Best-practices alignment and literature-driven improvement leads for SpatialTranscriptFormer, with competitive context from the spatial transcriptomics literature.
These areas already follow industry best practices:
- Global Pathway Targets — Pre-computed using a standardized offline script (
stf-compute-pathways) and stored as slide-stationary matrices, eliminating gene-to-pathway conversion latency during training. - Spatially Variable Pathway Targets — Ground-truth targets represent curated MSigDB Hallmark pathway activity scores, which exhibit much higher spatial coherence than noisy individual gene expression values.
- Spatial Coherence Validation — Diagnostic Moran's I is computed for each pathway across slide spots to confirm predicted patterns retain biological spatial autocorrelation.
- Spatial Context via Neighbourhoods — KD-tree-based neighbour aggregation in
HEST_FeatureDatasetprovides local spatial context to patch features. - Coordinate Standardisation —
normalize_coordinates()prevents spatial scale bias between slides from different platforms (Visium, Visium HD, etc.). - Direct Pathway Supervision — Decoupled pathway target supervision eliminates circular auxiliary losses and noise from high-dimensional gene reconstruction.
- Histology-Pathway Integration — The quad-flow architecture follows recommended multi-modal integration patterns; pathway tokens act as a structured biological bottleneck analogous to the Perceiver cross-attention design.
Source: Yue et al. (2023), Spatial transcriptomics guidebook; Nat. Comms. benchmark 2025.
Visium is the right primary target. The guidebook classifies 22 ST technologies and confirms 10x Visium (55 µm spots, 1–10 cells/spot) as the dominant commercial platform driving the majority of downstream computational tool development. This validates HEST-1k as the appropriate training corpus.
The resolution ceiling is real but addressable. Visium spots are multi-cellular by design. GHIST (Nat. Methods, 2025), which leverages Xenium sub-cellular data, represents the current performance ceiling. Scaling to Visium HD or Xenium is an architectural non-event — it is purely a data availability and preprocessing question since the model is already coordinate-agnostic.
No method dominates across all criteria. The 2025 Nat. Comms. benchmark evaluated 11 methods across 5 datasets using 28 metrics in five categories: prediction accuracy, generalisability, translational potential, usability, and computational efficiency. No single method wins across all categories. This is the framework for external evaluation.
| Method | Venue | Relative Strength |
|---|---|---|
| HisToGene | — | Best generalisability across datasets |
| DeepSpaCE | — | Best generalisability across datasets |
| Hist2ST | NeurIPS 2023 | Generalisability; contrastive bi-modal training |
| EGNv2 | — | Highest per-platform accuracy on ST data |
| DeepPT | — | Highest accuracy on Visium specifically |
| GHIST | Nat. Methods 2025 | Single-cell resolution via Xenium |
STF's differentiators in this landscape are pathway-level interpretability and direct biological (pathway) supervision — both under-represented in the 2025 benchmark's translational potential category, where competing methods score poorly.
Session: March 2026. Sources: Schubert et al. 2018; Badia-I-Mompel et al. 2022; Dimitrov et al. 2022/2024; Tanevski et al. 2022; Müller-Dott et al. 2023.
The Saez-Rodriguez lab (EMBL Heidelberg / RWTH Aachen) maintains a family of interoperable prior knowledge resources and statistical tools that are directly relevant to the SpatialTranscriptFormer pathway bottleneck.
PROGENy (Schubert et al., Nat. Comms. 2018) — 14 cancer-relevant signalling pathways (EGFR, MAPK, PI3K, TGFb, TNFa, Hypoxia, VEGF, WNT, p53, NFkB, JAK-STAT, Androgen, Estrogen, Trail) derived empirically from a large compendium of perturbation experiments. Each gene carries a continuous signed weight (direction and magnitude of pathway response). This is fundamentally different from MSigDB Hallmarks, which use binary unweighted membership. PROGENy weights are a natural warm-start for the learnable pathway tokens (see item 9). Holland et al. (Genome Biology, 2020) benchmarked PROGENy on scRNA-seq and Visium data, confirming it performs well in the spot-level transcriptome regime.
decoupleR (Badia-I-Mompel et al., Bioinformatics Advances, 2022) — A Python/R
package that applies a family of statistical estimators (MLM, ULM, GSEA, AUCell,
consensus) to any (resource, method) combination. get_progeny() and run_mlm()
provide PROGENy activity scores out of the box. The official documentation includes
a Visium spatial transcriptomics notebook. This is the practical bridge for computing
weighted pathway activity scores from HEST-1k expression matrices — a richer
alternative to the current Hallmark mean-aggregation targets (see item 7).
CollecTRI (Müller-Dott et al., Nucleic Acids Research, 2023) — The recommended successor to DoRothEA for transcription factor regulons. Integrates 12 resources for 1186 human TFs with signed interactions, consistently outperforming other regulon collections in perturbation benchmarks. TF activity tokens are a longer-term architectural direction: complementary to pathway tokens, operating at a finer regulatory resolution.
LIANA+ (Dimitrov et al., Nature Cell Biology, 2024) — Extends LIANA's
ligand-receptor benchmarking to spatially resolved data using location-weighted cosine
similarity scoring. The spatially-weighted LR scoring is conceptually aligned with
the h2p / p2h attention flows in STF's quad-flow model. LIANA+ LR pairs from
OmniPath are the highest-quality resource for supervising intercellular communication
pathway tokens (see item 8).
MISTY / mistyR (Tanevski et al., Genome Biology, 2022) — A multiview spatial
analysis framework that decomposes how much of a marker's variance is explained by
intrinsic vs. juxtaposition vs. neighbourhood views. Not a prediction tool, but
useful as a diagnostic: running MISTY on HEST-1k training data would reveal
whether PROGENy pathway activities are primarily driven by local (intrinsic) or
neighbourhood (para) context, informing how strongly the h2h interaction flow
matters for pathway token learning.
OmniPath (Türei et al., Mol. Systems Biology, 2021) — The meta-database
underpinning the entire Saezlab ecosystem (PROGENy, DoRothEA, LIANA). Aggregates
100+ resources covering directed signed protein interactions, enzyme-PTM, protein
complexes, pathway annotations, and ligand-receptor pairs. Accessible via the
omnipath Python package as a single programmatic interface.
SPathDB (Li et al., Nucleic Acids Research, 2025) — Pre-computed pathway activity scores for 114,998 pathways across 695 spatial transcriptomics slices from 84 public datasets. Not a Saezlab product, but relevant as a reference atlas: STF pathway token outputs can be validated against SPathDB ground-truth pathway activity maps for the same tissue types.
Coverage trade-offs by resource:
| Resource | Pathways/TFs | Gene weights | Spatial validation |
|---|---|---|---|
| MSigDB Hallmarks (current) | 50 | Binary | No |
| PROGENy | 14 signalling | Continuous, signed | Yes (Visium, decoupleR) |
| CollecTRI | ~1186 TFs | Continuous, signed | Indirect |
| Combined (PROGENy + Hallmarks) | 64 total | Mixed | Partial |
The 14 PROGENy pathways cover signalling cascades; the 50 Hallmarks cover broader biological programs (proliferation, metabolism, immunity). A hybrid initialisation using PROGENy weights for signalling tokens and Hallmark membership for the remaining tokens has not been published but is a natural next step.
Status key: ✅ Implemented | 🔧 Open | 💡 Research lead
1. Mitochondrial gene filtering ✅ — Completed
Mitochondrial genes (MT-*) reflect cell health and apoptotic state, not spatial morphological patterns. Standard spot-level QC filters out mitochondrial reads.
Implementation: The stf-compute-pathways pipeline filters out mitochondrial reads based on a configurable max MT fraction (--qc-max-mt, default: 0.15) during target pre-computation.
2. SVG-weighted vocab rebuild 🚫 — Superseded
Superseded by the transition to a pathway-exclusive prediction paradigm, removing gene-level vocabulary alignment entirely.
3. Vectorise morans_i calculations 🚫 — N/A
Computing spatial autocorrelation over 50 pathway activities is computationally negligible, eliminating the need to vectorise high-dimensional gene-level Moran's I loops.
4. Dispersion-based (HVG) gene filtering 🚫 — Superseded
Superseded by the transition to a pathway-exclusive prediction paradigm.
5. Standardised library-size normalisation ✅ — Completed
Spots are normalised to 10,000 counts (CP10k) before applying log1p during the offline pre-computation step.
6. Per-spot quality control ✅ — Completed
Explicit QC thresholds (minimum UMI count, minimum detected genes, maximum mitochondrial fraction) are enforced at target pre-computation time via stf-compute-pathways.
Architectural direction: pathway activity as the primary task. ✅ Implemented. This was the central reframing of the project and is now the shipped architecture: the model predicts pathway activity directly via a dot-product + Softplus head, with no gene-reconstruction step. The rationale is kept here as a record of why the change was made.
Earlier prototypes predicted ~1000 individual gene expression values with pathway scores as an auxiliary output. That framing had two compounding problems:
-
Noisy learning signal. Even after log-normalisation and SVG-based vocabulary selection, MSE over 1000 genes is dominated by high-expression genes — not necessarily the spatially informative ones. Moran's I weighting fixes gene selection but not the gradient during training.
-
The (former)
AuxiliaryPathwayLosswas circular. Its targets were computed from the same gene expression being predicted (MSigDB membership sums), so it constrained the intermediate representation to match a re-parameterisation of the output, adding no independent biological information.
The fix was to invert the task hierarchy:
PRIMARY: H&E → pathway activity maps (spatially coherent, interpretable)
SECONDARY: pathway scores → gene expression (optional; not currently implemented)
Pathway activity targets (currently the mean of log1p CP10k expression over each MSigDB Hallmark's member genes, pre-computed offline) are:
- Spatially cleaner than individual genes — Moran's I on pathway aggregates is typically 3–5× higher than on individual gene expression
- Adaptable — the offline preprocessing step can be swapped for any biologically informed prior (PROGENy for signalling, Hallmarks for broader programs, LIANA+ for intercellular communication, CollecTRI for TF activity) without changing the model
- Genuinely supervised — the pathway scores are first-class outputs, not constraints on an intermediate representation
The remaining open leads below concern richer target backends (decoupleR/PROGENy) and an optional secondary gene head — not the core reframing, which is done.
7. Richer pathway-activity targets (decoupleR + PROGENy) 💡 — High priority, medium effort
The offline target pipeline (stf-compute-pathways) currently aggregates MSigDB
Hallmark membership (mean log1p CP10k). A higher-quality alternative: for each HEST
sample, run decoupler.run_mlm(expression, net=get_progeny()) to produce a
(spots × 14) signed, continuous pathway activity matrix and cache it as an
alternative target set. Because targets are decoupled from the model, this is a
preprocessing-only change — no model edits required.
The decoupleR MLM estimator fits: expression[:, g] = a_p · W[g, p] + b_p + ε and
solves for a_p — the pathway activity at each spot — via OLS. The result is
continuous, signed, and validated against known pathway perturbation experiments. The
prior knowledge source (PROGENy, Hallmarks, LIANA+ LR pairs, CollecTRI TF regulons)
can be swapped at the preprocessing step without touching the model.
8. Moran's I weighted gene loss 🚫 — Superseded (pathway-exclusive)
Only relevant if an optional secondary gene-reconstruction head is ever added. In that case each gene's loss contribution could be weighted by its Moran's I so spatially coherent genes drive the gradient:
L_gene = sum(w_g * MSE(pred_g, target_g) for g in genes) / sum(w_g) # w_g = MoranI_gFor the current pathway-exclusive model this does not apply — there is no gene head.
9. PROGENy pathway-token initialisation 💡 — Medium priority, medium effort
The learnable pathway tokens are currently randomly initialised. A lead: warm-start them from PROGENy's continuous signed footprint weights (each gene carries a directional, perturbation-validated weight) instead of from scratch.
Hybrid strategy: initialise the first 14 tokens from PROGENy weights (signalling) and
the rest from Hallmark membership (metabolic, immune, proliferation). get_pathway_init
in data/pathways.py already builds a membership matrix; this would re-introduce a
token-initialisation pathway into the model and a corresponding option in
training/builder.py (the earlier --pathway-init flag was removed in the
pathway-exclusive cleanup and would need re-adding).
10. Cell–cell interaction pathway tokens (LIANA+) 💡 — Low priority, high effort
Under the primary-task framing, a subset of pathway tokens could be supervised by LIANA+ spatially-weighted ligand-receptor scores (from OmniPath) rather than intracellular pathway activity. This would produce communication tokens encoding intercellular signalling alongside the intracellular pathway tokens — a natural extension of the adaptable preprocessing approach in item 7.
11. Cell-type deconvolution secondary head 💡 — Medium priority, high effort
Pre-computed cell2location or RCTD deconvolution proportions could serve as an additional secondary prediction head alongside gene expression — useful for datasets where cell-type spatial organisation is the primary question.
12. Evaluate on the Nat. Comms. 2025 benchmark suite 🔧 — High priority, medium effort
The 2025 benchmark provides a ready-made evaluation framework: 5 public datasets, 28 metrics, 5 categories. SpatialTranscriptFormer has not yet been evaluated on this suite. Submitting results, particularly in the translational potential category (where pathway interpretability is the differentiator), is the clearest path to situating STF in the published competitive landscape.
13. Visium HD / Xenium data 💡 — Low priority, high effort
The architecture is platform-agnostic. Scaling to sub-cellular resolution platforms (Xenium: ~10 µm, Stereo-seq: 0.22 µm) is blocked only by data availability and preprocessing pipelines, not the model itself. GHIST demonstrates that Xenium-derived training labels substantially improve resolution.
14. Preprocessing data contract ✅ — Largely implemented
Each stf-compute-pathways output .h5 records its provenance as attributes
(format_version, QC thresholds, spot counts, scored-pathway count; see
PATHWAY_MAPPING.md), and save_pretrained stamps
pathway_format_version into config.json so load_pretrained / --resume refuse
checkpoints trained against incompatible targets. Remaining nicety: surface the same
provenance (QC params, --pathways subset, run date) in a per-run summary for quick
human inspection.
| # | Direction | Priority | Effort | Status |
|---|---|---|---|---|
| 1 | MT-gene filter in stf-compute-pathways |
High | Low | ✅ Implemented |
| 2 | Rebuild vocab (Superseded by pathway targets) | High | Low | 🚫 Superseded |
| 3 | Vectorise morans_i calculations |
Medium | Low | 🚫 N/A (Low Dim) |
| 4 | HVG dispersion-based gene filtering | Medium | Medium | 🚫 Superseded |
| 5 | Library-size normalisation (CP10k) | Medium-High | Medium | ✅ Implemented |
| 6 | Per-spot QC thresholds | Medium | Medium | ✅ Implemented |
| 7 | Pre-compute pathway targets (decoupleR + PROGENy) | High | Medium | 💡 Lead |
| 8 | Moran's I weighted gene loss | High | Low | 🚫 Superseded |
| 9 | PROGENy pathway token initialisation | High | Medium | 💡 Lead |
| 10 | Cell–cell interaction tokens (LIANA+) | Low | High | 💡 Lead |
| 11 | Cell-type deconvolution secondary head | Medium | High | 💡 Lead |
| 12 | Nat. Comms. 2025 benchmark evaluation | High | Medium | 🔧 Open |
| 13 | Scale to Visium HD / Xenium | Low | High | 💡 Lead |
| 14 | Preprocessing data contract doc | Low | Low | ✅ Implemented |
- Yue et al. (2023). A guidebook of spatial transcriptomic technologies, data resources and analysis approaches. CSBJ, 21, 940–955. DOI | PMC10781722
- Benchmarking the translational potential of spatial gene expression prediction from histology. Nat. Comms. 2025, 16(1):1544. DOI | PMC11814321
- GHIST: Spatial gene expression at single-cell resolution from histology. Nat. Methods 2025. PMC12446070
- Deep learning in spatially resolved transcriptomics: a comprehensive technical view. Briefings in Bioinformatics 2024. Oxford Academic
- Single-cell Best Practices (Theis Lab): sc-best-practices.org
- Squidpy: squidpy.readthedocs.io
- Schubert et al. (2018). PROGENy: perturbation-response genes reveal signaling footprints in cancer. Nat. Comms. 9:20. DOI | PMC5750219
- Holland et al. (2020). Robustness of TF and pathway tools on single-cell RNA-seq. Genome Biology 21:36. DOI | PMC7017576
- Badia-I-Mompel et al. (2022). decoupleR: ensemble methods to infer biological activities. Bioinformatics Advances 2(1):vbac016. DOI | PMC9710656
- Müller-Dott et al. (2023). CollecTRI: expanding TF regulon coverage for activity estimation. Nucleic Acids Research 51(20):10934. DOI | PMC10639077
- Dimitrov et al. (2022). LIANA: comparison of cell-cell communication methods. Nat. Comms. 13:3224. DOI | PMC9184522
- Dimitrov et al. (2024). LIANA+: all-in-one framework for cell-cell communication. Nature Cell Biology 26(9):1613. DOI | PMC11392821
- Tanevski et al. (2022). MISTY: multiview framework for spatial relationships. Genome Biology 23:97. DOI | PMC9011939
- Türei et al. (2021). OmniPath: integrated intra- and intercellular signaling knowledge. Mol. Systems Biology 17(3):e9923. DOI | PMC7983032
- Li et al. (2025). SPathDB: spatial pathway activity atlas. Nucleic Acids Research 53(D1):D1205. DOI | PMC11701687