Interpretable Machine Learning to Decipher Myelodysplastic Syndrome-Associated Alterations of the Extracellular Matrix by Time-of-Flight Secondary Ion Mass Spectrometry Machine learning (ML) accelerates progress in many areas, including biomedical and clinical research. ML algorithms provide powerful options for efficiently analyzing multivariate data sets. We develo...

What would you probe with interpretable ToF-SIMS? 🤔

🧬 Early MDS from biopsies

🔬 Drug response via matrix shifts

🩸 Bone marrow biomarkers

Paper: pubs.acs.org/doi/10.1021/...

#MachineLearning #CancerBiology #NewMethod (3/3)

The problem: ECM changes in MDS are real but invisible.

ToF-SIMS sees the molecular fingerprints, but neural nets stay black-boxed.

Fix: Bayesian-tuned networks + SHAP surface the m/z peaks driving each class 🔬 (2/3)

Bone marrow whispers in its matrix and now we learned to listen.

From raw ion spectra to disease signatures — Bayesian neural nets + SHAP decode ECM chemistry 🧬🔬

A new @JASMS paper turns black-box classifiers into diagnostic windows on MDS 🧵 (1/3)

Quantifying uncertainty in protein representations across models and tasks - Nature Methods A model-agnostic empirical framework is proposed to measure the uncertainty associated with protein embeddings and to assess the biological relevance of these embeddings in order to improve model reli...

What would RNS flag in your PLM pipeline? 🤔

🧬 Filter low-confidence ESM-2 embeddings

🔬 QC for variant-effect prediction

🧠 Sanity-check novel/orphan proteins

Paper: www.nature.com/articles/s41...

#CompBio #DeepLearning #Bioinformatics (3/3)

The problem: protein embeddings are opaque.

Dense vectors from ESM-2 & ProtT5 — no way to know if they capture real biology or collapsed to noise.

Fix: RNS — fraction of synthetic random sequences among a protein's nearest neighbors 🔬 (2/3)

Your protein embedding looks fine. But is it biology—or noise?

From ESM-2 & ProtT5 to synthetic nearest-neighbors — a model-agnostic uncertainty score. 🧬📏

New NatureMethods: the Random Neighbor Score (RNS) tells you which to trust. 🧵 (1/3)

Highly accurate ab initio gene annotation with ANNEVO - Nature Methods ANNEVO advances accurate and scalable ab initio gene annotation of evolutionarily diverse genomes using deep learning approach modeling sequence evolution and long-range dependencies and mixture of ex...

What would you annotate with ANNEVO first? 🤔

🧬 Novel non-model genomes

🔬 Fix gaps in reference annotations

🧠 Phylogenetically rare species

Paper: www.nature.com/articles/s41...

Code: github.com/xjtu-omics/A...

#CompBio #DeepLearning #NewMethod (3/3)

The problem: ab initio gene annotation has lagged.

HMM-based tools miss long-range signals + can't learn across diverse genomes.

Fix: ANNEVO — a MoE genomic language model learning distal + evolutionary signal 🧬 (2/3)

Ab initio gene callers used to lag evidence-based pipelines. Not anymore.

From raw DNA → introns → gene structures across 566 species — one MoE language model. 🧬🔬

NatureMethods: ANNEVO jointly models distal + evolutionary signal 🧵 (1/3)

Multimodal foundation transformer models for multiscale genomics - Nature Methods This Perspective overviews recent and emerging developments in building and using multimodal foundation models based on transformers for analyzing various types of genomics data.

Top barrier in multimodal genomic AI? 🤔

🧬 Tokenization across DNA/cells/space

🔬 Cross-scale alignment

🧠 Interpretability

Share thoughts 👇

Paper: www.nature.com/articles/s41...

Code: github.com/Translationa...

#CompBio #DeepLearning #SpatialOmics (3/3)

❌ Siloed models miss cross-scale context

❌ Tokenization for DNA/cells/space still unsolved

✅ Super Transformer bridges all 3 scales via cross-attention

✅ Modular design → transfer learning across biology 🔬 (2/3)

One framework to unify genomics across scales 🧬🔭

"Super Transformer": modular cross-attention for multiscale biology

- DNA, scRNA-seq & spatial omics
- Unimodal → multimodal roadmap
- Masked modeling + contrastive learning

🧵 (1/3)

SMASH Imaging: A Serial Matrix-Assisted Laser Desorption/Ionization Mass Spectrometry Strategy for High-Resolution Imaging Facilitates Dual-Polarity and MS2 Spatial Lipidomics on a Single Tissue Secti... Matrix-assisted laser desorption/ionization mass spectrometry imaging (MALDI-MSI) is a key technology in spatial lipidomics that provides high sensitivity and spatial resolution. Because of the ioniza...

What would you SMASH image first? 🤔

🧠 Brain lipids at 5 µm

🫀 Cardiac lipid markers

🔬 Tumor lipid mapping

Share your ideas 👇

Paper: pubs.acs.org/doi/10.1021/...

#SpatialOmics #Lipidomics #Bioinformatics

(3/3)

❌ Standard MALDI-MSI: one polarity, one scan, limited info

❌ High-res imaging destroys sample fast

✅ SMASH reuses same pixels across 8 serial layers

✅ PASEF enables MS2 lipid ID at 5 µm

✅ Spatial correlation validates across layers

(2/3)

🧬🔬 SMASH Imaging maps 400+ lipids on ONE tissue section!

Serial MALDI-MSI strategy:

- Dual-polarity + MS2 on same pixels
- 30 µm → 5 µm resolution
- 8 imaging layers from 1 section

🧵 (1/3)

High-Resolution Microscope-Mode Secondary Ion Mass Spectrometry Imaging We report the development of a secondary ion mass spectrometry (SIMS) microscope-mode imaging instrument suitable for a wide range of applications in which high throughput is an advantage. By coupling...

Where would you use sub-5µm SIMS?

🧠 Brain lipid circuits

🦠 Single-cell metabolomics

💊 Drug tissue distribution

🔬 Correlative imaging

Reply with your use case 👇

Paper: pubs.acs.org/doi/10.1021/...

#SpatialOmics #BioImaging #Bioinformatics (4/4)

Challenges in SIMS imaging:

❌ Raster SIMS: slow, sequential

❌ High-res: tiny FOV

❌ Speed vs. resolution tradeoff

New microscope-mode solves it:

✅ All ions at once

✅ <5 µm + m/Δm ~6900

✅ mm tissue in minutes

Maps lipids + amino acids (3/4)

Old SIMS: good pixels, poor chemistry (~2000 m/Δm).

New instrument reveals:

WHERE: entire mm-scale tissue sections

WHEN: minutes per image

HOW: pulsed TOF + fast scintillator → 6900 m/Δm

Lipids + amino acids, both resolved. (2/4)

🔬 Map every lipid in a brain slice at sub-5 µm resolution — in minutes.

Microscope-mode SIMS: all ions imaged simultaneously, no raster scan.

- <5 µm spatial resolution
- m/Δm ~6900 mass resolution
- mm-scale tissue in minutes

🧵 (1/4)

stVCR: spatiotemporal dynamics of single cells - Nature Methods stVCR models and reconstructs single-cell dynamics of cell differentiation, proliferation and migration using time-series spatial transcriptome data.

What dynamics would you reconstruct?
🧠 Brain development/regeneration
🫀 Heart morphogenesis
🦎 Limb regeneration
🧬 Tumor invasion over time
🪺 Organoid growth

www.nature.com/articles/s41...

Key innovation: End-to-end spatiotemporal reconstruction

Result: Continuous trajectories showing differentiation + proliferation + migration simultaneously

Not just "pseudotime"—actual physical-space dynamics

Why this matters for developmental & regenerative biology:
- Track WHERE cells migrate in physical space
- See HOW phenotype transitions couple with spatial movement
- Handle real data messiness (in different coordinate systems)
- Works in 3D reconstruction

Reconstruct cell migration, differentiation & proliferation from spatial snapshots 🎬

stVCR: A generative deep learning framework that "plays back" tissue dynamics from time-series spatial transcriptomics

Aligning samples across different coordinate systems or time points

MSInet: A Self-Supervised CNN Framework Integrating Global and Local Context for Robust Mass Spectrometry Imaging Segmentation Mass spectrometry imaging (MSI) enables label-free molecular mapping in tissues but presents challenges for spatial segmentation due to high dimensionality, nonlinear spectral variation, and tissue he...

What's your biggest segmentation challenge?

🧠 Complex brain anatomy (many subregions)
🔬 Tumor heterogeneity (cancer vs. necrosis vs. healthy)
📊 Choosing cluster numbers (k=?)
🌀 Fragmented results

pubs.acs.org/doi/10.1021/...

#MassSpec #SpatialOmics #DeepLearning #Metabolomics

(4/4)

Key innovation: Dual-consistency learning

Challenge: MSI has high dimensions + nonlinear variation + tissue heterogeneity

MSInet solves both:
- Contrastive learning → global semantic relationships
- Superpixel refinement → local spatial consistency

(3/4)

Why this matters for spatial metabolomics:

- No predefined cluster numbers needed (self-supervised)
- Preserves spatial continuity (no fragmented segments)
- Works across platforms: MALDI-MSI, DESI-MSI
- Biologically meaningful boundaries (matches histology)

(2/4)

Mass spectrometry imaging segmentation just got a self-supervised upgrade🔬

MSInet: Annotation-free spatial segmentation combining:
- Patch-wise contrastive learning (global context)
- Superpixel-guided refinement (local boundaries)

Outperforms t-SNE+k-means, CNNAE, GCN-based methods by ~26%

(1/4)

AF2BIND: predicting small-molecule binding sites using the pair representation of AlphaFold2 - Nature Methods AF2BIND is a logistic regression model trained on AlphaFold2 pair features to predict small-molecule binding-site residues in proteins, without multiple sequence alignments, homology models or knowled...

What "undruggable" target would you tackle first?

Reply with:
🎯 Your target protein
🧬 Disease area
💊 Why it's been hard to drug

www.nature.com/articles/s41...

#DrugDiscovery #AlphaFold

Innovation: Repurposes AlphaFold2's learned representations for binding site prediction

Challenge: Previous methods need homology models OR known ligands OR multiple sequence alignments

AF2BIND: Just needs protein sequence → AF2 embeddings → logistic regression → binding sites

Why this matters for drug discovery:

- Predict binding sites WITHOUT known similar proteins (true de novo)
- Works on "undruggable" targets with no structural homologs
- Thousands of new therapeutic sites identified
- Predicts ligand chemical properties from binding pocket features