line) and interquartile range (box), with whiskers extending to 1.5× the interquartile range (IQR). Outliers beyond 1.5× IQR were removed for visualization. B Kernel density estimates of DDA-BERT scores for target and decoy PSMs after random perturbation of the m/z values in the spectra, while peptide sequence information was preserved. C Estimated peptide-level FDP as a function of the nominal FDR threshold for representative human benchmark files, using three FDP estimation strategies: the combined method, the lower-bound method, and the paired method, with entrapment sequences generated by FDRBench. The vertical dashed line indicates the 1% FDR threshold. Text annotations report the total number of peptide identifications and the corresponding estimated FDP at 1% FDR for each method. Source data are provided as a Source data file.

Robust peptide-level false discovery proportion control by DDA-BERT

To evaluate the peptide-level false discovery proportion (FDP) control of DDA-BERT, we followed the methodology of Wen et al.²⁷ and applied three estimation strategies: combined method, lower bound, and paired method (Fig. 6C). The evaluation was conducted using Jurkat cell lysate data²⁸ acquired on an Orbitrap Exploris 480 mass spectrometer with a 90-min chromatographic gradient, with entrapment sequences generated using the default FDRBench configuration. Across all analyses, the estimated FDP increased monotonically with increasing FDR thresholds. The combined and paired methods produced highly similar FDP curves across all runs, whereas the lower-bound method consistently yielded more conservative estimates. At the nominal 1% FDR threshold (vertical dashed line), FDP estimates from the combined and paired methods ranged from approximately 0.90% to 0.99% and 0.88% to 0.97%, respectively, while the lower-bound method produced lower values in the range of 0.45–0.50%. Importantly, across all files and estimation strategies, the observed FDPs remained at or below the nominal 1% level, indicating that DDA-BERT maintains effective control of false discoveries and exhibits robust behavior under stringent error-rate constraints.

Discussion

Recent successes of transformer-based models in de novo peptide sequencing^18,29,30 motivated the development of DDA-BERT as an end-to-end rescoring framework for DDA proteomics. Unlike existing data-driven rescoring pipelines, such as MS²Rescore¹², Oktoberfest³¹, AlphaPeptDeep⁷, and MSBooster¹⁵, which combine handcrafted features with property-specific predictors, DDA-BERT performs joint representation learning over peptide sequences and full MS/MS spectra within a unified architecture. This design allows the model to capture global and contextual dependencies that extend beyond individual predicted properties, enabling more flexible and data-adaptive scoring behavior.

A central algorithmic feature of DDA-BERT is its use of contiguous-span (block-wise) masking during training. In contrast to token-wise random masking, span-level masking substantially increases the difficulty of reconstructing or memorizing complete peptide sequences, as the model can no longer rely on a few unmasked local tokens or short-range co-occurrence patterns to infer the masked content. Instead, it is forced to integrate broader sequence context together with consistent spectral evidence. This design effectively reduces the likelihood of shortcut learning based on sequence memorization. When combined with spectrum dropout, the model is trained under simultaneous sequence uncertainty and spectral incompleteness, encouraging it to rely on stable peptide-spectrum correspondences rather than local sequence motifs, individual fragment peaks, or dataset-specific artifacts.

It is important to place DDA-BERT in the context of existing per-file or per-project rescoring workflows, which remain highly effective and widely used in practice. Such approaches offer several advantages, including flexible deployment at the file or project level, efficient model fitting on dataset-specific PSMs, relatively low computational overhead, and robust performance in CPU-only environments. Moreover, modern rescoring pipelines often leverage deep learning-based predictors of peptide physicochemical properties, including RT and fragment ion intensities, that are trained once on large and diverse corpora and exhibit strong cross-dataset generalization. At the same time, rescoring models trained exclusively on PSMs observed within a single dataset or project are inherently constrained by the size, quality, and diversity of that dataset. While project-level training can yield well-calibrated models for large and homogeneous studies, performance may vary when substantial heterogeneity exists across runs, arising from differences in instrumentation, fragmentation methods, acquisition settings, or sample complexity. In such cases, information learned in one dataset cannot be readily transferred to others, potentially limiting sensitivity for less abundant peptides or spectra with atypical fragmentation patterns.

DDA-BERT is motivated by the complementary idea that cross-project training on large-scale DDA data can raise the performance ceiling of rescoring models. By integrating spectra from multiple species, instruments, and experimental conditions into a unified training corpus, DDA-BERT learns more generalizable peptide-spectrum matching representations that can be applied across datasets. Importantly, DDA-BERT is not intended to replace per-file or per-project workflows, but rather to complement them by providing a scoring function that encodes broader proteomic knowledge and can be directly applied to diverse search outputs. Although the current version of DDA-BERT is trained on a finite collection of DDA-MS datasets, spanning multiple species and platforms, we observe improved identification performance compared to alternative rescoring strategies across multiple independent benchmarks. These gains are achieved without compromising error control, suggesting that the learned decision boundary remains stable under stringent FDR constraints. As larger and more diverse DDA datasets become available, the cross-project training paradigm underlying DDA-BERT is expected to further improve.

However, DDA-BERT does have several limitations. As a deep learning-based rescoring model operating on full MS/MS spectra, it incurs higher computational overhead than lightweight feature-based tools, such as MSBooster or MS²Rescore. We benchmarked its end-to-end performance on a workstation equipped with an NVIDIA A100 (40GB) GPU and 128 GB of RAM. Using a DDA file containing ~23,000 spectra as an example, the complete workflow, including database search (FragPipe (MSBooster), Sage, and AlphaPept), PSM rescoring, and protein inference, was completed in approximately 29.2 min with a single GPU and 20 CPU cores, of which the rescoring step itself required < 2 min. We also observed reduced scoring confidence for peptides longer than ~35 amino acids, consistent with prior observations³² that transformer architectures may be less effective for very long sequences^33,34. In addition, DDA-BERT currently does not explicitly incorporate predicted spectral features (e.g., fragment m/z and intensity) in the way that MSBooster-like approaches do. In the future, one may explore auxiliary learning objectives, such as spectrum intensity prediction, to improve learning efficiency and interpretability while preserving end-to-end modeling. Although our analyses did not reveal evidence of systematic score bias between training and held-out sets, and target-decoy separation collapsed when informative spectral evidence was removed, these results do not fully exclude the possibility of more subtle forms of bias. In particular, end-to-end models may capture higher-order statistical patterns that are not easily detected by standard validation strategies. Therefore, continued development of rigorous evaluation frameworks is necessary to assess potential overfitting, decoy-related artifacts, and other failure modes. As a rescoring method, DDA-BERT is also bounded by the quality of candidate PSMs generated by upstream database search engines, motivating future integration of search and rescoring. It is also necessary to further evaluate the generalizability of DDA-BERT across additional fragmentation strategies (e.g., trap-type CID, EAD, and ETD).

DDA-BERT provides a generalizable and structurally extensible framework for PSM rescoring. By combining cross-project training, structured multimodal regularization, and end-to-end representation learning, it complements existing rescoring paradigms and offers a methodological foundation for scalable, AI-driven peptide identification in DDA proteomics. DDA-BERT, together with a few other software tools of its kind published recently, heralds the emerging era of end-to-end training models for MS-based proteomics data analysis.

Methods

Dataset preparation

The DDA-BERT model was trained on 12,285 DDA-MS raw files collected from 110 independent projects, spanning 11 species and acquired using 10 different models of high-resolution mass spectrometers (Supplementary Data 1). An independent immunopeptidomics (HLA) dataset (accession number: PXD019643), comprising 190 raw files, was used as a validation set during training to assess generalization performance, guide model selection, and mitigate overfitting. Although data from the same project were also included in the training pool, the raw files used for validation were entirely distinct from those used for training. In addition, to ensure strict independence between the training and validation sets, cleaned peptide sequences present in the training set were explicitly excluded from the validation set, ensuring no peptide-level overlap between the two subsets.

Database searching

Database searches were conducted using Sage v0.14.6³⁵, FragPipe v21.1¹⁶ (with MSBooster enabled), and AlphaPept v0.5.3³⁶ under consistent parameter settings. Prior to database searching, raw MS/MS files were converted to mzML format using the msconvert GUI (ProteoWizard, v3.0.19014)³⁷ for all tools except AlphaPept, which performed database searching directly on the raw files.

For conventional tryptic proteomics datasets, searches were performed with a peptide length range of 7–50 amino acids, a precursor mass tolerance of 20 ppm, and trypsin digestion allowing up to two missed cleavages. Variable modifications included methionine oxidation, N-terminal acetylation, and deamidation of asparagine and glutamine, while carbamidomethylation of cysteine was specified as a fixed modification.

For HLA datasets, enzyme specificity was set to nonspecific, with peptide lengths restricted to 8–15 amino acids. The same workflows and parameter settings were consistently applied to both the training and validation HLA datasets. For all searches, up to ten candidate PSMs were retained per MS/MS scan for downstream processing.

Search configuration

FragPipe (MSBooster)

Database searches were performed using MSFragger with the MSBooster module enabled to predict peptide RT, IM, and MS/MS fragment ion intensities. For conventional tryptic proteomics datasets, search parameters were set as follows: peptide length range of 7–50 amino acids, precursor mass tolerance of 20 ppm, and tryptic digestion allowing up to two missed cleavages. Variable modifications included methionine oxidation, N-terminal acetylation, and deamidation of asparagine and glutamine, while carbamidomethylation of cysteine was specified as a fixed modification. All other search parameters were kept at the default settings of FragPipe. For HLA immunopeptidomics data, database searches were conducted using the nonspecific HLA-C57 workflow provided by FragPipe, with peptide lengths restricted to 8–15 amino acids, while all remaining parameters followed the default HLA-C57 configuration. For result reporting, FragPipe was configured to retain the top 10 highest-scoring PSMs for each MS/MS scan and to output the corresponding search scores.

AlphaPept

Searches were performed using the same peptide length range, precursor mass tolerance, digestion rules, and modification settings as described above to ensure parameter consistency across tools. Custom scripts were used to extract initial database search results from the first search top-level group in the “ms_data.hdf” file, which contains PSM outputs, including matched peptide sequences, numbers of matched fragment ions, RT, and other metrics. The “ids.csv” file from the search results was then processed with custom scripts to calculate FDR and generate PSM-, precursor-, and peptide-level identifications for downstream statistical analyses.

Sage

Database searches with Sage were conducted using the same parameter settings as for FragPipe and AlphaPept to ensure comparability. These included peptide length range, precursor mass tolerance, digestion rules, modification settings, and the number of candidate PSMs reported per MS/MS scan. All identifications were processed using custom scripts following the same criteria applied to the other tools.

AlphaPeptDeep

AlphaPeptDeep v1.3.1 extracts candidate PSM lists from the raw database search results of AlphaPept and employs deep learning models to predict MS/MS spectra and RT. These predictions are used to compute additional features, which are then combined with the original search scores and processed with the post-search rescoring tool Percolator to improve PSM ranking and enhance identification performance.

MS²Rescore

MS²Rescore v3.1.5 was applied to the raw search outputs from FragPipe, Sage, and AlphaPept. The standard processing workflow was followed, in which a set of engineered features, such as spectral matching scores, mass errors, and RT deviations, were computed and subsequently fed into the default rescoring framework, mokapot¹³, to optimize PSM ranking and improve identification performance.

Training data construction and negative sampling

To construct reliable positive training examples, target PSMs identified at a 1% FDR threshold were first extracted. Only PSMs present in the complete Sage search results were retained for model training, as the predicted RTs were directly obtained from Sage. An additional empirical curation step based on RT was then applied: PSMs were ranked according to the difference between experimental and predicted retention times (ΔRT), and the top 2% with the largest deviations were excluded to remove potentially lower-confidence matches.

To improve model generalization and reduce overfitting to dataset-specific patterns, we constructed a diverse and challenging training dataset by integrating multiple sources of negative samples, including decoys and false targets. Decoy-based negatives were derived from two complementary sources. First, high-scoring decoy PSMs (score > 0.2) generated in earlier iterations of DDA-BERT were included as difficult negatives near the decision boundary. Second, for each training file, peptide sequences were independently shuffled to generate file-specific decoy sequences, which were subsequently rescored using a previous model version. Because shuffling was performed independently for each file, this strategy produced unique, non-repetitive decoy sets across datasets, preventing the model from exploiting shared decoy templates. From these candidates, both high-confidence and low-confidence decoys were selectively retained to maintain diversity and balance within the training distribution.

To further challenge the classifier, we introduced false target samples as hard negatives. These were peptide sequences reported by Sage or FragPipe at a nominal 1% FDR but excluded from the final filtered PSM lists because they were not the top-ranked matches for the corresponding MS/MS spectra. Although identical to true targets at the sequence level, they lack supporting spectral evidence and therefore constitute incorrect matches at the PSM level, making them effective hard negatives for shaping the decision boundary. False targets with relatively high model scores were preferentially retained, with their number capped at approximately 20% of the true target set. For cases in which a false target shared the same modified sequence as a true target, the deltaRT value was inherited from the corresponding true target; otherwise, it was sampled from a beta (1, 5) distribution.

Model architecture

DDA-BERT comprises two interconnected transformer encoders: a spectra encoder for processing peak m/z and intensity values from MS2 scans, and a peptide precursor encoder that processes peptide sequences along with precursor-level features, including precursor m/z, charge state, predicted RT, and ΔRT (Fig. 1C). During training, the predicted RT and ΔRT are obtained from the Sage search results. During inference, the predicted RT is derived from both Sage and AlphaPeptDeep⁷ (for peptides uniquely identified by AlphaPept). Following the approach used in Casanovo¹⁸, sinusoidal position embeddings and learned linear layers were used to encode MS2 m/z values and intensities, while learned embedding layers were applied to encode the peptide precursor’s m/z and charge. The sum of the learned embedding of each amino acid and its position embedding is concatenated to encode the amino acid sequence. Causal masking was not applied in the peptide precursor encoder, as DDA-BERT is not designed for autoregressive sequence generation. Instead, cross-attention layers were introduced between the two encoders, enabling context-aware integration of spectral and precursor-level information and facilitating the alignment and correlation of relevant features across modalities.

Model training, regularization, and optimization

To enhance robustness and promote generalization across heterogeneous spectra, we introduced joint regularization on both the spectral and sequence modalities during training. On the spectral side, spectrum dropout was employed, in which 10% of the (m/z, intensity) peaks in each spectrum were randomly removed to simulate realistic experimental noise and missing fragment ions commonly observed in DDA data. To avoid introducing systematic bias, 10% of spectra in each batch were left unaltered.

In parallel, contiguous-span masked language modeling (MLM) was applied to peptide token sequences. For each selected sequence, one to three nonoverlapping masked spans were randomly generated, together covering approximately 20% of the non-padding tokens. The masking procedure was applied uniformly to both target and decoy peptide sequences, without distinction. This strategy compels the model to capture contextual dependencies between peptide sequences and their fragmentation patterns, rather than relying on surrounding token patterns. By coupling spectrum dropout with span-level MLM, each training iteration exposes the model to both spectral incompleteness and sequence uncertainty. This dual corruption scheme serves as an effective multimodal regularizer, mitigating overfitting to dataset-specific spectral artifacts and improving generalization across diverse proteomic conditions.

All experiments were conducted using a transformer architecture comprising nine layers, an embedding dimension of 768, and sixteen attention heads, totaling approximately 95 million parameters. Training was performed with a batch size of 128 spectra, a weight decay of 5 × 10⁻⁵, and a peak learning rate of 5 × 10⁻⁵, scheduled with a linear warm-up over 150,000 steps followed by cosine decay. The model was trained for 30 epochs on 16 NVIDIA A100 GPUs, completing in approximately 7 days. The checkpoint corresponding to the lowest validation loss was selected for downstream evaluation.

Metrics and benchmarks

We performed a systematic comparison of DDA-BERT with five state-of-the-art published rescoring tools or search frameworks, using eight publicly available datasets. These benchmark datasets span multiple species and diverse experimental conditions, allowing for a comprehensive evaluation of each method’s performance across diverse analytical scenarios. Detailed descriptions of the datasets are provided below:

Dataset 1: lysates from Jurkat cells were analyzed using an Orbitrap Exploris 480 mass spectrometer with a 90-min chromatographic gradient. The raw MS/MS files are available on PRIDE (accession number: PXD036024)²⁸.

Dataset 2: formalin-fixed, paraffin-embedded (FFPE) tissue samples from colorectal cancer patients were analyzed using an Orbitrap Exploris 480 mass spectrometer with a 120-min gradient. The dataset is accessible via PRIDE (accession number: PXD041625)²².

Dataset 3: proteomics data for samples containing 0.4, 2, 4, and 40 cells were acquired using a Q Exactive HF mass spectrometer. This dataset can be accessed through PRIDE (accession number: PXD047627)²⁵.

Dataset 4: single HeLa cell lysate analyzed using an Orbitrap Exploris 480 mass spectrometer. The dataset can be accessed on PRIDE (accession number: PXD041879)²⁶.

Dataset 5: yeast proteomics data acquired by nanoLC-MS/MS in DDA mode with HCD fragmentation on an Orbitrap Exploris 480 instrument. This dataset is available via PRIDE (accession number: PXD025934)³⁸.

Dataset 6: Arabidopsis seedling peptides from 16 biological samples were pooled and fractionated into 10 fractions before analysis on a Q Exactive HF instrument. The dataset is accessible through iProx (accession number: IPX0003458001)²⁴.

Dataset 7: Drosophila samples were analyzed on a Q Exactive HF mass spectrometer using a 50-min LC gradient in DDA mode. The dataset is available via PRIDE (accession number: PXD040158)²³.

Dataset 8: HLA class I immunopeptidomics data from the metastatic melanoma patient Mel15 were retrieved from the PRIDE repository (accession number: PXD004894)³⁹. Data were acquired on a Q Exactive mass spectrometer using higher-energy collisional dissociation (HCD) fragmentation, and database searches were performed against the reference FASTA provided by Yang et al.¹⁵.

FDR control and statistics

FDRs at the PSM, precursor, peptide, and protein group levels were controlled at 1% FDR using a target-decoy strategy. FDR estimation was performed independently for each LC–MS/MS run (run-wise FDR). Specifically, for each run, all candidate PSMs, including decoy matches, were ranked by score, and the score threshold corresponding to a 1% FDR was determined. Target matches exceeding this threshold were retained as the final run-level identifications. Final identifications at the precursor, peptide, and protein group levels were obtained using a unified FDR control script together with a standardized protein inference script. Protein inference was performed following the method described by Mann et al. in AlphaPept³⁶ to obtain the final protein identifications. This unified post-processing workflow was applied consistently across all datasets and tools, ensuring comparability of results and reproducibility of the analysis.

Reporting summary

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

Author contributions

T.G., Y.C. conceptualized the research and provided oversight. Y.C. designed the methodology and laid the foundation for research implementation. J.A., P.L., and J.L. managed the overall experimental pipeline. Z.N. developed the data preprocessing component and the DDA-BERT software system. J.A., J. Lin, X.Z., Y.Z., Y.S., and Z.X. curated the data. J.A., P.L., J. Lin, J. Liu, Z.Y., and X.Z. conducted model parameter tuning and performed analyses. J.A. designed the benchmark experiments and drafted the initial manuscript with contributions from J. Lin, P.L., and X.Z., J.A., Y.C., Y.S., and P.L. revised the manuscript. All authors have reviewed and T.G. approved the final version of the manuscript. T.G. acquired funding and supervised the project.

Peer review

Peer review information

Nature Communications thanks Ralf Gabriels and the other anonymous reviewer(s) for their contribution to the peer review of this work. A peer review file is available.

Data availability

The training data used in this study were primarily obtained from the ProteomeXchange Consortium PRIDE repository⁴⁰ and the iProX database^41,42, with a small portion of the data generated internally by our Lab. Accession numbers, file names, species information, and assignment to the training or validation sets for all datasets are summarized in Supplementary Data 1. To ensure data reliability, consistency, and reproducibility, only datasets with complete metadata and well-defined experimental conditions were included. The inclusion criteria for the training data were as follows: (i) all data were acquired from raw mass spectrometry files generated using data-dependent acquisition (DDA) mode; (ii) all experiments followed a label-free experimental design; and (iii) data were acquired using Q Exactive or newer instruments (e.g., Q Exactive HF-X, Orbitrap Exploris 480). To further increase peptide sequence diversity in the training set, 450 raw files from a human leukocyte antigen (HLA) immunopeptidomics dataset (accession number: PXD019643)⁴³ were additionally incorporated into the training data. After standardized preprocessing, all datasets were integrated to construct a high-quality training resource comprising more than 270 million PSMs. The validation set comprised 190 HLA immunopeptidomics raw files (accession number: PXD019643)⁴³. To ensure strict independence, stripped peptide sequences present in the training set were excluded from the validation set, resulting in no peptide-level overlap between the two datasets. All datasets used for DDA-BERT training and for validation during model selection have been deposited in the PRIDE repository under accession number PXD072999. The dataset is provided in a preprocessed.pkl format, consistent with the input used during model development, to facilitate reproducibility and further methodological investigation, and will be made publicly available upon acceptance of the manuscript. Benchmark datasets used for performance evaluation are also publicly available under the following accession numbers: PXD021559, PXD036024, PXD041625, PXD047627, PXD040158, PXD041879, PXD025934, and iProX: IPX0003458001. All Source data supporting the plots and results presented in this paper are available in the Zenodo repository at https://zenodo.org/records/19020927. The raw benchmarking outputs for all algorithms are archived in Zenodo (10.5281/zenodo.18907066) as Supplementary Data 2, which provides the underlying data for the Source data files.

Code availability

The DDA-BERT source code, model weights, and containerized implementation are available on GitHub (https://github.com/guomics-lab/DDA-BERT), with a permanent version archived on Zenodo (10.5281/zenodo.18973835)⁴⁴. A Docker image for reproducible deployment is provided at https://hub.docker.com/r/guomics2017/dda-bert, and an executable version is accessible via https://guomics.com/software/dda-bert. Inference requires GPU support for bfloat16 (bf16) precision (e.g., NVIDIA A100). The code is distributed under a custom license restricting commercial use and redistribution.

Competing interests

T.G. is the shareholder of Westlake Omics Inc. P.L. is the employee of Westlake Omics Inc. The remaining authors declare no competing interests.

Supplementary information

The online version contains Supplementary material available at 10.1038/s41467-026-72246-6.