acceptodds
Under review as a conference paper at ICLR 2027

Trust-VC: Multi-View Clustering for Transcriptomics via Trust-Aware Virtual-Prior Residual Fusion

Abstract

Multi-view clustering (MVC) of paired multi-omics data must reconcile complementary but unequally reliable views, and the problem becomes harder when external representations from pretrained foundation models are introduced. We present Trust-VC, a reliability-aware MVC framework that treats a foundation-model embedding as a virtual-cell prior rather than an equally trusted observed view. RNA is directly measured in all evaluated datasets. We use a fixed RNA-derived representation as the reference for fusion, and express contributions from the measured auxiliary modality and the virtual prior as gated residual corrections relative to that reference. For each cell, virtual-prior trust combines cross-modal consistency, a learned latent-dispersion proxy, and a self-supervised assignment confidence gain proxy. A reference-preservation objective penalizes excessive net displacement from the fixed RNA reference, while partial-to-complete distillation aligns the two residual-enhanced partial views with the complete view without updating the RNA reference. Extensive experiments on eight paired single-cell multi-omics datasets demonstrate that Trust-VC not only achieves the best performance in 28 of 32 dataset–metric comparisons, with up to a 10.8 percentage-point ACC improvement, but also regulates transferred corrections in fixed-checkpoint prior-permutation diagnostics.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Single-cell multi-omics, multimodal clustering, single-cell foundation models, trust-aware fusion, residual learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_jGynOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Trust-VC, an unsupervised multi-view clustering framework for paired single-cell multi-omics data. It integrates non-trainable scRNA reference representation with observed auxiliary modality and single-cell foundation model embeddings via gated residual adjustments that combines cross-modal consistency, latent dispersion and self-supervised clustering assignment-confidence.

Critical strengths
  1. With the accumulation of single-cell pretrained foundation-models, it is timely to design proper methods to quantify reliability weights of different pretrained foundation-model representations when incorporating them for paired multi-omics integration.
  2. The authors conducted extensive benchmark over eight datasets spanning diverse biological tissues, five single-cell foundation models, and ten baseline paired multi-omics integration algorithms.
  3. Fixed-checkpoint permutation experiments show the critical role of prior gates.
Critical weaknesses
  1. Both the RNA reference zirz_i^r and the FM representation xivx_i^v are deterministic transformations of the exact same RNA expression profiles. Why bifurcating the same biological signal into a static PCA baseline and an additive neural prior? In particular, since the authors utilize xivx_i^v from multiple foundation models, why not treat zirz_i^r as a new one? Moreover, the authors directly added non-linear MLP-generated residuals into PCA-reduced space, which seems to be unprincipled since Euclidean displacements in unconstrained neural latent spaces do not directly correspond to semantic shift vectors within a linear PCA space.
  2. I am quite confused about how the virtual encoder is framed as learning a diagonal Gaussian distribution to capture latent dispersion. Since the encoders of fixed pretrained foundation models are deterministic maps and no stochastic reparameterization or sampling is performed, sivs_i^v is merely a parallel feedforward output constrained only by a standard Gaussian KL regularizer and an ad-hoc dispersion score. Hence, it would be unsubstantiated to claim that this captures model uncertainty or reliability. I consider it essentially an uncalibrated, deterministic heuristic dressed in variational notation.
  3. The clustering are based on a predefined cluster number optimized via DEC-style t-distribution sharpening, which is not practical in real single-cell analysis. Why not evaluating with more widely-used graph-based methods like Leiden clustering? Also, the main manuscript contains no manifold visualizations (like UMAP or t-SNE), and I do not observe clear benefits of Trust-VC in the UMAPs in the appendix. In fact, the representation quality of Trust-VC appears poor in Fig. 17, where the clusters are excessively dispersed as small, isolated spheres. This is likely due to the DEC-style approach forcing an overfitted contraction toward cluster centers, thereby losing the continuity of the true single-cell biological manifold.
  4. Benchmark seems to be comparing against some very rare methods while ignoring standard tools, such as Seurat, scVI, totalVI, MultiVI, MOFA+, etc.
Questions for authors

Does Figure 2 report standalone foundation model clustering on transcriptomics alone, or Trust-VC using each model as a prior? If former, it conflates algorithmic design with the data advantage of including complementary measured modalities; if latter, what are the standalone clustering scores without auxiliary modalities or fusion?

How does Trust-VC perform under misspecified cluster numbers or Leiden clustering? Is the performance robust?

I am also confused on the simple addition of latent vectors and the validity of dispersion that need the authors to clarify (see weakness).

Reviewer_h5ELOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The authors present Trust-VC, a multi-view clustering approach that combines a fixed RNA representation with corrections from an auxiliary modality and a foundation-model embedding. While the problem is relevant, I have major concerns regarding the central claim and evaluation.

Critical strengths

see above

Critical weaknesses

My main concern is that the authors do not demonstrate that the improvement comes from useful biological information in the foundation-model prior. In their own permutation diagnostic on GSE128639, scrambling the cell–prior correspondence improves ARI from 0.523 to 0.569, while mean trust only decreases from 0.933 to 0.767. Yet removing the prior reduces ACC by 15.4 percentage points in the main ablation. These experiments use different training configurations, which further complicates their interpretation. Taken together, they leave the source of the improvement unresolved. Comparisons with shuffled priors, random features and conventional RNA representations, trained under the same protocol, would be necessary to distinguish transferred knowledge from an architectural or regularisation effect. Moreover, the “usefulness” target measures increased assignment confidence, which can also increase for incorrect assignments.

Second, validation labels are used to select the foundation model separately for each dataset. This is a substantive supervised model-selection step in an otherwise unsupervised clustering evaluation. It is not clear how this selection would be performed in practice without annotations, or whether the baselines receive comparable tuning. Results with a fixed prior source or a label-free selection procedure are needed.

I also miss comparisons to standard single-cell multimodal approaches, particularly WNN, totalVI/multiVI and MOFA+, followed by appropriate downstream clustering. Given that the work is explicitly motivated by single-cell data integration, these are essential comparisons for assessing its practical usefulness.

There are further concerns regarding the reporting of the experiments. On PBMC5K, both ACC and ARI are reported with ±0.0 variation across repeated runs, despite considerably larger variation in the ablations. The number of independent runs, what was randomised and the unrounded results should be provided. Likewise, the sensitivity analysis should report the actual evaluated parameter grid and scores, and clarify whether interpolation was used to construct the 3D surfaces.

Finally, the manuscript is unnecessarily dense and jargon-heavy. The the conceptual advance and evidence supporting the proposed trust mechanism remain unclear.

Questions for authors

see above

Reviewer_RPLUOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a multi-view clustering method for multi-omics data. It establishes a fixed, RNA-derived representation as a baseline for subsequent data fusion. A virtual-prior trust metric is then derived from three components: cross-modal consistency, a learned latent-dispersion proxy, and a self-supervised assignment-confidence gain proxy. The authors claim that this approach achieves superior performance across most datasets.

Critical strengths
  1. Using a fixed, RNA-derived representation as a baseline is a reasonable design choice; anchoring other information to this baseline helps ensure the reliability and stability of the algorithm.
  2. The overall model architecture is clear, and Figure 1 aligns well with the textual description, which enhances the paper's readability.
Critical weaknesses
  1. Why is Equation (34) enclosed in a box? Is this a remnant from processing by a large language model (such as GPT)? It looks quite odd.
  2. The datasets and experimental setup are not sufficiently clear. The authors appear to have relegated a significant amount of information which should have been included in the main text to the appendix. It is important to note that the main body of a paper must remain relatively self-contained and complete.
Questions for authors
  1. What is the relationship between the three components, cross-modal consistency, the learned latent-dispersion proxy, and the self-supervised assignment-confidence gain proxy? Is the information they capture mutually exclusive, or is there overlap?
  2. Could the embeddings from the base model introduce extraneous information, thereby compromising the fairness of the comparisons? The authors should clarify whether the method's core contributions truly align with the observed performance gains.
  3. I am curious about the robustness of this method. Given the use of prior knowledge and the three metrics for determining virtual-prior trust, does the method demonstrate a significant improvement in robustness compared to similar approaches? Could the authors provide some theoretical analysis on this?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.