DiNovo enables high-coverage and high-confidence de novo peptide sequencing via mirror proteases and deep learning

Protein sample preparation and LC-MS/MS analysis

Proteins of E. coli or yeast strains were extracted using a lysis buffer containing 8 M urea, 150 mM NaCl, 50 mM Tris-HCl (pH 7.5), 1 mM phenylmethysulfonyl fluoride (PMSF), 1 mM 2-chloroacetamide, and 1× cocktail (Roche, 11697498001). The proteins were reduced with 5 mM DTT at 45 °C for 30 min, and then fully alkylated with 15 mM IAA at room temperature in the dark for 30 min. Samples were processed by in-solution digestion or the FASP method as previously described, with slight modifications^33,34 After reducing the urea concentration in the lysate to <1 M by dilution or displacement, the proteins were aliquoted into four portions and digested with different proteases at an enzyme-to-protein ratio of 1:50. For trypsin/Lys-C or Lys-C alone digestion, the sample was solubilized in a buffer of 50 mM ABC (pH 8.3), and trypsin and/or Lys-C were added for overnight digestion at 37 °C. For LysargiNase/Lys-N digestion, the sample was solubilized in a buffer of 20 mM HEPES, 10 mM CaCl2, and 1 mM Zn(Ac)2 (pH 8.3); after digestion with LysargiNase at 37 °C for 4 h, Lys-N was added for overnight digestion. For Lys-N alone digestion, the sample was solubilized in a buffer of 50 mM ABC and 1 mM Zn(Ac)2 (pH 7.5–8.5), and Lys-N was added for overnight digestion at 37 °C. All of trypsin, LysargiNase, Lys-C and Lys-N were provided by Enzyme & Spectrum (Beijing, China)^23,35,36,37. After digestion, 1% formic acid (FA) was added to stop the reaction. The digested peptides were desalted using an optimized StageTip³⁸.

LC-MS/MS analysis was conducted using an Orbitrap Q-Exactive HF mass spectrometer (Thermo Fisher Scientific, USA) coupled with an EASY-nLC 1000 system (Thermo Fisher Scientific, USA). The resulting peptides were resolved in buffer containing 1% acetonitrile and 1% FA, then loaded onto a 75 μm I.D. ×20 cm LC column packed with 1.9 μm C₁₈ reverse phase packing particles (Dr. Maisch GmbH, Germany). LC separation was performed with a 120 min gradient from 6% to 45% buffer B (80% acetonitrile with 0.1% FA) at a flow rate of 300 nL/min. MS measurements were performed in data-dependent acquisition mode. For trypsin/Lys-C digested peptides, MS1 spectra were acquired with a survey scan (300–1400 m/z) at a resolution of 60,000 at 200 m/z, with an AGC target of 3e6 and a maximum injection time (MIT) of 30 ms. Precursors with an intensity >3.3e4 were selected for MS2 analysis, collected at an MIT of 60 ms or an AGC target of 5e4, then fragmented at a normalized collision energy (NCE) of 27. The resulting fragment ions were analyzed in the Orbitrap at a resolution of 15,000 at 400 m/z. For LysargiNase/Lys-N digested peptides, MS1 spectra were acquired with the same survey scan (300–1400 m/z) at a resolution of 60,000 at 200 m/z, an AGC target of 3e6 and an MIT of 50 ms. Precursors with an intensity >1.3e4 were selected for MS2 analysis, collected with an MIT of 150 ms or an AGC target of 1e5, then fragmented with an NCE of 27. The resulting fragment ions were analyzed in the Orbitrap at a resolution of 15,000 at 400 m/z.

Datasets and preprocessing

The LC-MS/MS analysis of the E. coli and yeast proteins digested by four proteases resulted in eight datasets of MS2 spectra. The number of spectra in each dataset is shown in Supplementary Table 3. In addition, we conducted a database search using pFind v3.2.0 and calculated the pairing rate of the dataset at both spectrum and peptide levels (Supplementary Table 3). For simplicity, we refer to trypsin/Lys-C mixed digestion as trypsin digestion, and LysargiNase/Lys-N mixed digestion as LysargiNase digestion.

Centroided spectra were extracted in Mascot generic format from RAW files using pParse³⁹. In order to better recognize the mirror spectra, all mass spectra were preprocessed with the following steps. First, we retained the most abundant N peaks (\(N=200\) by default) for each spectrum. Second, we detected isotopic clusters and transformed each cluster into a singly charged monoisotopic peak with intensities combined. To enhance the signals of backbone fragment ions, we also accumulated the intensities from their isotopic and neutral-loss (−H₂O and −NH₃) peaks. Third, we normalized the peak intensities and then took their square roots to reduce variance effects. Fourth, we performed local denoising by retaining the most abundant K peaks (\(K=7\) by default) per mass bin (with a default bin width of 100 Da). Additionally, we removed peaks corresponding to precursor ions and immonium ions. At last, to alleviate the problem of imbalanced ion types, we generated a complementary peak for each peak after denoising, while maintaining the same ion intensity.

It should be noted that these preprocessing steps were applied only to the mirror-spectra recognition process in DiNovo and not to the other algorithms in order to avoid any unexpected influence on their performance.

Mirror-spectra recognition algorithm

In DiNovo, we developed a novel mirror-spectra recognition algorithm, named MirrorFinder, which is free of prior peptide sequencing from separate spectra as done in pNovoM²³. The algorithm directly utilizes the information of precursor and fragment ions to recognize spectra generated from mirror peptides. To be specific, it first selects candidate spectral pairs whose precursor masses comply with the rule of mirror peptides (as outlined in Supplementary Table 4) and whose retention time difference falls in a reasonable range. The rule states that each type of mirror peptides has a theoretical precursor mass difference and two theoretical fragment ion mass differences. Each candidate spectral pair is then assigned a matching score, reflecting the confidence level that they are indeed a pair of mirror spectra. Notably, the parameters for preprocessing and mirror-spectra recognition were optimized through extensive testing to ensure better performance, as detailed in Supplementary Note 3. Furthermore, the results in Supplementary Note 3 show that MirrorFinder was more accurate than pMerge, the undisclosed mirror-spectra recognition algorithm used in pNovoM.

To begin with, we select candidate spectral pairs with precursor mass differences matching the theoretical values (as listed in Supplementary Table 4) within a specified tolerance. We set the tolerance to 20 ppm, and calculate the maximum and minimum values of precursor mass difference, which are

where \({m}_{1}\) and \({m}_{2}\) are the precursor masses of the two spectra, respectively. If one of the theoretical precursor mass differences in Supplementary Table 4 falls within the interval \(\left[{D}_{2},{D}_{1}\right]\), then the spectral pair is a candidate of a certain category (e.g., A-G).

After that, each candidate spectral pair is assigned a matching score, which is calculated through the following steps (as illustrated in Fig. 1b). First, a fixed mass range (−200 to +200 Da in this work) is divided into equal small intervals (e.g., 0.02 Da), with the total number of intervals denoted as \(N\). Next, for the \(i\)-th interval, the number of fragment ion pairs whose mass differences fall in this interval is counted, and their intensities are summed, denoted as \({{\mbox{count}}}_{i}\) and \({{\mbox{sum\_inten}}}_{i}\) (\(i=1,\cdots,N\)), respectively. Then, a statistic \({s}_{i}\) is calculated for the \(i\)-th interval by multiplying the number of pairs by the summed intensity, which is

Finally, two Bonferroni-corrected \(p\)-values are estimated for the intervals corresponding to two theoretical fragment ion mass differences for a certain category (as shown in Supplementary Table 4) from the distribution of the observed statistics⁴⁰. The smaller of the two \(p\)-values is denoted as \({p}^{*}\), and the matching score \(S\) is obtained by

Obviously, a higher matching score indicates stronger evidence that the spectral pair is a pair of mirror spectra.

Subsequently, we use a target-decoy strategy^25,26 to filter spectral pairs according to their matching scores. In MirrorFinder, the matching score is closely related to the theoretical fragment ion mass differences, allowing us to directly use the score derived from a random mass difference as the decoy score to compete with the corresponding original (target) score. For each spectral pair, if the original score is higher than the decoy score, the spectral pair is labeled as a target spectral pair, and its final score is set as the original score. Otherwise, it is labeled as a decoy spectral pair, and its final score is set as the decoy score. Then the final scores are sorted in descending order, and the FDR of target spectral pairs with final scores above a given threshold t can be estimated as

where T(t) and D(t) are the numbers of target and decoy spectral pairs, respectively. Through this strategy, we can filter target spectral pairs and control the FDR under a specified level, e.g., 2%. This approach eliminates the need to construct additional decoy spectra and offers a reasonable quality control standard for the recognized mirror spectra. We also verified that our quality control method satisfies the assumption of the target-decoy strategy in Supplementary Note 4.

DNN-based de novo sequencing algorithm for mirror spectra

We proposed a DNN-based de novo sequencing algorithm called MirrorNovo, which fully leverages complementary fragment ions in mirror spectra without the need to explicitly merge the spectra. In MirrorNovo, we preprocess each spectrum in a mirror spectral pair individually, padding the peak list with zeros if there are fewer than the top N peaks. To improve computational efficiency and reduce intensity variance, we convert the m/z values of all peaks to neutral masses and normalize their square-root-transformed intensities to the base peak.

As illustrated in Fig. 1c, the input to the MirrorNovo model consists of two peak-ion matching matrices, each corresponding to one spectrum in a mirror spectral pair, along with the category of the spectral pair. We compute the two matrices by comparing the observed peak masses in each spectrum to 18 types of theoretical fragment ion masses (i.e., singly and doubly charged b, y, and a ions, their singly charged neutral-loss ions, and 6 types of internal ions), and then concatenate them along the peak dimension. Theoretical fragment ions are generated by enumerating all possible subsequent amino acids based on the already predicted prefix sequence. To ensure meaningful matrix concatenation, we align the two matrices according to the corresponding fragmentation sites in the overlapping regions of the mirror peptides. For example, when iteratively predicting a trypsin-digested peptide of length L with either lysine or arginine (identified by MirrorFinder) at its C-terminal end, we align its b_i (y_L–i) fragments with the b_i + 1 (y_L–i-1) fragments derived from the LysargiNase-digested peptide, which has either lysine or arginine at its N-terminal end.

MirrorNovo employs a T-Net architecture¹³ with three one-dimensional convolutional layers to extract peak features from the concatenated matching matrices. A subsequent GRU layer captures both short- and long-term dependencies across peaks while maintaining the dimensionality of the input and output matrices. Next, three fully connected (FC) layers are utilized to learn higher-level feature representations, with batch normalization applied between them to enhance training stability and accelerate convergence. The Rectified Linear Unit (ReLU) activation function introduces non-linearity to prevent the collapse of adjacent linear layers. Finally, a softmax layer outputs probabilities for 20 amino acids, where isoleucine and leucine are not distinguished.

MirrorNovo predicts complete peptide sequences from mirror spectral pairs through iterative amino acid inference. At each iteration, it employs a beam search strategy to avoid local optima and utilizes a knapsack algorithm to reduce the number of candidate amino acids. A bidirectional prediction strategy is integrated to mitigate the loss of correct sequences. The iteration terminates when the difference between the prefix mass and precursor mass falls within 20 ppm. The top \(M\) (\(M=10\) by default) candidate peptides are retained based on peptide probability scores, calculated as the average of individual amino acid probabilities. To enhance data utilization, DiNovo also includes a single-protease de novo sequencing algorithm, GCNovo (the updated version of Denovo-GCN¹⁸), which allows peptide sequencing from spectra that cannot form mirror spectral pairs. As an independent sequencing algorithm, the running speed of GCNovo is similar to that of PointNovo and Casanovo, while GraphNovo and pNovo3 have significantly faster speeds (Supplementary Fig. 5a). To process the sequencing results, we first resolved conflicts where a single spectrum was paired with different spectra to produce different sequences. After assigning a unique mirror sequence to each spectrum, we further filtered out results where the mirror sequence was inconsistent with the sequence derived from the single spectrum.

In this study, MirrorNovo was trained on three high-resolution tandem mass spectrometry datasets from protein samples of Vero, MC2155, and human testis. Annotated mirror spectral pairs were constructed based on pFind (v3.2.0) search results at 1% FDR. Since a pair of mirror peptides may correspond to multiple mirror spectral pairs, and to ensure high-quality PSMs and sequence diversity, we ranked the mirror spectral pairs by pFind’s final scores and selected up to the top \(K\) highest-scoring pairs. Only spectral pairs of types A, B, and C with charge states of 2+ or 3+ were retained for the final annotated dataset. The dataset was split into training and validation sets at a 9:1 ratio based on peptide sequences (as detailed in Supplementary Table 5), ensuring no sequence overlap between the sets. The single-protease sequencing model, GCNovo, was trained on all trypsin- and LysargiNase-digested PSMs from the same dataset.

Graph-based de novo sequencing algorithm for mirror spectra

In addition to the DNN-based sequencing algorithm MirrorNovo, we also developed a graph-based sequencing algorithm, pNovoM2, which generates candidate peptides through spectrum graphs. Building on pNovoM²³, we have enhanced the performance of peptide sequencing and expanded the types of mirror peptides that can be sequenced, enabling pNovoM2 to deliver more results than pNovoM. Furthermore, we have optimized the speed and memory management of pNovoM2, facilitating rapid de novo peptide sequencing on large-scale datasets without the need for GPU support.

Supplementary Fig. 1 illustrates the schematic diagram of pNovoM2. Unlike pNovoM, pNovoM2 interprets not only the merged mirror spectra but also the separate spectra. This strategy allows pNovoM2 to sequence peptides even when MirrorFinder fails to find the mirror spectral pairs. In the workflow of pNovoM2, all spectra are first preprocessed, with each generating a spectrum graph similar to pNovo⁸. If MirrorFinder recognizes the corresponding mirror spectrum for the input spectrum, the spectrum graphs of the two spectra will be merged and used for subsequent analysis. Otherwise, the spectrum graph of the input spectrum will be used directly. For the merged spectrum graph, an edge is created between two nodes if their mass difference matches the total mass of any two or three amino acids. However, for the spectrum graph from a separate spectrum, the maximum number of amino acids is increased to five, due to the relatively lower abundance of signal peaks. Our scoring function aligns with that of pNovo, but we’ve made slight adjustments to better accommodate the merged spectrum graph. If the mass of a node in one spectrum graph differs from the mass of a node in another spectrum graph by a specific value, such as the mass of the amino acid K or R in the case of trypsin and LysargiNase, the peaks corresponding to these nodes are likely signal peaks of fragment ions. These nodes are then assigned a higher score. The pDag⁴¹ algorithm is subsequently used to identify the top \(N\) highest-scoring paths from the source node to the target node. For the merged spectrum graph, the value of \(N\) is set to 20; for the spectrum graph from a separate spectrum, it is set to 40. All peptides corresponding to these paths from the spectrum graphs are rescored together, and a final list of candidate peptides is generated.

De novo sequencing results evaluation

In this paper, we invented a simple yet effective method for evaluating de novo sequencing results, called TD mapping (as illustrated in Fig. 2a). This method requires the existence of a sequence database of proteins to be sequenced, but eliminates the need for peptide identification by traditional database search. Through the introduction of a decoy database, it addresses the lack of quality control in existing mapping-based methods¹⁶.

Building on the TD method^25,26 we first generate a decoy protein database by simply reversing or shuffling the sequences in the original database. Then, we map all the de novo sequences separately to the protein sequences in the target and decoy databases. Those successfully mapped peptides are retained and subjected to subsequent analysis. Unmatched peptides are excluded, as well as those mapped to both target and decoy sequences. Next, mapped peptides are ordered according to their masses, and the FDR of target peptides with masses above a threshold m is estimated as

where T(m) and D(m) are the numbers of target and decoy peptides, respectively. Finally, an appropriate mass threshold is selected so that the FDR is controlled under the desired level.

Configuration of de novo sequencing and database search

To fairly compare the performance of de novo sequencing software tools, we configured different tools with the same parameters. In addition, to perform a horizontal comparison between de novo sequencing and database search, all raw files were analyzed using pFind, MSFragger, and MS-GF+ for peptide identification. Both precursor and fragment ion mass tolerances were set to 20 ppm. Carbamidomethyl of cysteine (C) was specified as a fixed modification, and oxidation of methionine (M) was considered as a variable modification. The detailed configurations for each software tool are as follows:

• DiNovo: We offer two optional de novo sequencing algorithms in DiNovo, i.e., MirrorNovo and pNovoM2. By default, MirrorNovo is used for optimal performance, unless otherwise specified. The MirrorNovo model was trained on an NVIDIA GeForce RTX 3090 GPU with the following parameters: optimizer = Adam, loss function = Focal, maximum epochs = 5, batch size = 20, learning rate = 1 × 10⁻³, decay factor = 0.5. For the mirror-spectra recognition experiment, we first retained the 200 most abundant peaks and then selected 7 peaks per 100 Da. For the sequencing experiment, we only retained spectra with a precursor mass in the range of [300, 3500] Da. When comparing these two algorithms (as discussed in “Results”), we set the number of processes of pNovoM2 as 15, and the batch size of MirrorNovo as 350. Additionally, MirrorNovo was run on an NVIDIA GeForce RTX 4090 GPU, while pNovoM2 was executed on an Intel(R) Xeon(R) Silver 4410Y CPU.

• pNovo3: pNovo3¹² uses the pDag⁴¹ algorithm to find the maximum scoring path within the spectrum graph, after which it rescores candidate peptides using an SVM model. The peptide with the highest score is then reported as the output. For our analysis in “Results”, we used the latest version of pNovo3, pNovo v3.1.5, which was released on December 6, 2023.

• PointNovo*: PointNovo¹³ is a neural network model that employs a T-Net structure, specifically designed for order-invariant data. This design allows it to be applicable to mass spectra of varying resolutions without increasing computational complexity. Since the authors of PointNovo did not provide a trained model file, and the original 9-species training dataset used in¹³ included the yeast species, we excluded the yeast species and trained a PointNovo model from scratch on the remaining 8-species dataset for our analysis.

• GraphNovo: We acquired the commercial software Peaks-Online 12 and used its built-in de novo peptide sequencing algorithm to analyze E. coli and yeast datasets in “Results”. Peaks-Online 12 utilizes an integrated deep learning-based de novo peptide sequencing algorithm (GraphNovo¹⁷), which addresses the missing fragments problem in de novo sequencing by using graph deep learning to accurately reconstruct full peptide sequences from incomplete mass spectrometry data. However, the specifics of the training set used to train the model remain undisclosed to us.

• Casanovo* and CasanovoV2: Casanovo^14,15 is a deep learning algorithm that employs a transformer neural network architecture to translate the sequence of peaks in a tandem mass spectrum into the corresponding amino acid sequence of the generating peptide. Notably, Casanovo does not constrain the peptides it generates by precursor mass, so the precursor masses of its generated peptides may not be equal to those provided by the spectra. In contrast, all the other algorithms mentioned above use precursor mass as a constraint for de novo peptides. Therefore, it would be unfair to Casanovo to directly use the highest-scoring results without considering the precursor mass error. However, if we filter out peptides generated by Casanovo with precursor mass errors exceeding 20 ppm, some low-quality spectra will not yield any sequencing results. As a result, these spectra will be discarded in the results of Casanovo, and the TD mapping method will be applied only to the remaining spectra with higher quality. This could unfairly underestimate the FDR of Casanovo compared to other software tools. For a fairer comparison, we adopted a compromise strategy: we filtered out peptides generated by Casanovo with precursor mass errors exceeding 20 ppm and applied the same peptide mass thresholds to the Casanovo results as we did to DiNovo. For our analysis in “Results”, we used two models of Casanovo 4.2.0, which was released on May 14, 2024. The first model, referred to as Casanovo*, was trained by us from scratch on the 8-species dataset, similar to the approach used with PointNovo. The second model, referred to as CasanovoV2, is the latest version provided by its authors, trained on a large-scale dataset derived from the MassIVE Knowledge Base (MassIVE-KB), which consists of 30 million PSMs. It is important to note that CasanovoV2 is a trained model that does not allow for the specification of modifications. For a fair comparison, we selected the highest-ranking peptide without unexpected modifications as the result for each spectrum.

• pFind: In our analysis, we used pFind v3.2.0^27,28,29 for database search. The E. coli database includes 4789 protein entries along with contaminant protein sequences from UniProt. The yeast database contains 7021 protein entries from UniProt, also with contaminant protein sequences. For the closed search parameters, the maximum number of modification sites per peptide sequence was limited to 3. The length of identified peptides was set to 6 to 100. The digestion mode was set as full specificity with a maximum of 3 missed cleavage sites allowed. The search results were filtered to achieve a 1% FDR at both the PSM and protein levels.

• MSFragger: In our analysis, we used FragPipe 23.0, in which the version of MSFragger³⁰ is v4.2. The searched E. coli and yeast databases were the same as those used in pFind, and the basic search parameters were consistent with those of pFind.

• MS-GF+ and MS-GF + *: In our analysis, we used two models of MS-GF+ v20240326³¹. The first model, referred to as MS-GF + , was the original model provided by its authors. The second model, referred to as MS-GF + *, was retrained on peptides digested by four different proteases. Specifically, we used the program provided in the original text to train scoring parameters for each protease. The searched E. coli and yeast databases were the same as those used in pFind, and the basic search parameters were consistent with those of pFind.

Design of DiNovo software

Previously, the only software available for mirror-protease de novo sequencing was pNovoM, which required users to manually carry out step-by-step operations for pre-sequencing, mirror-spectra recognition, and de novo sequencing (note that the programs for the first two steps were not publicly released). This makes its use of it cumbersome and not user-friendly. As a result, there remains a need for software that can recognize and sequence mirror spectra with a simple, one-click operation.

To meet this need, we integrate the various algorithms developed in this paper into a complete software solution named DiNovo, making it easily accessible to participants in the field. The DiNovo software includes the following modules: configuration settings, MS data I/O, spectra preprocessing, mirror-spectra recognition, de novo sequencing, and real-time log output during the process. As illustrated in Supplementary Fig. 14, parallelization is fully supported in each analytical component of DiNovo, significantly improving computational efficiency and reducing overall processing time.

In DiNovo, error-tolerant precursor mass indexes are built for all spectra to enable high-speed mirror-spectra recognition. As a result, the time complexity of DiNovo for selecting candidate spectral pairs involving \(n\) trypsin-digested spectra and \(m\) LysargiNase-digested spectra is \(O(m+n)\), in contrast to \(O({mn})\) by brute force. For example, DiNovo takes approximately 20 s to select candidate spectral pairs from a pool of 10 million spectral pairs, 1000 times faster than brute force. On the 16 million spectral pairs of E. coli dataset, DiNovo performed mirror-spectra recognition 2.1 and 46.5 times faster than pMerge (the mirror-spectra recognition module of pNovoM, unreleased) with 1 and 30 processes on a workstation, respectively. In combination of mirror-spectra recognition and sequencing, DiNovo completed the whole task four to five times faster than pMerge and pNovoM.

Reporting summary

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