Replacing a single fingerprint with a band of spectrum-informed queries set new state-of-the-art results on two mass spectrometry structure elucidation benchmarks
AspireAI community · arXiv
Molecular structure elucidation from tandem mass spectra is a central inverse problem, and most approaches remain tied to reference libraries or predefined candidate sets, while de novo methods generate structures directly from spectra. The common de novo route predicts a molecular fingerprint from the spectrum and decodes structures from it, which creates a training-inference mismatch: the decoder is trained on oracle fingerprints computed from molecules but at inference is queried with a noisy spectrum-induced posterior collapsed to a single thresholded fingerprint. MS-GPT recasts fingerprint-mediated de novo elucidation as spectrum-induced posterior querying of a conditional molecule-language model, conditioning on fingerprints and formulas and converting the spectrum-induced posterior into a band of fingerprint queries near the oracle-fingerprint manifold through active-bit density calibration, then pooling candidates across the band and ranking by generation frequency and posterior similarity. A lightweight LoRA adapter mitigates domain-specific posterior bias while preserving the pretrained molecular prior. On NPLIB1 and MassSpecGym, MS-GPT sets a new state of the art with Top-1 and Top-10 exact-match accuracy of 17.1% and 27.6% and 10.2% and 13.2% respectively, and candidate-pool scaling improves recovery with a near-linear increase in inference cost. The limitation is that even the state-of-the-art exact-match rates are low, so the method is a meaningful advance on a problem that remains largely unsolved.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.