ARO: Aligned Representation learning for multi-Omics data

Amogh Singh; Yash Shah; Chiara D'Ercoli; Arash Mehrjou; Patrick Schwab; Timothy Jones; Pietro Liò

cs.LGOct 5, 2026

ARO: Aligned Representation learning for multi-Omics data

Authors: Amogh Singh, Yash Shah, Chiara D'Ercoli, Arash Mehrjou, Patrick Schwab, Timothy Jones, Pietro Liò

Organizations: Department of Computer Science, University of Cambridge, Cambridge, UK · DIAG, University of Rome “Sapienza”, Rome, Italy · Max Planck Institute for Intelligent Systems

Abstract

The high cost of functional molecular assays, and prevalence of missing modalities and unmatched samples in computational biology, create significant barriers to comprehensive multi-omic profiling, essential for capturing and reasoning over molecules, cells, tissues, and organisms. This work proposes a model that learns meaningful representations from multi-omics cancer data supporting the reconstruction of missing and unpaired modalities. Contrary to increasingly complex, larger models, e.g. Foundation Models (FMs), ARO prioritizes practical applicability in limited or incomplete data settings. ARO optimally reconstructs missing modalities (MSE of $0.15$ on the validation and test data in the Unmasked settings), with its learned latent embeddings enabling a downstream cancer classification task. Our findings indicate that analyzing diverse molecular layers as a single integrated system offers a reliable and cost-efficient approach, reducing dependence on large-scale experimental testing, while still supporting multi-omic exploration in limited data settings.

Figures & tables

Figure 1 : Overview of ARO. Multi-omics cancer datasets are first harmonized and aggregated. Next, pre-processing involving median imputation, Winsorization, normalization, and log transformation is performed. Features are then optionally masked (shown by dotted outlines) and passed to the model. The loss is computed, and learned latent representations can be used for downstream tasks.

Panel of Figure 2 — Figure 2 : ARO performance in discriminating each omics-modality within the latent space, visualized using PCA (Figure 2(a) ) as well as across downstream tasks including binary cancer classification (Figure 2(b) ) and cancer laterality prediction (Figure 2(c) ).

	Train	Validation	Test
ARO Masked	$0.046$	$0.14$	$0.14$
ARO Unmasked	$0.064$	$0.15$	$0.15$
PCA	$3.1e-30$	0.14	0.14

Table 1 : Reconstruction performance using MSE Loss averaged over 5 runs

	Train	Validation	Test
ARO Masked	$0.022$	$0.070$	$0.068$
ARO Unmasked	$0.032$	$0.071$	$0.07$
PCA	$1.5e-30$	$0.068$	$0.065$

Table 2 : Reconstruction performance using Huber Loss averaged over 5 runs

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

Layer	Shape	Parameters
inp_lin.weight	(2048, 71094)	145,600,512
inp_lin.bias	(2048, )	2048
encoder_layers.0.weight	(2048, 2048)	4,194,304
encoder_layers.0.bias	(2048, )	2048
encoder_layers.1.weight	(2048, 2048)	4,194,304
encoder_layers.1.bias	(2048, )	2048

Appendix

Table 3 : Parameter numbers of each block of the autoencoder.

Panel of Figure 3 — Figure 3 : Iterations for hyperparameter tuning performed over validation loss, optimizing for dropout configuration, hidden dimensions, learning rate, modality-specific masking, number of hidden layers, batch normalization usage and weight decay. The configuration with the lowest validation loss is chosen from each setup.

	Train	Validation	Test
ARO Masked	$0.046\pm 56.0e-04$	$0.14\pm 31.0e-05$	$0.14\pm 21.0e-05$
ARO Unmasked	$0.064\pm 31.0e-05$	$0.15\pm 21.0e-05$	$0.15\pm 24.0e-05$
PCA	$3.1e-30$	0.14	0.14

Appendix

Table 4 : Reconstruction performance using MSE Loss averaged over 5 runs

	Train	Validation	Test
ARO Masked	$0.022\pm 28.0e-04$	$0.070\pm 5.6e-05$	$0.068\pm 20.0e-05$
ARO Unmasked	$0.032\pm 16.0e-05$	$0.071\pm 10.0e-05$	$0.07\pm 6.1e-05$
PCA	$1.5e-30$	0.068	0.065

Appendix

Table 5 : Reconstruction performance using Huber Loss averaged over 5 runs

Figure 4 : t-SNE cluster of the learned embeddings from ARO

Panel of Figure 5 — Figure 5 : Confusion matrices of downstream tasks using PCA projections.

Data set	Samples	Feature Count
( Li et al., 2024 ) - Transcriptomics	108	45886
( Li et al., 2024 ) - Proteomics	109	10998
( Li et al., 2024 ) - Metabolomics	87	67
( Wang et al., 2021 ) - Transcriptomics	240	26941
( Wang et al., 2021 ) - Proteomics	333	12309
( Wang et al., 2021 ) - Metabolomics	36	216

Appendix

Table 6 : Details of the Multi-omics datasets considered and analysed in the study.

Explore similar work

Probing, Fusion, and Trustworthiness: A Systematic Evaluation of Foundation Model Representations for Multimodal Cancer Analysis
Jun 15, 2026Jingyu Hu, Giuseppe Tripodi, Reed Naidoo +2Pathology Foundation Models Multimodal Clinical Data
The Alan Turing Institute, London, United Kingdom · University of Manchester, Manchester, United Kingdom · The Institute of Cancer Research, London, United Kingdom +1
Latent World Recovery for Multimodal Learning with Missing Modalities
Jun 10, 2026Hui Wang, Tianyu Ren, Joseph Butler +3Missing Modalities Multimodal Learning
Queen’s University Belfast, Belfast, United Kingdom
DoGMA: A Central-Dogma-Guided Foundation Model for Multi-Omics Alignment and Multi-Task Learning in Oncology
Aug 8, 2026Junfei Ling, Bangzheng Pu, Bingsen Xue +3Multi-Omic Integration Multidisciplinary Tumor Boards
Institute of Medical Robotics, Shanghai Jiao Tong University · Institute of Medical Robotics, Shanghai Jiao Tong University, Shanghai 200240, China · Institute of Data Science, The University of Hong Kong +1

Jun 15, 2026, cs.LGJ/K move · Enter open · S save

Probing, Fusion, and Trustworthiness: A Systematic Evaluation of Foundation Model Representations for Multimodal Cancer Analysis

Jingyu Hu, Giuseppe Tripodi, Reed Naidoo, Sarah F. McGough+1

The Alan Turing Institute, London, United Kingdom · University of Manchester, Manchester, United Kingdom · The Institute of Cancer Research, London, United Kingdom+1

Pathology Foundation Models Multimodal Clinical Data Multi-Omic Integration Foundation Model

Open paper AskPDF arXiv

Foundation models (FMs) have emerged as powerful representation extractors for medical data, yet their generalizability to datasets under distribution shift remains underexplored. This work systematically evaluates FM-based representations on a suite of computational pathology tasks across two real-world commercial cohorts, IH-BC and IH-NSCLC, drawn from the licensed in-house (IH) oncology dataset. The analysis focuses on two modalities, whole-slide images and transcriptomic profiles, drawn from the IH multimodal data. We first benchmark unimodal probing performance across five FMs on eight downstream classification tasks, and find that image and omics representations carry complementary predictive signals. Then we investigate whether multimodal fusion can yield additional gains over unimodal baselines by comparing three image-omics fusion strategies built on paired representations. The trustworthiness of selected unimodal and multimodal pipelines is further assessed through conformal prediction. Our results show that FM representations achieve competitive performance on out-of-distribution data and that multimodal fusion helps mainly when no single modality dominates the signal. Conformal prediction reveals that in the majority of cases where a point prediction fails, the true diagnosis remains recoverable within the prediction set, reinforcing the value of uncertainty-aware inference for clinical support.

Abstract

Explore similar work

Probing, Fusion, and Trustworthiness: A Systematic Evaluation of Foundation Model Representations for Multimodal Cancer Analysis

Latent World Recovery for Multimodal Learning with Missing Modalities

DoGMA: A Central-Dogma-Guided Foundation Model for Multi-Omics Alignment and Multi-Task Learning in Oncology