[Paper Review] ContraNovo: A Contrastive Learning Approach to Enhance De Novo Peptide Sequencing

Zhi Jin, Sheng Xu+7 more|arXiv (Cornell University)|Dec 18, 2023

Advanced Proteomics Techniques and Applications7 citations

TL;DR

ContraNovo uses contrastive learning to bridge spectra and peptides and incorporates prefix/suffix mass information into decoding, achieving state-of-the-art de novo peptide sequencing on benchmark datasets.

ABSTRACT

De novo peptide sequencing from mass spectrometry (MS) data is a critical task in proteomics research. Traditional de novo algorithms have encountered a bottleneck in accuracy due to the inherent complexity of proteomics data. While deep learning-based methods have shown progress, they reduce the problem to a translation task, potentially overlooking critical nuances between spectra and peptides. In our research, we present ContraNovo, a pioneering algorithm that leverages contrastive learning to extract the relationship between spectra and peptides and incorporates the mass information into peptide decoding, aiming to address these intricacies more efficiently. Through rigorous evaluations on two benchmark datasets, ContraNovo consistently outshines contemporary state-of-the-art solutions, underscoring its promising potential in enhancing de novo peptide sequencing. The source code is available at https://github.com/BEAM-Labs/ContraNovo.

Motivation & Objective

• Motivate improving de novo peptide sequencing accuracy beyond sequence-to-sequence framing.
• Leverage contrastive learning to align spectrum and peptide representations.
• Incorporate amino acid masses via prefix/suffix information to guide decoding.
• Develop a decoder that utilizes mass-aware embeddings and an integrated loss for joint training.

Proposed method

• Encode spectra and peptides with self-attention-based encoders to obtain global spectrum/peptide features.
• Train using a contrastive loss to align spectrum and peptide representations in a shared space.
• Decode peptides with a transformer-like decoder that uses prefix-suffix mass information and a mass-aware amino acid embedding lookup.
• Combine cross-entropy decoder loss with contrastive loss in a joint training objective, with a dynamic weighting parameter.
• Incorporate peak, precursor, and amino acid mass information into the embedding and decoding process.

Experimental results

Research questions

• RQ1 Can contrastive learning between spectra and peptides improve de novo sequencing accuracy beyond traditional seq2seq approaches?
• RQ2 Does incorporating prefix and suffix mass information into decoding improve amino acid and peptide-level predictions?
• RQ3 How does mass-aware amino acid embedding impact decoding decisions for close-mass amino acids?
• RQ4 What is the impact of combining contrastive loss with standard decoding loss on overall performance?

Key findings

• ContraNovo achieves higher amino acid and peptide precision than Peaks, DeepNovo, PointNovo, CasaNovo, and CasaNovoV2 on the 9-species-V1 benchmark.
• On average across nine species, ContraNovo improves amino acid precision by 4.6% over CasaNovoV2 and peptide precision by 8.2% over CasaNovoV2.
• On 9-species-V2 with richer modifications, ContraNovo exceeds CasaNovoV2 by 3.3% in amino acid precision and 3.7% in peptide precision.
• ContraNovo shows strong performance on human data, surpassing CasaNovoV2 by 17.6% at the amino acid level.
• Ablation studies show beam search, contrastive learning, amino acid embedding, and prefix/suffix mass information each contribute to performance gains, with notable drops when omitted.
• ContraNovo better handles amino acids with similar masses (e.g., M(Oxidation) vs. F) than CasaNovoV2.