Marker Ion Ranking And Filtering

marker-ion-ranking-and-filtering

Summary

Iteratively rank and select marker ions from imaging mass spectrometry data using latent low-dimensional peak features learned by a graph-attention autoencoder, stopping when a target number of markers is reached or diminishing returns are observed. This skill transforms learned feature representations into an ordered list of biologically or chemically significant m/z peaks for downstream analysis.

When to use

You have extracted latent low-dimensional peak features from imaging mass spectrometry (IMS) data using a graph-attention autoencoder and need to identify a ranked subset of marker ions that best represent spatial metabolomic patterns. Use this skill when your goal is to reduce dimensionality from hundreds or thousands of m/z peaks to a manageable, interpretable set of biomarker candidates ranked by importance or variance in latent space.

When NOT to use

• Latent features have not yet been extracted by a graph-attention autoencoder or equivalent dimensionality-reduction model; use the feature-extraction step first.
• Raw imaging mass spectrometry peak data is provided without prior feature learning; this skill operates on learned representations, not raw m/z intensities.
• Input is already a pre-curated, manually validated set of marker ions; this skill is for automated discovery and ranking, not validation of known markers.

Inputs

• latent low-dimensional peak features (HDF5 .h5 file or NumPy array from graph-attention autoencoder)
• m/z identifiers or peak indices corresponding to the latent features
• target number of marker ions (integer parameter)
• ranking criterion (e.g., 'variance', 'importance_score', or custom scoring function)

Outputs

• ranked list of marker-ion m/z identifiers (CSV or tabular format)
• ranked scores or importance weights for each selected marker ion
• iteration history / convergence log (optional)

How to apply

Load the latent low-dimensional peak features (typically stored in HDF5 or NumPy format) output by the graph-attention autoencoder. Initialize an empty marker-ion list and iteration counter. In each iteration, apply a ranking criterion—such as feature importance scores or variance across latent dimensions—to identify the top m/z peaks not yet selected. Add newly selected peaks to the marker list and increment the iteration counter. Repeat until a stopping criterion is satisfied: either the target number of markers is reached, or a diminishing-returns threshold is detected (e.g., marginal gain in variance explained falls below a threshold). Export the final ranked list of m/z identifiers and their scores to a CSV file for validation and downstream use.

Related tools

• pandas (handle and export ranked marker-ion list to CSV; manage iteration state and scoring DataFrames)
• h5py (load latent low-dimensional peak features from HDF5 files output by graph-attention autoencoder)
• Graph-attention autoencoder (upstream feature extractor that produces the latent low-dimensional peak representations used for ranking) — https://github.com/zhanglabtools/SmartGate
• scanpy (general single-cell and spatial omics analysis framework integrated with SmartGate for downstream marker validation)

Evaluation signals

• Ranked list is exported as valid CSV with m/z identifiers, iteration count, and importance scores; no missing or malformed rows.
• Number of selected markers equals or approaches the target threshold (or stopping criterion is documented and justified).
• Ranking scores are monotonically decreasing or non-increasing across iterations, indicating consistent application of selection criterion.
• Selected m/z peaks show biological or chemical relevance (e.g., known metabolite masses or spatial co-localization patterns) when validated against reference databases or imaging data.
• Diminishing-returns stopping criterion is verifiable: marginal gain in variance or importance between successive iterations falls below the predefined threshold.

Limitations

• Skill is entirely dependent on the quality of latent features learned by the upstream graph-attention autoencoder; poor feature learning will produce uninformative marker rankings.
• No guidance is provided in the source material on how to set the target number of markers or the diminishing-returns threshold; these are data- and application-dependent hyperparameters.
• Stopping criterion (target number or diminishing returns) must be chosen a priori; the paper does not describe adaptive or data-driven stopping strategies.
• Ranking is based solely on latent-space statistics (variance, importance); biological validation or cross-validation against experimental metadata is external to this skill.

Evidence

• [other] SmartGate obtains latent low-dimension peak features from a graph-attention autoencoder and applies these features in an iterative automatic peak-picking process to identify marker ions in imaging mass spectrometry datasets.: "SmartGate obtains latent low-dimension peak features from a graph-attention autoencoder and applies these features in an iterative automatic peak-picking process to identify marker ions"
• [other] In each iteration, apply a peak-selection criterion (e.g., ranked by feature importance or variance in latent space) to identify the top-scoring m/z peaks not yet selected.: "apply a peak-selection criterion (e.g., ranked by feature importance or variance in latent space) to identify the top-scoring m/z peaks not yet selected"
• [other] Repeat steps 3–4 until a stopping criterion is met (e.g., target number of markers reached, or diminishing returns observed).: "until a stopping criterion is met (e.g., target number of markers reached, or diminishing returns observed)"
• [other] Export the final ranked list of marker-ion m/z identifiers to a CSV file.: "Export the final ranked list of marker-ion m/z identifiers to a CSV file"
• [readme] SmartGate could get latent low dimension peak features by Graph-attention autoencoder which help us achieve automic peak picking iteratively to find the marker ions.: "latent low dimension peak features by Graph-attention autoencoder which help us achieve automic peak picking iteratively to find the marker ions"