DoGMA

Self-attention treats every pair of inputs symmetrically: in a multi-omics fusion model, a methylation feature attends to a protein abundance exactly as freely as the reverse. For bulk tumor profiles that permissiveness is a liability, because the molecular layers are ordered — DNA-level state constrains transcription, transcription constrains protein abundance, and feedback the other way is comparatively weak. A model free to route information backwards can fit a correlation that holds on one cohort's assay platform and lose it on the next.

DoGMA writes the central dogma into the attention computation itself. Each attention head adds a structural bias to its logits before the softmax: edges running against central-dogma flow carry a fixed negative penalty shared across all weak-feedback directions, and every omics pair additionally gets a zero-initialized learnable residual, so the prior sets a coarse direction while the data refines individual routes. It was developed at the Institute of Medical Robotics, Shanghai Jiao Tong University, with collaborators at The University of Hong Kong and the University of Shanghai for Science and Technology, and posted to arXiv in August 2026.

It sits apart from sequence-level central-dogma work such as Life-Code, which folds DNA, RNA and protein into one nucleotide alphabet. DoGMA instead operates on bulk per-patient measurements — methylation, mutation calls, copy-number alterations, expression and protein abundance — constraining which molecular layer may inform which.

#Key Features

• Directed cross-omics attention: A binary topology matrix marks weak-feedback edges and penalizes them with a shared negative bias; a learnable, head-specific residual per omics pair recovers pair-specific regulatory structure the coarse prior misses.
• Modality-specific expert routing: After attention, each omics representation passes through a mixture-of-experts layer whose gate weights expert outputs per patient, letting the contribution of each molecular layer vary with the assays actually available.
• Learnable missing-modality tokens: Absent assays are represented by learned embeddings rather than zeros, so patients with partial panels are retained instead of dropped to complete cases.
• Masked hierarchical reconstruction: Pretraining combines masked self-reconstruction within each modality, cross-omics reconstruction of one layer from the fused representation, and adversarial domain alignment with contrastive learning.
• Counterfactual edge intervention: With parameters frozen, individual omics edges can be masked at inference to test which routes the trained model actually relies on.

#Technical Details

Training proceeds in three stages. Modality-specific VQ-VAEs first quantize each high-dimensional profile into a 256-dimensional latent against a learned codebook. Six Transformer-MoE blocks with eight attention heads then pretrain on 8,194 TCGA specimens spanning 32 cancer types, split 4,916/1,639/1,639, using AdamW at 5 × 10⁻⁵, batch size 128, a 0.15 masking ratio and 200 epochs; clinical text is encoded by a frozen PubMedBERT. Downstream adaptation keeps lower blocks frozen.

Trained on TCGA-BRCA and tested on the independent METABRIC cohort (1,689 patients), DoGMA reaches 0.955 average test accuracy across four subtype endpoints, against 0.890 for the identical architecture trained from scratch, 0.943 for DeePathNet and 0.918 for pretrained TMO-Net; its pretrained embedding's cancer-type silhouette score is 0.628 versus TMO-Net's 0.368 before any fine-tuning. On the MetaCancer distant-metastasis cohort (399 specimens from 318 cases, split at case level) it reports 0.9747 accuracy and 0.9750 F1, and on a 997-patient institutional colorectal cohort the best C-index among compared methods. Cross-omics reconstruction of expression from genomic inputs reaches Pearson r = 0.78 versus 0.64, and driver-mutation inference AUC 0.75 versus 0.57. A Dogma Directionality Index quantifying forward-versus-reverse attention flow rises to 0.739 for the full model against 0.265 for standard attention, and knocking out forward edges costs 0.1242 C-index — more than any random-matched deletion.

#Applications

The pretrained backbone is aimed at clinical cohorts too small to train a fusion model from scratch: molecular subtyping, distant-metastasis classification and prognostic risk stratification all start from one checkpoint and a task head. Because missing assays are tolerated by construction, it suits hospital cohorts where proteomics or methylation were run on only a fraction of patients. The cross-omics reconstruction head additionally imputes transcriptomic state from genomic inputs and infers driver-mutation status from downstream profiles, useful where an assay was never ordered.

#Impact

DoGMA is an argument that domain structure belongs in the architecture rather than only in the loss: the ablations show the directional prior and the reconstruction objectives are each insufficient alone, and that the prior becomes operational only once pretraining teaches the model to use it — a framing portable to other ordered biological systems. The work remains a preprint awaiting peer review, and no code or model weights have been released, so the reported numbers cannot yet be reproduced independently; the institutional colorectal cohort is likewise unreleased pending data-release procedures. Evaluation is confined to bulk human tumor cohorts, and the causal chains recovered by in-silico intervention are presented as hypotheses rather than validated mechanisms.