Latent Imputation before Prediction: A New Computational Paradigm for De Novo Peptide Sequencing

We acknowledge with gratitude for the reviewer's appreciation that our work's motivation is quite clear and supported by the experimental results, as well as the recognition that the design of proposed algorithm is generally reasonable. We address the reviewer's concerns in detail below.

Q1: The weakness of this work may can be the overall algorithm design is not that impressive and cannot be adapted to other fields. However, personally I like such application paper with simple yet effective approach.

Thanks for the constructive feedback and appreciation of our simple yet effective approach. While our algorithm is specifically designed for peptide sequencing, it may offer insights into other computational biology areas facing similar challenges. For instance, in single-cell transcriptomics [3], expression values appear undetected due to technical limitations rather than true biological absence, and the number of missing expression values varies across cells. Our method could serve as inspiration for developing analogous solutions in such contexts.

Q2: The imputation may also introduce additional time consumption. Could the authors also provide the time portion of imputation in the entire pipeline.

Thanks for the valuable suggestion. We evaluate the computational overhead of the imputation module in Table 4, which reports the inference time and proportion (averaged over the Nine-Species test set) of each module in LIPNovo. The results show that the imputation module accounts for only 0.46% of the total pipeline runtime.

In Table 5, we also provide the training time comparison, which shows that our LIPNovo adds 8 minutes per training epoch compared to the baseline model. All experiments were conducted on a server equipped with:

• CPU: Intel(R)Xeon(R)Gold 5418Y @ 2.00GHz (96 cores,AVX2)

• GPU: NVIDIA GeForce RTX 4090 (24GB)

• OS: Ubuntu 20.04.6 LTS

Table 4: Inference time (ms) and proportion of each module in LIPNovo.

Table 5: Training time (minutes) for an epoch ($499,402$ samples with batch size=$32$) compared to the baseline model.

[3]. Deng et al. Scalable analysis of cell-type composition from single-cell transcriptomics using deep recurrent learning, Nature Methods, 2019.

We sincerely appreciate the reviewer for the recognition that the proposed method is novel, and the evaluation is fair, which addresses the key issue of missing fragments in de novo peptide sequencing and can advance progress in this field. We address the reviewer's question as follows.

Q: It is better to demonstrate the experimental results to see whether this method is orthogonal to HelixNovo. which is also a method for solving the missing fragment problem. Using the complementary spectra as the LIPNovo's input.

Thanks for the thoughtful suggestion. $\pi$-HelixNovo and LIPNovo actually address different aspects of the missing fragmentation issue.

$\pi$-HelixNovo handles the case where at least one ion ($b$ or $y$) in a pair is present, using the complementary spectrum. Specifically, if the $b$ ion is observed but the paired $y$ ion is missing, $\pi$-HelixNovo can recover the missing $y$ ion by subtracting the $b$ ion's mass from the precursor mass, which generates a complementary spectrum. However, $\pi$-HelixNovo cannot handle cases where a pair of $b$ and $y$ ions are all missing, which is how the missing ratio is calculated according to [2].

In contrast, LIPNovo is designed to impute all types of missing scenarios. This is achieved by leveraging an imputation module that predicts the theoretical spectrum in a latent space, thus it effectively addresses problems that $\pi$-HelixNovo cannot resolve. This has been further evaluated through ablation experiments (as shown in Table 3). Specifically, "Baseline+ Complementary Spectrum" improves peptide precision by 0.8%. In contrast, "Baseline + our Imputation mechanism" achieves a +4.0% increase. Moreover, "Baseline + Imputation + Complementary Spectrum" further leads to 1.3% increase, because the complementary spectrum can provide additional information for the imputation process.

Table 3: Experiments regarding the complementary spectrum used in $\pi$-HelixNovo.

[2]. Zhou et al. NovoBench: Benchmarking Deep Learning-based De Novo Peptide Sequencing Methods in Proteomics, NeurIPS 2024 D&B track.

Thanks for the positive feedback on our work's novelty and the recognition that our method significantly improves the robustness of the model, achieves SOTA performance, and demonstrates generalization and efficiency. Our responses to the concerns are as follows.

Q1: Limitations of theoretical spectrum generation.

We make some clarifications about the theoretical spectrum generation.

In our work, the theoretical spectra are computed from ground truth peptide sequences in the training set, where PTMs are explicitly considered. This ensures that peptides with different PTMs yield distinct theoretical spectra, and the mass deviations introduced by modifications are properly accounted for during training.

Importantly, theoretical spectra are only used during training. Our method does not require theoretical spectrum computation during inference.

LIPNovo is trained to map incomplete observed spectra to their corresponding theoretical spectrum representations. This acts as a signal enhancement mechanism, helping to mitigate deviations between observed and theoretical spectra and clarify spectral patterns, ultimately improving peptide sequence prediction.

In more complex fragmentation scenarios (e.g. Electron Transfer Dissociation producing $c/z$ ions), the theoretical spectra can be extended to include these additional ion types. LIPNovo remains compatible with such scenarios without requiring changes to its computational paradigm, further demonstrating its flexibility.

Q2: Computational efficiency analysis.

Thanks for the valuable suggestion. We evaluate the computational overhead compared to the baseline in Table 1, which reports the inference time and GPU memory usage. The values are the mean across examples in the Nine-Species test set. The results show that LIPNovo increases the inference time by 14.82 ms (+1.94%) per example, and consumes an additional 98MB of GPU memory. All experiments were conducted on a server equipped with:

• CPU: Intel(R)Xeon(R)Gold 5418Y @ 2.00GHz (96 cores,AVX2)

• GPU: NVIDIA GeForce RTX 4090 (24GB)

• OS: Ubuntu 20.04.6 LTS

Table 1: Comparison of computational overhead compared to baseline.

Q3: Improvement direction for extreme missing scenarios.

Thanks for the thoughtful suggestion. In Lines 425-436, we have discussed such limitations and introduced possible improvement directions. As indicated, one potential way is utilizing cross-sample information for missing peak imputation. To investigate this, we conduct a preliminary exploration to incorporate cross-sample learning into LIPNovo. Specifically, for each spectrum we search for the top one most similar spectrum (i.e., reference spectrum) from the training set with mass constraints, following [1]. Spectral similarity implies possible peptide similarity, where shared fragments can offer additional information for imputation. The reference spectrum is encoded by the spectrum encoder, and the resulting representation is fed into the imputation module to assist imputation.

The results are presented in Table 2. As shown, under the high missing ratio range [0.7, 0.8), while the LIPNovo improves the baseline by only 0.86%, the simple extension LIPNovo+Cross-Sample achieves a 2.22% improvement. On average, this extension enhances baseline performance by 8.38% (v.s. 6.58% by LIPNovo).

Table 2: Experiments on LIPNovo's improvement direction for extreme missing scenarios. Amnio-acid precision (%) is reported.

[1]. Xia et al. SearchNovo, ICLR 2025.