Foundation model for mass spectrometry proteomics

TL;DR

This work introduces Casanovo Foundation, a foundation model for tandem mass spectrometry proteomics that pre-trains a spectrum encoder on de novo sequencing to learn generalizable spectrum representations. By applying lightweight task heads to frozen embeddings, the model improves performance on spectrum quality, chimericity, phosphorylation, and glycosylation status prediction, with multi-task fine-tuning yielding further gains. The results show strong benefits when labeled data are limited (notably in phosphorylation with small datasets) and demonstrate that a shared spectrum encoder can serve as a versatile starting point for diverse proteomics tasks. The approach has practical implications for data quality assessment and acquisition planning in proteomics experiments, while highlighting the need for substantial compute resources and potential future directions toward unsupervised pre-training.

Abstract

Mass spectrometry is the dominant technology in the field of proteomics, enabling high-throughput analysis of the protein content of complex biological samples. Due to the complexity of the instrumentation and resulting data, sophisticated computational methods are required for the processing and interpretation of acquired mass spectra. Machine learning has shown great promise to improve the analysis of mass spectrometry data, with numerous purpose-built methods for improving specific steps in the data acquisition and analysis pipeline reaching widespread adoption. Here, we propose unifying various spectrum prediction tasks under a single foundation model for mass spectra. To this end, we pre-train a spectrum encoder using de novo sequencing as a pre-training task. We then show that using these pre-trained spectrum representations improves our performance on the four downstream tasks of spectrum quality prediction, chimericity prediction, phosphorylation prediction, and glycosylation status prediction. Finally, we perform multi-task fine-tuning and find that this approach improves the performance on each task individually. Overall, our work demonstrates that a foundation model for tandem mass spectrometry proteomics trained on de novo sequencing learns generalizable representations of spectra, improves performance on downstream tasks where training data is limited, and can ultimately enhance data acquisition and analysis in proteomics experiments.

Figures (6)

• Figure 1: Benchmarking Casanovo Foundation on downstream tasks. ROC curves and the area under the curve (AUC) reported for: (A) Spectrum quality prediction, (B) Chimericity prediction, (C) phosphorylation detection, and (D) glycosylation status prediction.
• Figure 2: Phosphorylation task experiments. (A) Learning curve showing the performance of Casanovo Foundation and baselines on the phosphorylation prediction task when trained on datasets of varying size. (B) PCA plot visualizing the spectrum embeddings from the pre-trained encoder for spectra in the phosphorylation task test set. (C) Embeddings for the same spectra after multi-task fine-tuning. The percentage of variance explained by each component is indicated in parentheses.
• Figure S1: Glyco precision-recall curve. Precision-recall curve for each model on the glycosylation status prediction task.
• Figure S2: Learned spectrum embeddings. PCA plots visualizing the learned embeddings for test set spectra. The first row contains embeddings from the pre-trained encoder for spectra from (A) the quality prediction task and (B) the chimericity prediction task. The second row (C and D) shows embeddings from the multi-task fine-tuned encoder for the same two datasets. The percentage of variance explained by each component is indicated in parentheses.
• Figure S3: Multi-task loss curves. (A) Training and (B) validation loss curves for each of the four tasks during the multi-task fine-tuning of the spectrum encoder. Train losses curves are smoothed by averaging over the last 10,000 training steps.