BGPT: Author Review: Raf Van De Plas

Explore by Goal

Quick Explanation Copied

Raf Van De Plas — scientific strengths (from 2 provided raw-data–grounded papers): rigorous, interpretable ML for imaging mass spectrometry (IMS), explicit attempts to separate “content” from “signal strength,” and chemistry-aware dimensionality reduction with automated isotope-candidate detection. Key methods: β-VAE latent-space decomposition (SiSAL) and chemistry-informed NMF-based DR (IsoSpace).

Long Explanation

Author Scientific Strength Review: Raf Van De Plas

Evidence scope: only the 2 provided IMS/ML research records (SiSAL; IsoSpace). OpenAlex retrieval timed out, so author citation metrics are not assessed here.

What is “known” from the provided material (and only that):

SiSAL: a β-VAE workflow that learns a 2D latent space intended to disentangle molecular content from signal strength, with latent regions mapped back to tissue structures and ranked molecular differentiators in human kidney IMS datasets (HuBMAP) plus a synthetic validation case.
IsoSpace: a chemistry-informed dimensionality reduction for IMS that combines NMF-based spatial/spectral decomposition with isotope-candidate pattern extraction using an m/z pattern matrix, targeting isotope-ladder–like spectral relationships while preserving chemical interpretability across mouse, rat, and human kidney IMS datasets.

VISUAL 1 — Dataset scale + peak complexity (from the provided records)

Pixel counts (analysis granularity) and number of measured peaks (feature dimensionality) across the two projects’ listed datasets.

Color key: deep blue = pixels; light blue = peaks.

Note on interpretation: these are counts reported in the provided summaries (not re-derived from raw files here). For any ML conclusions, generalization hinges on preprocessing, instrument settings, and peak-picking consistency (uncertain from the provided excerpts).

VISUAL 2 — Claimed conceptual separation vs chemistry-aware isotope structure

A compact “mechanism comparison” diagram translating each work’s stated core idea into a reproducible analytical objective.

The “failure modes” line blends what is explicitly stated in the provided record with cautious interpretation about what such methods typically require; any statement of method limitations is therefore grounded only where the provided summaries mention them.

Scientific strength appraisal (critical, evidence-based)

1) Methodological clarity + reproducibility intent

Both records explicitly describe algorithmic components (β-VAE latent space with uncertainty; KDE-defined latent ROIs; latent traversal mapping for SiSAL) and (NMF decomposition; isotope-like m/z pattern filtering via 1±ω; IPC scoring; missing-peak handling via 2ω) for IsoSpace.
SiSAL is linked to a stated code artifact (“SiSAL Python package” on GitHub) and includes public dataset DOIs for case studies via HuBMAP.

Skeptical note: code availability and public datasets help, but “reproducibility” still depends on exact preprocessing/peak-picking, hyperparameter choices (ω, KDE bandwidth, β-VAE architecture), and whether full m/z spectra are available beyond selected differentiators—which is not fully verifiable from the provided excerpts.

2) Biological interpretability attempts (not just prediction)

SiSAL claims latent regions map to renal tissue structures (e.g., glomeruli, proximal tubules) and that latent differences can be traced back to lipid m/z features that differentiate ROIs across multiple donors/case studies.
IsoSpace claims isotope-candidate pattern extraction that preserves chemical interpretability (i.e., m/z distance–structured families) while reducing dimension; it reports specific dataset-scale results (e.g., 11.5% dimensionality reduction in the provided figure summary).

Skeptical note: interpretability via decoder/traversal (SiSAL) is sensitive to model mis-specification; isotope-candidate detection (IsoSpace) may generate chemical-plausible but biologically irrelevant patterns if ω/IPC thresholds are not carefully calibrated and validated across instrument conditions.

3) “Evidence strength” limits from the provided record

No incoming citation counts were provided for either record; thus claims like “impact” or “community uptake” cannot be assessed here. (OpenAlex timeout prevented citation metric retrieval.)
Both records emphasize methodological validation (synthetic case for SiSAL; cross-dataset demonstrations for IsoSpace), but the provided excerpts do not include full statistical power details (e.g., distributional assumptions, number of independent experimental replicates beyond pixel-level counts, and uncertainty quantification for biological mapping).

Confidence note: given the excerpted nature of evidence, scientific rigor is judged mainly on (i) algorithmic specificity, (ii) interpretability mechanisms, and (iii) stated limitations—not on full end-to-end reproducibility and statistical validation.

VISUAL 3 — SiSAL: reported “structure separation” footprint (from provided m/z differentiators list)

This chart summarizes the provided illustrative lipid m/z differentiators and which tissue subregion they are said to be associated with (only as listed in the record; not complete spectrum).

This visualization is limited to the explicit m/z examples provided (not a full set of discriminants). For full evidence, the complete differentiator tables and uncertainty/ranking criteria would need to be checked in the original manuscript materials.

Science-actions (optional next steps in BGPT)

Feedback:

Updated: March 25, 2026