Transferable Mass Spectrum Prediction via Reference-Guided Test-time Specialization

Yunhua Zhong; Runting Li; Yifan Li; Pan Liu; Zhiwen Yang; Zikun Wang; Yixuan Tang; Jun Xia

cs.LGSep 28, 2026

Transferable Mass Spectrum Prediction via Reference-Guided Test-time Specialization

Authors: Yunhua Zhong, Runting Li, Yifan Li, Pan Liu, Zhiwen Yang, Zikun Wang, Yixuan Tang, Jun Xia

Organizations: The Hong Kong University of Science and Technology (Guangzhou) · The University of Hong Kong · South China University of Technology · Hong Kong Polytechnic University · Jinan University · The Hong Kong University of Science and Technology

Abstract

Tandem mass spectrum prediction supports compound identification across metabolomics, natural-product discovery, and environmental analysis. However, pretrained predictors often degrade under shifts in chemical space and acquisition conditions, while retraining domain-specific models from scratch is costly. We introduce SPARC, a retrieval-guided test-time specialization framework that adapts a pretrained predictor using a spectral reference library without accessing test-query spectra. For each target query, SPARC retrieves chemically related reference spectra to recalibrate fragment intensities within the learned fragmentation space. During Transfer, SPARC combines reference-guided spectral adaptation with reliability-aware consistency, using reconstruction behavior on retrieved spectra to selectively preserve trustworthy predictions during continual specialization. Across MassSpecGym, NPLIB1 and application-specific GNPS libraries, SPARC improves spectral prediction under multiple transfer settings. These results establish retrieval-guided test-time specialization as a practical strategy for extending pretrained MS/MS predictors to specific chemical and acquisition domains, with continual test-time training providing further refinement during deployment.

Figures & tables

Figure 1: Overview of SPARC . Top: A predictor pretrained on a broad source collection is specialized to a target research domain. Bottom: Target molecules retrieve acquisition-compatible and chemically similar support spectra. The fragment generator remains frozen, while support reconstruction errors calibrate bin-wise reliability of an exponential moving average teacher. The student intensity model optimizes similarity-weighted support reconstruction and reliability-weighted teacher–student consistency, with stochastic restoration and rollback constraining adaptation drift.

Figure 2: Prediction performance is locally structured in spectral and molecule spaces.

Method Entropy MSE EntSim CosSim MSG

\rightarrow

MSG-MNa (MSG initialized) NEIMS 4.150 14.461 0.072 0.072 GrAFF-MS 4.134 15.822 0.075 0.071 FIORA 3.015 1.020 0.063 0.071 Iceberg–MSG 3.461 1.242 0.180 0.280 Iceberg–MNa 2.485 1.006 0.262 0.348 SPARC (val) 2.715 0.895 0.274 0.370 SPARC -TTT 2.923 0.882 0.284 0.376 Iceberg(tent) 2.077 1.183 0.229 0.290 Iceberg(cotta) 3.419 1.189 0.191 0.288

Method Entropy MSE EntSim CosSim MSG

\rightarrow

NPLIB1 (MSG initialized) NEIMS 3.568 20.246 0.392 0.363 GrAFF-MS 3.396 21.170 0.330 0.280 FIORA 1.900 1.317 0.369 0.412 Iceberg-MSG 3.586 1.269 0.470 0.517 Iceberg-NPLIB1 3.516 1.217 0.521 0.586 SPARC (val) 3.112 1.063 0.559 0.630 SPARC –TTT 3.128 1.041 0.575 0.650 Iceberg(tent) 3.052 1.252 0.490 0.517 Iceberg(cotta) 3.560 1.245 0.472 0.518

Table 1: Test performance for MSG

\rightarrow

MSG-MNa and MSG

\rightarrow

NPLIB1 transfer.

Method Entropy MSE EntSim CosSim MSG M+H (MSG initialized) NEIMS 3.982 16.180 0.184 0.162 GrAFF-MS 3.756 16.713 0.169 0.137 FIORA 1.742 0.917 0.434 0.460 Iceberg–MSG 3.477 0.853 0.462 0.529 SPARC (val) 2.866 0.907 0.494 0.532 SPARC –TTT 2.794 0.880 0.505 0.547

Method Entropy MSE EntSim CosSim NPLIB1 (NPLIB1 initialized) NEIMS 3.349 20.652 0.324 0.280 GrAFF-MS 3.521 19.689 0.384 0.362 FIORA 1.596 1.237 0.419 0.484 Iceberg-NPLIB1 3.516 1.217 0.521 0.586 SPARC (val) 2.925 1.107 0.559 0.616 SPARC –TTT 2.944 1.092 0.565 0.628

Table 2: Test performance for MSG M+H and NPLIB1.

Figure 3: Representative example of prediction from SPARC . Each displayed spectrum is normalized, and peaks with intensities below 1% are removed.

	DoA			3HAA			ECG			GNPS-A/B			SC-FDA
Method	MSE	EntSim	Cos	MSE	EntSim	Cos	MSE	EntSim	Cos	MSE	EntSim	Cos	MSE	EntSim	Cos
NEIMS	19.90	0.198	0.127	12.84	0.386	0.382	34.33	0.184	0.145	18.05	0.409	0.242	14.75	0.148	0.110
GrAFF-MS	19.98	0.149	0.084	15.23	0.238	0.173	34.49	0.145	0.117	19.84	0.324	0.179	15.57	0.157	0.107
FIORA	1.078	0.427	0.458	0.935	0.395	0.397	2.836	0.140	0.117	1.009	0.393	0.488	0.893	0.362	0.405
Iceberg	1.057	0.415	0.483	0.904	0.490	0.509	2.621	0.178	0.223	0.960	0.504	0.511	0.925	0.322	0.379
SPARC	1.058	0.482	0.509	0.901	0.508	0.514	2.654	0.205	0.246	0.919	0.534	0.569	0.839	0.389	0.435

Table 4: Performance on Drugs of Abuse, 3HAA, ECG, GNPS-A/B, and SC-FDA. All models are initially trained on the MSG dataset.

Method	MSE	EntSim	CosSim
-similarity-based support retrieval	0.9859	0.2595	0.3441
-support-calibrated bin confidence	0.8889	0.2637	0.3514
-stochastic restoration	0.8840	0.2720	0.3692
-per-query rollback	0.8977	0.2656	0.3631
SPARC	0.8804	0.2742	0.3755

Table 5: Ablation Study of SPARC Components.

Explore similar work

FlexMS: A Unified Public Benchmark for Molecule Tandem Mass Spectrum Prediction
Feb 26, 2026Yunhua Zhong, Yixuan Tang, Yifan Li +5Tandem Mass Spectra Metabolic
The Hong Kong University of Science and Technology (Guangzhou). · The University of Hong Kong. · Yangzhou University. +3
GLACIER: Rethinking Mass Spectrum Prediction as an Object Detection Problem
Jun 28, 2026Rui-Xi Wang, Runzhong Wang, Connor W. ColeyTandem Mass Spectra Detection Transformer
Massachusetts Institute of Technology Cambridge, MA 02139
PepSpecBench: A Unified Evaluation Benchmark for Peptide Tandem Mass Spectrometry Prediction
May 3, 2026Zhiwen Yang, Pan Liu, Yifan Li +2Tandem Mass Spectra Multiple Sclerosis
The Hong Kong University of Science and Technology (Guangzhou). · Yangzhou University. · The Hong Kong University of Science and Technology.

Feb 26, 2026, cs.AIJ/K move · Enter open · S save

FlexMS: A Unified Public Benchmark for Molecule Tandem Mass Spectrum Prediction

Yunhua Zhong, Yixuan Tang, Yifan Li, Pan Liu+4

The Hong Kong University of Science and Technology (Guangzhou). · The University of Hong Kong. · Yangzhou University.+3

Tandem Mass Spectra Metabolic Multiple Sclerosis

Open paper AskPDF arXiv

Tandem mass spectrometry (MS/MS) is central to small molecule identification, but current deep learning systems for spectrum prediction still remain difficult to evaluate and deploy in practice. While novel architectures constantly claim state-of-the-art performance, inconsistent metadata conditioning and entangled preprocessing pipelines hinder fair architectural comparisons. Besides, existing evaluations are often restricted to curated datasets, failing to capture the heterogeneity and cross-domain shifts of real-world metabolomics. Furthermore, current benchmarks lack difficulty-aware diagnostics and leave blind to how models behave under specific compute or data constraints. To address this, we present FlexMS, a modular public-data benchmark framework that standardizes MS/MS prediction across public resources while keeping molecular encoders, metadata conditioning, predictor heads, and downstream retrieval under one protocol. FlexMS establishes a fair evaluation playground which significantly lowers the barrier for integrating new predictive tools. Rather than solely optimizing for average scores, FlexMS augments aggregate accuracy with difficulty-aware diagnostics, providing actionable guidance on model selection across different compute constraints, data scales, and downstream retrieval objectives. Ultimately, FlexMS provides the community with a reproducible standard to identify which algorithmic conclusions are stable and which operating points are most viable in practice.

Abstract

Explore similar work

FlexMS: A Unified Public Benchmark for Molecule Tandem Mass Spectrum Prediction

GLACIER: Rethinking Mass Spectrum Prediction as an Object Detection Problem

PepSpecBench: A Unified Evaluation Benchmark for Peptide Tandem Mass Spectrometry Prediction