PaperlayerView exact revision on arXiv

Source-linked AI summary

OmicSync: Reliability-Aware Spatial Multi-Omics Clustering with Evidence-Constrained LLM Reasoning

Rabeya Tus Sadia, Qiang Ye, Qiang Cheng

arXiv:2608.22785v1cs.CV

TL;DR

Spatial domain discovery often lacks assignment reliability, modality attribution, and auditable explanations. OmicSync couples multimodal clustering with evidence-constrained reasoning, while OmicSync-R feeds reasoning-quality rewards back through REINFORCE; OmicSync ranks best on three of four benchmarks, and OmicSync-R improves Human Breast Cancer ARI from 45.73 to 46.72.

  • Problem

    Existing spatial domain discovery methods generally provide cluster assignments without indicating reliability, modality contributions, or why individual domain decisions should be trusted.

  • Method

    OmicSync combines multimodal spatial clustering with model-derived confidence, epistemic uncertainty, and modality-routing evidence for constrained per-spot reasoning; OmicSync-R uses reasoning-quality scores as REINFORCE rewards.

  • Results

    OmicSync achieves the best average rank on three of four CytAssist FFPE benchmarks, while OmicSync-R improves Human Breast Cancer ARI from 45.73 to 46.72 and surpasses existing methods on six of nine metrics.

  • Takeaways & Limitations

    Together, the methods provide reliability-aware, spot-level auditable spatial-domain discovery guided by evidence-constrained reasoning.

  • Takeaways & Limitations

    Modality-routing weights are proxies rather than causal measures of modality contribution, and OmicSync-R is resource intensive and evaluated on one dataset.

Abstract

from arXiv · show

Spatial multi-omics technologies jointly profile gene expression, surface proteins, and histology at each tissue spot, yet most spatial domain discovery methods provide only cluster assignments, without indicating assignment reliability, modality contributions, or why a domain decision should be trusted. We present OmicSync, a reliability-aware spatial multi-omics framework that couples unsupervised domain clustering with evidence-constrained LLM reasoning using model-derived per-spot signals, including assignment confidence, epistemic routing uncertainty, and modality-routing weights. These signals are converted into structured evidence dictionaries and used to generate standard, stepwise, counterfactual, contrastive, and uncertainty-focused explanations. OmicSync integrates a KAN-GCN backbone with spatial encoding, cross-modal fusion, uncertainty-aware routing, cell-type supervision, and missing-modality imputation. We further introduce OmicSync-R, which closes the reasoning-clustering loop by using automatically computed reasoning-quality scores as REINFORCE rewards, allowing reasoning coherence to shape the latent structure without backpropagating through the language model. Across four 10x CytAssist FFPE spatial proteomics benchmarks, OmicSync achieves the best average rank on Human Tonsil (1.44), Glioblastoma (1.78), and Tonsil Add-on (1.22), and second-best on Human Breast Cancer (2.33). OmicSync-R further improves ARI on Human Breast Cancer from 45.73 to 46.72 and outperforms existing methods on six of nine clustering metrics. Together, OmicSync and OmicSync-R enable reliability-aware, spot-level auditable spatial domain discovery guided by evidence-constrained reasoning.

1 Introduction

Spatial multi-omics domain discovery supports biologically coherent tissue analysis, but existing methods generally expose only domain partitions. OmicSync adds per-spot reliability signals and evidence-constrained reasoning to make assignments auditable.

  • Spatial multi-omics jointly profiles gene expression, surface proteins, histology, and spatial context at tissue spots.
  • Existing spatial domain methods generally return flat spot partitions without assignment reliability, modality contributions, or assignment explanations.
  • Treating all assignments as equally reliable can propagate clustering errors into downstream biological conclusions, especially for boundary or degraded spots.
  • OmicSync extracts assignment confidence, epistemic routing uncertainty, and modality-routing weights for each spot.
  • OmicSync-R uses automatically computed reasoning-quality scores as REINFORCE rewards so reasoning coherence can shape latent structure without language-model backpropagation.

2 Background and Related Work

OmicSync builds on multimodal spatial-domain methods and extends them with reliability estimation, evidence-constrained reasoning, and reasoning-guided refinement. Its design combines interpretable model signals with non-differentiable reasoning feedback.

  • Spatial multi-omics domain discovery: GROVER provides OmicSync’s closest architectural foundation through KAN graph convolutions over spatial and feature adjacencies with attention-based multimodal fusion.
  • Spatial multi-omics domain discovery: OmicSync adds reliability estimation, uncertainty-aware routing, evidence-constrained reasoning, and reasoning-guided model refinement to this backbone.
  • Explainability in spatial omics: Existing explainability methods can identify influential inputs but generally do not provide spot-specific clustering explanations grounded in multiple internal reliability signals.
  • Uncertainty-aware multimodal learning: MC dropout estimates epistemic uncertainty by repeatedly evaluating a dropout-equipped mixture-of-experts gating network.
  • Reinforcement learning with non-differentiable feedback: OmicSync-R treats automatically computed reasoning-quality scores as REINFORCE rewards that influence clustering without gradients through the language model.

3 Problem Formulation

The problem formulation represents each spatial spot with multimodal features, coordinates, adjacency structures, latent representations, domain probabilities, and reliability signals. It also defines optional annotation and evidence-constrained reasoning tasks.

  • Inputs and graph structures: Each spot is represented by RNA, ADT, histology, and spatial-coordinate inputs together with spatial and modality-specific feature adjacency matrices.
  • Task A: Spatial domain clustering and reliability estimation: Task A learns a shared latent representation and a soft assignment matrix whose rows are probability distributions over spatial domains.
  • Task A: Spatial domain clustering and reliability estimation: Each spot receives a hard domain assignment and confidence derived from the largest soft-assignment probability.
  • Task A: Spatial domain clustering and reliability estimation: The model also estimates epistemic routing uncertainty and modality-routing weights as core reliability signals.
  • Task B: Optional annotation and auxiliary regularisation: When annotations are unavailable, RNA-derived Leiden pseudo-labels make Task B an auxiliary consistency regularizer rather than an independently validated cell-type task.
  • Task C: Evidence-constrained reasoning: Task C constructs per-spot evidence dictionaries from model outputs and intermediate quantities, then generates evidence-supported reliability reports.

4 OmicSync Architecture

OmicSync preprocesses multimodal inputs, encodes spatial and modality-specific structure, fuses modalities with attention, routes evidence under uncertainty, and produces clustered latent representations with reliability outputs.

  • Overview: The workflow converts RNA, ADT, histology, and coordinates into modality-specific features and neighborhood graphs before shared representation learning.
  • Input preprocessing: RNA uses library-size normalization, log transformation, highly variable-gene selection, and PCA; ADT uses CLR normalization followed by PCA.
  • Input preprocessing: Histology is represented by L2-normalized 1,024-dimensional UNI image embeddings extracted from 224 × 224 H&E patches.
  • Spatial position encoding: Sinusoidal spatial encoding is added before graph convolution so identical molecular profiles at different tissue locations receive distinguishable representations.
  • Modality-specific encoding: Shared-weight KAN-GCN encoders process spatial and feature adjacency matrices, and within-modality attention merges the resulting embeddings.
  • Cross-modal fusion: A Transformer lets RNA, protein, and histology tokens from the same spot attend to one another before fusion.
  • Uncertainty-aware routing: MC-Dropout evaluates the uncertainty-aware gating network across stochastic samples to estimate epistemic uncertainty during modality routing.
  • Uncertainty-aware routing: The router thresholds mean gate values, prunes low-weight experts, and records routing weights as interpretable proxies for modality contributions.

4.7 Component 6: CellTypeHead and ImputationHeads

OmicSync combines cell-type prediction and modality-specific imputation with spatial, contrastive, clustering, and reconstruction objectives. Adaptive spatial exclusion and contrastive-weight scheduling promote coherent representations while preserving domain boundaries.

  • 4.7 Component 6: CellTypeHead and ImputationHeads: CellTypeHead predicts per-spot class logits, while three ImputationHeads reconstruct masked entries from the fused latent representation.Imputation regularizes the latent space by preserving information useful for recovering partially observed modalities.
  • 4.8 Training objective: OmicSync’s training objective combines reconstruction, contrastive, clustering, cell-type, imputation, and routing-uncertainty losses.The contrastive term uses topology-aware InfoNCE, while reconstruction and imputation use mean-squared-error objectives.
  • 4.8 Training objective: Spatial exclusion removes nearby spots from contrastive negatives, using nhops = 2 for homogeneous tissue with average neighbor similarity above 0.6.This reduces the risk of treating spatially adjacent, biologically similar spots as negatives.
  • 4.8 Training objective: The contrastive weight stays at its base value during early training and then decays linearly to λc/3, shifting later emphasis toward clustering and reconstruction.The schedule establishes local spatial consistency without overly constraining RNA-defined domain boundaries.

5 Task B: Cell-type regularisation and spatial prediction maps

Task B uses pseudo-labels for semi-supervised cell-type regularization rather than independent supervised annotation. After training, CellTypeHead outputs are presented as spatial pseudo-cell-type prediction maps.

  • 5 Task B: Cell-type regularisation and spatial prediction maps: Because matched external cell-type annotations are unavailable, CellTypeHead functions as a semi-supervised regularization branch.It is not treated as an independent supervised cell-type annotation benchmark.
  • 5 Task B: Cell-type regularisation and spatial prediction maps: Leiden RNA communities provide pseudo-labels, with 70% used for auxiliary training and 30% held out.The auxiliary cross-entropy loss encourages the shared latent space to preserve local pseudo-cell-type organization.
  • 5 Task B: Cell-type regularisation and spatial prediction maps: CellTypeHead predictions are reported qualitatively through spatial pseudo-cell-type maps rather than as an independent supervised benchmark.The head is applied to all spots after training.

6 Task C: Coupled Reasoning

Task C converts model-derived spot-level reliability and biological-context signals into evidence-constrained explanations. It supports multiple explanatory formats while keeping the language model limited to structured evidence supplied by the clustering model.

  • 6 Task C: Coupled Reasoning: Task C is a post-hoc module that translates quantitative spot-level outputs into human-readable explanations without adding a trainable component.Inputs include domain assignment, confidence, modality weights, uncertainty, marker evidence, and neighborhood composition.
  • 6 Task C: Coupled Reasoning: The language model receives only an evidence dictionary and is instructed not to introduce unsupported biological entities or claims.The dictionary includes the runner-up domain, confidence margin, modality-routing vector, dominant modality, epistemic uncertainty, top evidence features, and five-nearest-spot composition.
  • 6 Task C: Coupled Reasoning: Five prompt templates generate standard, stepwise, counterfactual, contrastive, and uncertainty-focused explanations of domain assignments.The uncertainty template asks whether an assignment should be trusted for downstream analysis.
  • 6 Task C: Coupled Reasoning: Llama-3.3-70B generates justifications under a strict evidence-only prompt, but the design does not guarantee biological correctness.Faithfulness is evaluated using grounding rate, confidence alignment, and spatial consistency.
  • 6 Task C: Coupled Reasoning: Grounding rate measures the fraction of supplied named marker genes or proteins mentioned in each justification.Spots whose dominant evidence consists of unnamed image features are excluded from this metric.

7 OmicSync-R: Reasoning-Guided Training via Policy Gradient

OmicSync-R feeds automatically scored reasoning quality back into clustering through REINFORCE, while treating the language model and scoring procedure as black boxes. The reward acts as a secondary regularizer that favors assignments producing grounded, calibrated, and spatially consistent justifications.

  • 7 OmicSync-R: Reasoning-Guided Training via Policy Gradient: OmicSync-R uses reasoning-quality scores as REINFORCE rewards for the ClusteringHead without differentiating through generated text.The scalar reward weights the log-probability of sampled cluster assignments.
  • 7 OmicSync-R: Reasoning-Guided Training via Policy Gradient: For each sampled domain label, OmicSync-R builds a domain-conditioned evidence dictionary, generates a justification, and scores it automatically.The score combines grounding rate, confidence alignment, and spatial consistency.
  • 7 OmicSync-R: Reasoning-Guided Training via Policy Gradient: Grounding rate receives weight 0.5, confidence alignment 0.3, and spatial consistency 0.2 in the reasoning-quality reward.Grounding receives the highest weight because it is the most directly verifiable faithfulness criterion.
  • 7 OmicSync-R: Reasoning-Guided Training via Policy Gradient: A positive advantage increases the probability of assignments producing well-grounded justifications, whereas a negative advantage decreases probabilities for poorly grounded, calibrated, or spatially inconsistent explanations.The reward and baseline are treated as constants during backpropagation through stop-gradient.
  • 7 OmicSync-R: Reasoning-Guided Training via Policy Gradient: Reasoning updates begin after a 50-epoch warm-up and use 10 sampled spots every 25 epochs, producing 22 update steps and 220 LLM queries over 600 epochs.This schedule avoids using unstable early-training explanations as reward signals.

8 Experiments

OmicSync is evaluated across four CytAssist FFPE spatial proteomics datasets against representative multimodal clustering baselines. Experiments also specify shared clustering evaluation, implementation settings, and a resource-conscious Human Breast Cancer case study for OmicSync-R.

  • Datasets and baselines: The four datasets contain RNA, protein, spatial-coordinate, and H&E image information, with 3,460–4,194 spots per dataset.Histology embeddings are extracted with UNI.
  • Datasets and baselines: OmicSync is compared with GROVER, MISO, SpatialGlue, and COSMOS across four multimodal spatial proteomics datasets.The datasets are Human Tonsil, Human Breast Cancer, Human Glioblastoma, and Human Tonsil with Add-on Antibodies.
  • OmicSync-R setup: OmicSync-R uses repeated LLM explanation and reward evaluation, so it is assessed on Human Breast Cancer as a representative case study.The variant uses λR = 0.02, NR = 25, α = 0.9, and samples 10 spots per update.
  • Evaluation protocol: Task A uses nine clustering metrics and curated pseudo-reference labels because expert-annotated spot-level spatial domains are unavailable.The reported metrics include ARI, NMI, FMI, SilC, AMI, Jaccard, CHI, Purity, and DBI.
  • Evaluation protocol: For fair comparison, all methods use k-means on learned latent representations over K ∈ {6, 7, 8, 9, 10}, while soft assignments support reliability analysis and reasoning.The best result over the shared K range is reported for each method.

9 Results

OmicSync performs strongly on clustering and produces structured, reliability-aware explanations whose quality varies by strategy, dataset, confidence, and modality routing. OmicSync-R improves most Human Breast Cancer clustering metrics, while modality-routing patterns remain exploratory rather than causal.

  • Task A: Clustering performance: Best average ranks are 1.44 on Human Tonsil, 1.78 on Human Glioblastoma, and 1.22 on Tonsil Add-on; Human Breast Cancer ranks second-best at 2.33.OmicSync achieves the best ARI on all four datasets: 46.81, 45.73, 45.74, and 53.80, respectively.
  • Task A: Clustering performance: ARI reaches 53.80 on Tonsil Add-on versus 46.5 for GROVER, the strongest baseline, alongside improved intrinsic separation metrics.The authors relate this result to richer antibody information and uncertainty-aware multimodal fusion.
  • Task A: Clustering performance: Human Tonsil shows a Purity trade-off: 57.71 for OmicSync versus 69.4 for GROVER despite better ARI, SilC, CHI, and DBI.The authors interpret overall advantage using complementary external-agreement and intrinsic-structure metrics.
  • Task C: Reasoning quality: Stepwise reasoning achieves GR = 1.00, CA = 1.00, and SC = 0.975 across 40 explained spots, the strongest joint strategy performance.Standard and Contrastive also achieve full grounding, while Uncertainty has GR = 0.074.
  • Task C: Reasoning quality: Tonsil Add-on has the highest spatial consistency (0.84) and lowest average routing uncertainty (0.0526), whereas Human Breast Cancer has the lowest spatial consistency (0.58).Human Tonsil has the highest confidence alignment at 0.94, showing dataset-dependent reasoning quality.
  • Modality routing: RNA dominates routing for 7 of 9 high-confidence spots, while image and ADT are equally frequent dominant modalities among 8 low-confidence spots.At dataset level, RNA dominates 50% of explained Glioblastoma spots; routing weights are proxies, not causal modality contributions.
  • Reliability audit: High-confidence spots have mean uncertainty 0.034 versus 0.111 for low-confidence spots and 75.6% versus 60.0% neighbourhood homogeneity.Confidence and routing uncertainty are moderately negatively correlated (r = −0.42, p < 0.01) but measure different properties.

10 Discussion

OmicSync’s discussion emphasizes a coupled audit trail that links interpretable clustering signals to evidence-grounded explanations, while OmicSync-R uses reasoning quality as a non-differentiable training signal. The framework remains bounded by reward sensitivity, computational cost, evidence constraints, and limited current validation.

  • Coupled audit trail: The coupled audit trail exposes confidence, modality routing, marker evidence, uncertainty, and neighbourhood composition for evidence-grounded spot-level explanations.These signals connect clustering outputs to human-understandable reasoning rather than treating interpretability as a separate visualization step.
  • Reasoning as a training signal: OmicSync-R uses reasoning-quality scores with REINFORCE to align reasoning coherence and spatial domain quality during training.The method does not back-propagate through the language model and instead treats reasoning quality as an auxiliary reward.
  • Adaptive spatial smoothing: Adaptive nhops smoothing uses larger exclusion radii in homogeneous tissue and smaller radii in heterogeneous tissue to preserve spatial boundaries.This addresses the ARI/SilC trade-off and reduces reliance on one manually fixed smoothing configuration across datasets.
  • Limitations: Base OmicSync explains clustering after training, whereas OmicSync-R is sensitive to cluster count, resource intensive, and currently evaluated on one dataset.The reasoning module is also limited to five model-supplied evidence types and inherits uncertainty from those upstream signals.
  • Future work: Future work will extend OmicSync-R across datasets, improve neighbourhood-aware rewards and k-robustness, and support additional modalities and higher-resolution platforms.The proposed extensions include ATAC-seq, spatial metabolomics, Xenium, and MERFISH.

11 Conclusion

OmicSync couples spatial domain clustering with evidence-constrained LLM reasoning, while OmicSync-R feeds automatically scored reasoning back into clustering through REINFORCE. Across four benchmarks, OmicSync ranks best on three, and OmicSync-R improves Human Breast Cancer performance.

  • 11 Conclusion: OmicSync and OmicSync-R couple domain clustering with five-strategy LLM reasoning and REINFORCE-based reasoning-quality rewards.The framework uses shared per-spot interpretability signals to support auditable explanations and reasoning-guided training.
  • 11 Conclusion: OmicSync achieves the best average rank on three of four CytAssist FFPE benchmarks while jointly providing reliability signals and auditable justifications.The reported signals include assignment confidence, epistemic uncertainty, and modality attribution.
  • 11 Conclusion: OmicSync-R improves Human Breast Cancer ARI from 45.73 to 46.72 and surpasses the strongest non-OmicSync baseline on six of nine metrics.The result supports partial alignment between evidence-grounded reasoning coherence and clustering quality.
Loading 2608.22785v1…