Can aligned representation learning reconstruct missing multi-omics modalities?

ARO rebuilds proteome and metabolome from transcriptome alone

The paper's central claim is that a single observed omics layer is enough to recover the others. ARO reconstructs missing and unpaired modalities from one input, specifically enabling reconstruction of proteomic and metabolomic data from transcriptomic measurements. This matters because functional molecular assays are expensive, and missing modalities plus unmatched samples are common in computational biology, which blocks comprehensive multi-omic profiling.

The design choice that enables this is a single autoencoder shared across modalities, rather than one model per modality. ARO does not require pretraining on large external datasets and does not incorporate biological priors, pathway information, or auxiliary imaging data. Instead it relies on careful preprocessing, a unified latent representation, and masking to reconstruct masked data.

Two paired cancer cohorts feed one 115 by 74,002 matrix

The evidence comes from two publicly available multi-omics cancer datasets. The glioblastoma cohort includes gene expression (FPKM), proteome, phospho-proteome, and metabolome. The renal cell cancer cohort includes RNA-seq expression (TPM), protein abundance, and metabolite compound abundance. In both cohorts, multi-omics profiles are obtained for tumor and normal paired samples.

Each modality is represented as a feature matrix, and the matrices are concatenated along the feature dimension into a unified dataset of size 115 × 74,002, where each feature is a distinct molecular entity spanning transcriptomic, proteomic, and metabolomic measurements and each row is a patient. Feature identifiers are harmonized to ENSG gene IDs via g:Profiler and to KEGG metabolite IDs via MetaboAnalyst with stripped ENSEMBL version suffixes. Zero-variance features are removed, and train-validation-test splits are performed at the patient level at a 70/15/15 ratio.

The authors state that normalization was performed independently for each feature, standardizing each gene to its own distribution within its dataset, rather than converting both cohorts to a common expression metric. They describe the objective as methodological rather than bioinformatic, and say care was taken to prevent leakage of validation or test set distribution information into training.

MSE and Huber loss rank ARO differently against PCA

Reconstruction is reported with two losses averaged over five runs, and the two metrics do not agree. Under MSE, ARO Unmasked reaches 0.064 on train, 0.15 on validation, and 0.15 on test; ARO Masked reaches 0.046, 0.14, and 0.14. A PCA baseline reaches 3.1e-30 on train but 0.14 on validation and 0.14 on test.

Under Huber loss, ARO Masked reaches 0.022 on train, 0.070 on validation, and 0.068 on test; ARO Unmasked reaches 0.032, 0.071, and 0.07. The PCA baseline reaches 1.5e-30 on train, 0.068 on validation, and 0.065 on test.

The near-zero training values for PCA are not evidence that PCA reconstructs better in any useful sense: PCA is fit directly on the data and its validation and test errors land at the same level as ARO's. The masked and unmasked variants also differ only marginally on validation and test, so the headline MSE of 0.15 should be read as the Unmasked setting specifically, not as a single number covering both configurations.

A 303.9M-parameter autoencoder stays trainable where MODIS does not

ARO is a fully connected feedforward autoencoder. The encoder applies input dropout at p = 0.1, then a linear projection from 71,094 to 2,048 dimensions, followed by three fully connected hidden layers of dimension 2,048, each with ReLU activation and dropout at p = 0.0, producing a 2,048-dimensional hidden representation. The decoder projects back to the original feature space with a final linear layer and a softplus activation to enforce non-negative outputs after preprocessing and log-space normalization. The evaluated model has about 303.9M trainable parameters.

Training uses AdamW at a learning rate of 10^-5 with no weight decay, a ReduceLROnPlateau scheduler with factor 0.1 and patience of 10 epochs, 200 epochs, batch size 80, mixed bfloat16 precision with fp32 fallback, and torch.compile on CUDA devices. No batch normalization is used. The optimized objective is Huber loss, and hyperparameters were tuned by Bayesian optimization through Weights & Biases over dropout configuration, hidden dimensions, learning rate, modality-specific masking, number of hidden layers, batch normalization usage, and weight decay.

The comparison the authors wanted was against MODIS, which they identify as the closest approach to ARO. That comparison could not be completed: at the dimensionality of this data, MODIS scales to roughly 22B parameters, producing out-of-memory errors and making full training infeasible. The authors argue that shrinking MODIS would require discarding a substantial portion of the omics set and would change the problem setting. Their limited MODIS training reached MSE of 2.53 for transcriptomics, 0.55 for proteomics, and 0.43 for metabolomics. ARO, by contrast, remains tractable across roughly 50M to 700M parameters, with model size showing limited impact on reconstruction performance.

Latent embeddings separate cancerous from non-cancerous samples

Embeddings extracted from the trained model's latent space were used to generate PCA and t-SNE plots. These show distinct clusters of cancerous and non-cancerous samples, and of cancer laterality. The authors read the separation as evidence that the learned embeddings capture the underlying structure of the original omics data while retaining enough discriminatory information to distinguish clinical states.

For downstream evaluation, embeddings from the hidden layer of the ARO Unmasked model were paired with an MLP for binary classification (cancer versus non-cancer) and multi-label classification (cancer laterality). ARO achieved 100% accuracy on the binary task on test data. On laterality it reached 72.4% accuracy, which the authors describe as comparatively weaker and consistent with biological expectations, since laterality requires subtler spatial or micro-environmental signals that omics profiles alone cannot supply.

The PCA projections used for the same downstream tasks give 72.41% on binary classification and 58.62% on multi-label classification. The gap between 100% and 72.41% on the binary task is the clearest downstream separation in the paper, and it is a comparison between ARO's learned embeddings and raw PCA projections, not between ARO and MODIS.

115 samples limit generalization beyond these two cohorts

The authors state plainly that although the data are highly feature-dimensional, the number of training samples remains limited, which may constrain generalization to unseen data distributions and highlights the need for validation on external cohorts, possibly including other diseases. The 100% binary accuracy and 72.4% laterality accuracy are therefore results on held-out splits of these two cohorts, not estimates of performance on new patient populations.

The authors also state that the learned associations remain statistical regularities and carry no claims of biological plausibility, and that further analysis would be needed to substantiate potential biological meaning. The reconstruction of proteomic and metabolomic layers should be read as a statistical mapping learned from these aligned matrices, not as a demonstrated biological mechanism.

Two further scope conditions are explicit. ARO's low-capacity architecture is intended for limited-sample settings, and the work focuses on robust learning from aligned multi-omics matrices rather than heterogeneous multi-modal settings such as imaging or spatial transcriptomics. The authors also note that an extension could integrate generative settings such as variational autoencoders to enable full synthesis of proteomic and metabolomic profiles conditioned on transcriptomic inputs, which places full generative synthesis outside what was demonstrated here.

Reconstruction is worth it when transcriptomics is the only layer you have

The practical priority this paper establishes is a scenario split, not a general claim that ARO beats larger models. When a cohort has transcriptomic measurements but missing or unpaired proteomic and metabolomic layers, ARO offers a way to fill those layers from a single observed input without pretraining on large external datasets and without pathway priors. When the goal is architectural novelty or heterogeneous multi-modal fusion across imaging and sequence data, this paper explicitly does not target that setting.

The comparison that should guide a reader's judgment is the one the paper actually completed: ARO's embeddings against PCA projections on the same downstream tasks, where binary classification moves from 72.41% to 100% and multi-label laterality from 58.62% to 72.4%. The comparison the paper could not complete is against MODIS, which ran out of memory at roughly 22B parameters on this data. The most defensible reading is that ARO makes masked multi-omics reconstruction tractable and useful at 115 samples and roughly 303.9M parameters, and that its value is concentrated in limited or incomplete data settings rather than in scaling to larger cohorts or richer modality types.

Sources used in this answer