Sparse Autoencoders for Interpretability of InstaNovo
thesis · 2026
| Date | 2026-06-25 |
| Type | thesis |
| Publisher | MSc thesis |
| Contribution | adjacent |
| Supervisor | Konstantinos Kalogeropoulos, Jeroen Van Goey, Jemma Daniel |
| Link | https://jeroen.vangoey.be/files/theses/interpretability-sparse-autoencoders-instanovo.pdf |
Abstract
Transformer models for de novo peptide sequencing reconstruct peptide sequences from mass spectra with high accuracy but offer no insight into which spectral features drive their predictions. This work investigates whether sparse autoencoders can uncover chemically meaningful representations inside InstaNovo, a de novo peptide sequencer. We train sparse autoencoders at four depths of the model’s encoder and pair them with a fragment-ion annotation pipeline that labels each spectral peak with chemical concepts derived from first principles, allowing us to ask which learned features correspond to which chemical events. The trained autoencoders reconstruct the encoder’s internal representations with greater than 92% fidelity at every layer while preserving more than 99.5% of the model’s sequencing loss. A substantial fraction of features associate significantly with specific chemical concepts, and these associations are organised by depth: physical spectral properties are represented most sharply in early layers, while chemically specific properties such as cleavage specificity and modification localisation emerge in the deeper layers. Causal ablation experiments establish the degree to which these associations reflect genuine use by the model rather than coincidental correlation. The findings show that sparse autoencoders are a viable tool for mechanistic interpretability of transformer models used in proteomics, while exposing a structural limitation in how correlation-based feature ranking interacts with the chemical concepts that co-occur throughout mass spectrometry data.
Methods and tools
- InstaNovo: Knapsack beam search
- Sparse autoencoder interpretability of InstaNovo: Mechanistic-interpretability probe of a de novo sequencer: sparse autoencoders trained on the internal activations at four depths of InstaNovo’s spectrum encoder, paired with a fragment-ion annotation pipeline that labels every peak with chemical concepts derived from first principles, so learned features can be matched to chemical events. Reconstructs the encoder’s representations at over 92% fidelity per layer while preserving more than 99.5% of sequencing loss, and finds the feature-to-concept associations are organised by depth: physical spectral properties sharpest in early layers, cleavage specificity and modification localisation emerging deeper. Causal ablation tests how far those associations reflect genuine use by the model, and exposes a structural limit in correlation-based feature ranking when chemical concepts co-occur.