MS-DGFormer: Dictionary-Guided Aerosol MS Classifier

• The paper introduces MS-DGFormer, a dual-encoder transformer that leverages class-specific, SVD-denoised spectral dictionaries for single-shot classification of raw aerosol MALDI-MS spectra.
• It employs convolutional patchification and m/z-axis-based positional encoding to capture narrow, overlapping peak structures while minimizing noise from heterogeneous aerosols.
• Empirical results demonstrate significant gains in macro accuracy and inference speed, highlighting its potential for rapid, portable airborne threat detection.

Searching arXiv for the specified paper and closely related mass-spectrometry transformer work. Mass Spectral Dictionary-Guided Transformer (MS-DGFormer) is a dual-encoder transformer architecture for single-shot classification of raw aerosol Matrix Assisted Laser Desorption/Ionization Mass Spectrometry (MALDI-MS) spectra acquired by a portable, field-deployable instrument. It was introduced as the central method in “Unmasking Airborne Threats: Guided-Transformers for Portable Aerosol Mass Spectrometry” (Regan et al., 21 Nov 2025). The model addresses a deployment regime in which each laser shot may hit a different aerosol particle and averaging across shots can blur diagnostic peaks rather than clarify them. In this setting, MS-DGFormer operates on individual raw spectra, treats each spectrum as a 1D sequential signal, and conditions transformer-based sequence modeling on class-structured, Singular Value Decomposition (SVD)-denoised spectral dictionaries. The resulting system is designed for single-shot spectrum classification for airborne biomolecular threat identification, rather than conventional laboratory MALDI-MS workflows that rely on extensive sample preparation and multi-shot averaging (Regan et al., 21 Nov 2025).

1. Problem setting and task definition

MS-DGFormer was proposed for single-shot classification of raw aerosol MALDI-MS spectra in portable environmental monitoring, where autonomous sampling generates noisy spectra for unknown aerosol analytes and where spectra are minimally preprocessed (Regan et al., 21 Nov 2025). The core deployment argument is that conventional MALDI-MS identification typically relies on averaging multiple laser shots from the same analyte to improve signal-to-noise ratio before matching against a database, but that assumption breaks down in aerosol monitoring because each laser shot may strike a different particle and the stream is dominated by heterogeneous background particles such as dust. In that regime, averaging can mix distinct analytes and obscure rather than recover diagnostic structure.

The experimental task is framed operationally as classification of individual spectra into one of five classes: Arizona Road Dust, Bacillus globigii, E. coli, insulin, and ubiquitin (Regan et al., 21 Nov 2025). Although the abstract describes “accurate multi-label classification,” the implemented pipeline is effectively a single-spectrum, single-class prediction problem. The model predicts a patch-level peak pattern and then assigns the spectrum to the class whose class-specific peak template has maximum cosine similarity to the predicted patch vector. This supports the more precise interpretation of MS-DGFormer as a single-shot spectrum classifier.

Each acquired spectrum is represented by an intensity vector and a corresponding mass-to-charge vector: $$\mathbf{s} = [s_1, s_2, \dots, s_l]^T \in \mathbb{R}^l, \qquad \mathbf{m} = [m_1, m_2, \dots, m_l]^T \in \mathbb{R}^l.$$ For a batch of $n$ spectra, the paper stacks them into

$\mathbf{S} \in \mathbb{R}^{n \times l}.$

The manuscript explicitly treats the spectrum as a 1D sequential signal rather than a coarse binned histogram. This suggests a conceptual shift away from conventional spectral vectorization toward tokenized sequence modeling over the raw $m/z$-intensity trajectory.

2. Spectral representation and transformer input construction

The input pipeline adapts Vision Transformer-style patchification to 1D mass spectra through convolutional local embedding (Regan et al., 21 Nov 2025). Rather than linearly slicing fixed windows, MS-DGFormer uses a 1D convolution: $$p_{i,j} = \sum_{k=0}^{\rho-1} w_{j,k}s_{\gamma i+k} + b_j,$$ where $i$ indexes the patch, $j$ the output channel, $\rho$ is the kernel size, and $\gamma$ is the stride. Stacking the outputs yields

$\mathbf{P} \in \mathbb{R}^{N \times h}.$

The paper emphasizes that this is a key adaptation for mass spectra because local convolutional patch extraction captures narrow and overlapping peak structures while suppressing edge artifacts better than a pure linear projection.

The sequence length after patchification is

$n$0

In the reported experiments, $n$1, $n$2, and $n$3, which gives $n$4 tokens (Regan et al., 21 Nov 2025). The paper does not report additional intensity normalization beyond the model-side embedding and label smoothing; if any preprocessing beyond restricting $n$5 Da was used, it is not specified.

Because transformer attention is permutation-invariant, MS-DGFormer injects positional information derived from the physical $n$6 axis. The $n$7 vector is patched in parallel: $n$8 forming $n$9. A learnable linear projection maps each $\mathbf{S} \in \mathbb{R}^{n \times l}.$0 patch into the hidden dimension: $\mathbf{S} \in \mathbb{R}^{n \times l}.$1 with $\mathbf{S} \in \mathbb{R}^{n \times l}.$2 and $\mathbf{S} \in \mathbb{R}^{n \times l}.$3. This positional embedding is added to the spectral embeddings so that the transformer jointly observes local intensity morphology and physically meaningful $\mathbf{S} \in \mathbb{R}^{n \times l}.$4 position. The paper explicitly states that this preserves the structure of the raw $\mathbf{S} \in \mathbb{R}^{n \times l}.$5 axis while lifting it into latent space.

3. Dictionary-guided mechanism and SVD-denoised class priors

The defining feature of MS-DGFormer is its dictionary-guided design (Regan et al., 21 Nov 2025). The paper motivates the method using a sparse-representation perspective in which a spectrum is written as

$\mathbf{S} \in \mathbb{R}^{n \times l}.$6

with $\mathbf{S} \in \mathbb{R}^{n \times l}.$7 and sparse coefficients $\mathbf{S} \in \mathbb{R}^{n \times l}.$8, $\mathbf{S} \in \mathbb{R}^{n \times l}.$9. It references Basis Pursuit,

$m/z$0

but does not solve this optimization inside the network. Instead, the formulation motivates a class-structured spectral dictionary built directly from training spectra.

The dictionary is constructed from $m/z$1 training spectra, evenly distributed across $m/z$2 positive classes: $m/z$3 In the experiments, the dictionary includes only the four positive classes and excludes dust, with $m/z$4 total spectra and $m/z$5 spectra per class (Regan et al., 21 Nov 2025). The dictionary is therefore fixed from training data rather than learned as an adaptive external memory.

Each class sub-dictionary is denoised with SVD and truncated to a low-rank approximation. The paper writes the generic decomposition as

$m/z$6

and the classwise denoising step as

$m/z$7

The denoised global dictionary is then

$m/z$8

Experimentally, the rank is set to $m/z$9 (Regan et al., 21 Nov 2025). The intended idea is that spectra from the same class lie near a low-dimensional subspace, so the leading singular components preserve biologically meaningful peaks while suppressing shot noise. This suggests that the dictionary is not merely side information but a class-structured denoising prior tailored to single-shot deployment conditions.

The denoised dictionary enters the network through a separate embedding pathway. Each denoised spectrum in $$p_{i,j} = \sum_{k=0}^{\rho-1} w_{j,k}s_{\gamma i+k} + b_j,$$0 is patchified with its own convolutional embedding layer, producing

$$p_{i,j} = \sum_{k=0}^{\rho-1} w_{j,k}s_{\gamma i+k} + b_j,$$1

which stack into

$$p_{i,j} = \sum_{k=0}^{\rho-1} w_{j,k}s_{\gamma i+k} + b_j,$$2

The paper states that the input pathway and dictionary pathway use separate learnable weights, allowing one branch to encode noisy raw spectra and the other to encode denoised prior information.

4. Dual-encoder transformer architecture and selection attention

Both branches use transformer encoder blocks based on the Vaswani design (Regan et al., 21 Nov 2025). For the input branch, the patch embeddings after positional addition are projected to queries, keys, and values: $$p_{i,j} = \sum_{k=0}^{\rho-1} w_{j,k}s_{\gamma i+k} + b_j,$$3 with standard scaled dot-product attention: $$p_{i,j} = \sum_{k=0}^{\rho-1} w_{j,k}s_{\gamma i+k} + b_j,$$4 The reported hyperparameters are $$p_{i,j} = \sum_{k=0}^{\rho-1} w_{j,k}s_{\gamma i+k} + b_j,$$5, $$p_{i,j} = \sum_{k=0}^{\rho-1} w_{j,k}s_{\gamma i+k} + b_j,$$6, $$p_{i,j} = \sum_{k=0}^{\rho-1} w_{j,k}s_{\gamma i+k} + b_j,$$7, MLP intermediate dimension $$p_{i,j} = \sum_{k=0}^{\rho-1} w_{j,k}s_{\gamma i+k} + b_j,$$8, and $$p_{i,j} = \sum_{k=0}^{\rho-1} w_{j,k}s_{\gamma i+k} + b_j,$$9 encoder layers in each branch (Regan et al., 21 Nov 2025).

The dictionary encoder is structurally distinct. For each class sub-dictionary

$i$0

a learnable sequence

$i$1

is concatenated to the denoised class spectra. The tensor is then permuted to

$i$2

so that self-attention is performed across dictionary sequences at each patch position rather than along the spectral axis. This mechanism lets the learnable token gather patch-aligned information from all spectra in a class-specific sub-dictionary. After encoding, the model obtains $i$3 aggregated sequence tokens, one per class, each summarizing class-specific denoised structure across all $i$4 patches.

These class-level summaries are consumed by the model’s “selection attention” module, which functions as cross-attention (Regan et al., 21 Nov 2025). The encoded input spectrum acts as query, while the $i$5 aggregated dictionary sequences act as keys and values. The shapes are described as input sequence $i$6 and dictionary summaries $i$7. For each patch location, the model selects which class-specific dictionary features are relevant to the current input spectrum, while a residual connection preserves the original input representation. This is the operational realization of the dictionary-guided idea: the raw noisy spectrum does not map directly to class identity, but instead queries a denoised, class-organized prior memory aligned in patch space.

5. Prediction rule, optimization, and experimental setting

The final prediction is made at the patch level (Regan et al., 21 Nov 2025). For each class $i$8, the target is a binary peak-template vector

$i$9

where $j$0 if class $j$1 has a peak at patch $j$2. Given the fused transformer output $j$3, the classifier predicts

$j$4

with

$j$5

and $j$6. The predicted class is then chosen by cosine similarity to each class peak template: $j$7

Training uses binary cross-entropy between $j$8 and $j$9, together with binary label smoothing that maps positive labels from $\rho$0 to $\rho$1 and negatives from $\rho$2 to $\rho$3 (Regan et al., 21 Nov 2025). The paper reports no auxiliary loss beyond dropout and label smoothing.

The end-to-end pipeline is described explicitly as: start from a raw single-shot spectrum $\rho$4; patchify $\rho$5 by 1D convolution; patchify and project $\rho$6 for positional encoding; encode the input branch; construct classwise denoised dictionary $\rho$7 via SVD; embed the dictionary through a separate path; aggregate each class with a learnable sequence token; use cross-attention from input features to class summaries; predict the patch-level peak vector; optimize BCE; and classify by cosine similarity to class templates (Regan et al., 21 Nov 2025).

The dataset comes from a portable MALDI-Time-of-Flight aerosol mass spectrometer in the digitalMALDI prototype. Spectra are generated by aerosolized particles irradiated by a 349 nm UV laser, and the $\rho$8 range is restricted to $\rho$9 Da (Regan et al., 21 Nov 2025). The class counts are 630 Arizona Road Dust, 1500 B. globigii, 1500 E. coli, 1400 insulin, and 1500 ubiquitin, for a total of 6530 spectra split 80/20 into 5224 training and 1306 test spectra. All models were trained for 300 epochs with batch size 8, learning rate $\gamma$0, 10% warm-up, cosine annealing decay, and dropout probability 0.1. The optimizer is not specified.

6. Empirical performance, efficiency, and limitations

The reported baselines are RNN-6, LSTM-4, BiLSTM-6, MS-Former-3, and MS-Former-7 (Regan et al., 21 Nov 2025). The MS-Former variants remove the dictionary embedding, dictionary encoder, and selection attention, leaving a standard transformer backbone. This makes the strongest ablation evidence architectural rather than factorial: the benefit of the full dictionary-guided transformer is established, but SVD denoising is not isolated from the broader class-conditioned cross-attention mechanism.

The macro results are as follows:

Model	Macro Accuracy	Macro F1
RNN-6	0.560	0.491
LSTM-4	0.679	0.641
BiLSTM-6	0.939	0.915
MS-Former-3	0.709	0.664
MS-Former-7	0.862	0.824
MS-DGFormer	0.983	0.982

MS-DGFormer therefore improves macro F1 by 0.067 over the strongest baseline, BiLSTM-6, and by 0.158 over the parameter-matched transformer baseline MS-Former-7 (Regan et al., 21 Nov 2025). The most practically salient per-class gain is on Arizona Road Dust: compared with MS-Former-7, dust F1 rises from 0.553 to 0.949, and compared with BiLSTM-6 from 0.736 to 0.949. The paper interprets this as evidence that the denoised positive-class dictionary helps the model avoid overreacting to noisy background spectra.

The paper also introduces an inference-optimized form, MS-DGFormer-E, based on the observation that the dictionary is constant and class-aggregated summary sequences can be precomputed (Regan et al., 21 Nov 2025). This removes the full dictionary pathway at inference while preserving classification performance. Parameters fall from 8.36M to 4.39M. At batch size 1, inference time drops from 72.27 ms to 12.31 ms and throughput rises from 13.84 to 81.23 spectra/s. At batch size 8, inference time changes from 127.17 ms to 66.33 ms and throughput rises from 62.91 to 120.60 spectra/s. The reported hardware-specific conclusion is “nearly a $\gamma$1” increase in mean inference speed and more than $\gamma$2 throughput, which strengthens the deployment claim for portable autonomous monitoring.

Several limitations are explicit (Regan et al., 21 Nov 2025). The dataset is limited to five classes, with only four positive classes included in the dictionary. Data are collected class-specifically in a laboratory setting to mimic deployment, rather than from naturally mixed open-world field samples with many unknown analytes. The paper does not report AUROC, calibration, explicit synthetic-noise robustness, limited-data scaling, or open-set detection for truly unknown aerosol analytes. It also does not separately quantify the contributions of SVD denoising and class-conditioned cross-attention. A plausible implication is that MS-DGFormer should be understood as a strong proof of concept for raw single-shot aerosol MALDI-MS classification rather than a complete solution to open-world airborne threat identification.

7. Position within transformer-based mass spectrometry research

MS-DGFormer occupies a distinct position within transformer-based mass spectrometry research because it is a raw-spectrum, dictionary-guided classifier for portable aerosol MALDI-MS rather than a molecular structure elucidation model (Regan et al., 21 Nov 2025). This differentiates it from structure-generation systems such as MS-BART, which maps predicted fingerprint tokens to molecular SELFIES through a BART-style encoder-decoder (Han et al., 23 Oct 2025), and from DiffMS, which combines a transformer spectrum encoder with a formula-restricted graph diffusion decoder for de novo small-molecule generation (Bohde et al., 13 Feb 2025). It also differs from forward models such as MassFormer, which predicts LC-ESI-MS/MS spectra from molecular graphs using a Graphormer backbone (Young et al., 2021).

The closest conceptual relation lies in the use of structured priors. In MS-DGFormer, the prior is a class-specific, low-rank spectral dictionary aligned in patch space. In MS-BART, the analogous prior is a fixed token vocabulary spanning fingerprint bits and molecular SELFIES (Han et al., 23 Oct 2025). In DiffMS, peak formula annotations and neutral-loss structure act as chemically grounded conditioning for graph generation (Bohde et al., 13 Feb 2025). MS-DGFormer is unusual in that it injects class-specific denoised exemplars directly into the attention pathway of a transformer over raw spectra, without converting the problem into molecular generation or symbolic spectrum-to-structure translation.

This suggests a broader methodological interpretation. MS-DGFormer shows that “dictionary guidance” in mass spectrometry need not mean retrieval from an external library at inference time. It can also mean conditioning a transformer on denoised, class-structured prior exemplars that preserve patch alignment and physically meaningful $\gamma$3 position (Regan et al., 21 Nov 2025). In that sense, the method stands as a specialized instance of dictionary-guided transformer design for high-noise, low-preprocessing, field-deployable mass spectrometry, with its primary contribution lying in how denoised class subspaces are fused with raw single-shot measurements through cross-attention.