Hyperbolic Contrastive Learning with Entailment for Spatial Transcriptomics
Daniela Vega, Paula C\'ardenas, Hannah Ceballos, Leonardo Manrique, Pablo Arbela\'ez
cs.CV
Quick verdict
Hyperbolic contrastive+entailment achieves SpaRED SOTA (-6% MSE, +8% PCC) for image-to-gene-expression prediction, though gains may conflate geometry with stronger encoders.
Summary
The paper introduces HyCLoST, a framework that predicts spatially resolved gene expression from H&E histopathology patches by aligning image and gene representations in a shared hyperbolic (Lorentz) space. Histology patches and expression vectors are encoded with pretrained foundation models (UNI2-h, Geneformer), projected to 512 dimensions, and mapped to a hyperboloid with learned curvature. Training combines a symmetric contrastive loss using negative hyperbolic distance as logits with a gene-to-image entailment loss built on hyperbolic semantic cones, encoding the claimed asymmetry that expression determines morphology. Inference is non-parametric: each test patch retrieves the k=80 nearest training gene embeddings and averages their expression vectors. On the 26-dataset SpaRED benchmark (14 human, 12 mouse; top 128/32 spatially autocorrelated genes; LGDiST preprocessing), HyCLoST reports average MSE 0.899 and PCC 0.431, a ~6% MSE reduction and ~8% PCC gain over the second-best of nine re-implemented baselines, ranking first on 10/26 datasets, with Friedman and Holm-corrected Wilcoxon tests (not significant vs SEPAL). Ablations attribute gains to entailment weighting, Geneformer, and hyperbolic-vs-Euclidean geometry (0.899 vs 1.043). Main limitations: no baseline re-runs with foundation-model encoders confound geometry with backbone quality; significance is tested for MSE only; averaging 80 retrieved neighbors may itself smooth predictions; hyperparameter selection and per-dataset detail are thin.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Projecting histology-patch and gene-expression embeddings into a shared Lorentz hyperbolic space with learned curvature improves contrastive alignment for spatial gene expression prediction.
- A gene-to-image entailment loss on hyperbolic semantic cones, encoding the asymmetry that expression determines morphology, improves accuracy, and the gene-to-image direction beats image-to-gene in 21 of 26 datasets (0.899 vs 0.919 MSE).
- HyCLoST reduces average MSE by ~6% and increases PCC by ~8% over nine re-evaluated methods across the 26 SpaRED datasets (MSE 0.899 +/- 0.436, PCC 0.431 +/- 0.176), ranking first on 10 of 26 datasets.
- With all other components fixed, hyperbolic geometry outperforms Euclidean space (average MSE 0.899 vs 1.043), supporting hierarchy-aware representation as the source of the gains.
Strengths
Unified, statistically tested benchmark re-evaluation
Nine SOTA methods are re-implemented across all 26 SpaRED datasets (Tab. 1) and compared with a Friedman test (chi2=117.02, p<10^-19) plus Holm-Bonferroni-corrected paired one-sided Wilcoxon tests on MSE, directly addressing the inconsistent prior evaluations flagged in Sec. 1.
Matched geometry ablation isolates the geometric claim
Tab. 2 rows 5 vs 6 swap hyperbolic for Euclidean space with all other components fixed (avg MSE 0.899 vs 1.043), which is the correct controlled comparison for the central hypothesis, rather than relying on cross-paper numbers.
Transparent handling of baseline evaluation flaws
Directionality of entailment is tested, not just asserted
Headline numbers are internally consistent and modest
Weaknesses
Backbone-method confound is not resolved
Statistical rigor is asymmetric across metrics
Retrieval averaging may reintroduce the failure mode the paper targets
Thin hyperparameter and validation protocol
Self-referential evaluation stack with no external validation
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is HyCLoST and what does it do?
HyCLoST is a deep-learning model that predicts spatially resolved gene expression directly from H&E-stained histopathology images. It encodes image patches and gene-expression vectors with pretrained foundation models (UNI2-h and Geneformer), aligns them in a shared hyperbolic space using a contrastive loss plus a directional gene-entails-image entailment loss, and predicts a test patch's expression by averaging the expression vectors of its 80 nearest gene embeddings in training data.
How much better is HyCLoST than previous methods?
On the 26 public datasets of the SpaRED benchmark it reports an average MSE of 0.899 and PCC of 0.431, roughly a 6% MSE reduction versus the second-best method (EGGN, 0.953) and an 8% PCC increase versus the best prior PCC (ST-Net, 0.400). It ranks first on 10 of 26 datasets and significantly beats 10 of 11 baseline configurations on MSE with corrected statistical tests; the exception is SEPAL.
Is HyCLoST's code and data available?
Yes. The paper states the source code is publicly available on GitHub (listed as BCV-Uniandes/HyCLoST in the abstract and dvegaa00/HyCLoST in the introduction). Training data come from the public SpaRED benchmark of 26 datasets, and the encoders are public pretrained models (UNI2-h, Geneformer), so the full pipeline is in principle reproducible.
Why use hyperbolic space for gene expression prediction?
The authors argue biological systems are hierarchical - gene regulation constrains tissue morphology - and hyperbolic geometry represents hierarchies more naturally than Euclidean space of the same dimension. In their ablation, swapping hyperbolic for Euclidean space while keeping everything else fixed raises average MSE from 0.899 to 1.043 across the 26 test datasets, which they take as evidence that the geometry itself contributes to accuracy.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1-2 could not be inspected; the qualitative over-smoothing claims in Sec. 3.3 rest on figure content that is not verifiable from the extracted text.
- Per-dataset result tables, individual Wilcoxon p-values, and any supplementary statistics were not part of the extracted text.
- Repository contents were not verified, and the two cited GitHub URLs differ from each other.
- Equation typography (notably Eqs. 6-7) was partially mangled by text extraction; the entailment-cone math was interpreted from context following reference [6] (MERU).
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.