CMGL: Confidence-guided Multi-omics Graph Learning for Cancer Subtype Classification

The gist: CMGL proposes a two-stage framework that estimates per-sample modality reliability through evidential deep learning and uses these frozen confidence scores to guide cross-omics fusion and graph construction for cancer subtype classification.

Motivation

Multi-omics integration can improve cancer subtyping, but modality informativeness and noise vary across cancer types and patients Existing graph-based methods optimize modality weights jointly with the classification objective and therefore lack independent reliability estimates low-quality omics distort patient similarity graphs and amplify noise through message passing Biologically, this means that the molecular signatures of true cancer subtypes, such as the hormone-receptor versus proliferation axis in BRCA, can be diluted by noise from uninformative modalities. The core question is therefore how to identify which omics modalities are trustworthy for each patient before fusion and graph reasoning.

CMGL Framework

CMGL adopts a two-stage design where Stage 1 estimates modality reliability and Stage 2 performs cross-omics feature fusion and graph-based classification under frozen confidence priors The key design principle is that reliability estimation precedes fusion and graph reasoning, and is not updated by downstream gradients.

Stage 1 involves Evidence Modeling where an encoder maps raw input into a hidden space, and an evidence head outputs non-negative class-wise evidence e(m) Following the EDL framework [Sensoy et al., 2018], Dirichlet parameters are constructed to yield the Dirichlet predictive mean π(m)i = E[p(m)i] and the epistemic uncertainty u(m)i = C/S(m). A Quality Estimator extracts three complementary signals: log evidence strength a(m)i, normalized predictive entropy H˜(m)i, and maximum predictive probability π(m)i,max. These are combined by a per-modality scoring network Qm(·) to yield the modality confidence r(m)i. The Stage 1 objective is the weighted sum of three terms: an EDL loss Ledl with an annealed KL regularizer, a confidence-weighted classification loss L(r)cls, and a diversity constraint Ldiv. After training, the confidence vector ri is frozen and used by Stage 2.

Confidence-Guided Fusion and Graph Reasoning

Stage 2 performs feature fusion under frozen confidence priors where independent encoders map each omics modality into a shared representation space, augmented with an omics-identity embedding pm Multi-head self-attention across the M = 4 modality tokens exchanges information feature-wise. The fused representation is calculated as zi = X sum m=1 r(m)i g(m)i⊙ v˜(m) Low-confidence modalities are suppressed at both levels simultaneously, preventing noisy omics from diluting subtypediscriminative features.

For graph construction, CMGL constructs a k-nearest-neighbor (k-NN) graph E(m) independently for each omics modality and retains only edges supported by all modalities: E∩ = M sum m=1 E(m). This ensures that noise from any single low-quality modality cannot reach message passing. The graph message passing uses a two-layer GraphSAGE network [Hamilton et al., 2017] on the consistency intersection graph, with a residual link adding the fused representation zi directly to the output layer.

Results and Interpretability

CMGL consistently improves over baselines on four MLOmics cancer-subtype tasks and the 32-class pan-cancer task, surpassing it by 4.03% in average accuracy on the four single-cancer tasks. On BRCA, the five predicted subtypes correspond closely to the established PAM50 intrinsic subtypes in both molecular signatures and sample proportions. The model also identifies a metabolically active subtype enriched in the tryptophan–kynurenine catabolic pathway. In cross-cancer transfer experiments, the BRCA-trained model transfers without fine-tuning to kidney renal clear cell carcinoma (KIRC), stratifying patients into three prognostically distinct groups whose pathway profiles recapitulate the established ccA/ccB molecular subtype framework. The analysis shows that the proliferation–differentiation axis learned on BRCA transfers as a conserved prognostic axis rather than a dataset artifact.

Conclusion

CMGL estimates modality reliability separately from classification, addressing distinct failure modes by filtering noisy modalities and combining feature-wise signals across modalities. The framework attains the highest accuracy and Macro-F1 on four MLOmics single-cancer tasks and the pan-cancer task. The BRCA embedding recovers the PAM50 subtypes and isolates a metabolic subtype, and the BRCA-trained model transfers to KIRC without fine-tuning, stratifying patients into three clusters matching the ccA/ccB framework. Missing-modality handling and adaptive graph construction are natural extensions of this work.

Data Availability

The datasets used in this study are from the MLOmics benchmark, available at https://github.com/chenzRG/CancerMulti-Omics-Benchmark. Code will be made publicly available upon acceptance.

References

A. R. Brannon, A. Reddy, M. Seiler, A. Arreola, D. T. Moore, R. S. Pruthi, E. M. Wallen, M. E. Nielsen, Huliu Liu and K L Nathanson Molecular stratification of clear cell renal cell carcinoma by consensus clustering reveals distinct subtypes and survival patterns Genes & Cancer 1:152–163 2010.

A. Cheerla and O. Gevaert Deep learning with multimodal representation for pancancer prognosis prediction Bioinformatics 35:i446–i454 2019.

E. Y. Chen, C. M. Tan, Y. Kou, Q. Duan, Z. Wang, G V Meirelles N R Clark and A Ma’ayan Enrichr interactive and collaborative HTML5 gene list enrichment analysis tool BMC Bioinformatics 14:128 2013.

T. Chen, S. Kornblith, M. Norouzi, and G Hinton A simple framework for contrastive learning of visual representations In International Conference on Machine Learning ICML pages 1597–1607 2020.

W. Du, L. Zhang, A Brett-Morris et al HIF drives lipid deposition and cancer in ccRCC via repression of fatty acid metabolism Nature Communications 8:1769 2017.

Y. Gal and Z Ghahramani Dropout as a Bayesian approximation Representing model uncertainty in deep learning In International Conference on Machine Learning ICML pages 1050–1059 2016.

C. Guo, G Pleiss, Y Sun, and K Q Weinberger On calibration of modern neural networks In International Conference on Machine Learning ICML pages 1321–1330 2017.

W. L Hamilton R Ying and J Leskovec Inductive representation learning on large graphs In Advances in Neural Information Processing Systems NeurIPS pages 1024–1034 2017.

Z Han, C Zhang, H Fu, and J T Zhou Trusted multi-view classification with dynamic evidential fusion IEEE Transactions on Pattern Analysis and Machine Intelligence 45:2551–2566 2023.

Y Hasin M Seldin and A Lusis Multi-omics approaches to disease Genome Biology 18:83 2017.

1. textbfConfidence-Guided Modality Reliability Estimation in Stage 1: Implementing EDL and Quality Estimator for Robustness Analysis (Section 2.3, Algorithm S1). Developing this capability allows AI systems to quantify modality reliability independently, addressing the issue where low-quality omics distort patient similarity graphs and amplify noise through message passing. This results in a system that can suppress noisy inputs before fusion by using the output of the quality estimator to derive per-modality confidence scores, which are then frozen for Stage 2.

2. textbfCross-Omics Feature Fusion via Frozen Confidence Gating (Section 2.4.2). The improved system will utilize a mechanism where Low-confidence modalities are suppressed at both levels simultaneously, as described in the text, by using the frozen confidence scores to gate feature aggregation, thereby ensuring that only reliable modalities shape subsequent information flow during fusion and preventing errors from propagating through noisy omics.

3. textbfConsistency Intersection Graph Construction (Section 2.5). The AI system will employ a method where edges are retained only if every modality independently treats the two patients as neighbors, denoted as E∩ = M m=1 E(m). This ensures that noise from any single low-quality modality cannot reach message passing, leading to a more biologically sound graph structure for patient reasoning.

4. textbfAdaptive Graph Structure Selection via Warm-up Procedure (Section 2.5.2). Instead of fixed parameters, the system will implement an adaptive neighbor count selection by running a short Stage-2 warm-up in which training updates use the train-only graph and validation Macro-F1 is scored on the train–validation graph. This allows the system to select k∗ = arg maxk F1val(k) dynamically, optimizing graph density for each fold.

5. textbfRobustness via Hyperparameter Sensitivity Analysis (Section 3.6). The AI framework will be designed to be robust by confirming that chosen operating points lie within a broad robustness region by demonstrating that Macro-F1 varies by less than ±0.025 across tested loss weights, thus providing a stable configuration for training.

Motivation: Multi-omics integration can improve cancer subtyping, but modality informativeness and noise vary across cancer types and patients. Most graph methods for multi-omics data learn modality contributions within the downstream classification objective, leaving predictive reliability for each patient implicit. As a result, uninformative modalities can weaken the fused representation, while unreliable omics can introduce noisy patient relationships into graph propagation. To address these two problems, we propose CMGL, which produces a separate reliability estimate before fusion and uses consensus patient neighborhoods for graph classification. Results: CMGL estimates modality confidence for each patient through evidential deep learning, fixes these values during fusion across omics, and performs classification on an independently specified consistency graph. On four MLOmics cancer-subtype tasks and the 32-class pan-cancer task, CMGL consistently improves over the strongest baseline, surpassing it by 4.03% in average accuracy on the four single-cancer tasks. Its representations recover the PAM50 intrinsic subtypes of breast invasive carcinoma (BRCA), and the model trained on BRCA transfers without fine tuning to kidney renal clear cell carcinoma (KIRC), stratifying patients into prognostically distinct groups.