Doctors increasingly want survival models that read a tumour's histology slide, its gene expression and the patient's clinical record together. The models that do this well are trained on hundreds of matched patients per cancer type. Many real cohorts, such as rare tumours, single-site trials and new indications, have fifty patients or fewer. The standard remedy is to borrow patients from larger, related cohorts. That only helps if you know which cohorts are related, and with fifty patients you cannot reliably estimate it from the data. Biology already holds a large, independent record of how diseases relate: millions of curated gene–disease, pathway–disease, symptom–disease and microRNA–disease links. This note proposes using those curated networks for two jobs. The first is to decide which source cohorts a tiny target cohort should borrow from, and how much. The second is to decide which molecular features the target model should trust before it has seen much data.
Stated plainly, the idea is a diffusion over the network. We spread probability from the target disease outward across a heterogeneous biological network, meaning a graph with several kinds of nodes (genes, biological processes, symptoms, microRNAs, diseases) and several kinds of edges. The resulting profile says which other diseases and which genes sit close to the target. We use the disease scores to weight source patients during multimodal pretraining and the gene scores to set per-gene regularization in the target model. Neither quantity needs a single target patient, so both remain usable where data-driven transfer weighting breaks down.
Core Claim
The claim is that knowledge-derived weights beat data-derived ones only below a certain cohort size. Below roughly 100 target patients, cohort weights and feature priors taken from an external disease network should give better-calibrated and more discriminative multimodal survival predictions than two alternatives. One is uniform pan-cancer pooling. The other is transfer weighted by empirical similarity between source and target patients. We expect the advantage to shrink and vanish as the target cohort grows, because data-estimated relatedness then becomes reliable. The contribution is this crossover, together with a direct test of whether network proximity predicts measured transfer benefit. We do not propose a new fusion architecture.
Prior work leaves this open from three sides. Boosting Transfer Learning with Survival Data from Heterogeneous Domains (TSB) re-weights source patients by their likelihood under the target feature distribution. It raised C-index from 0.500 for a target-only Cox model to 0.677 on a MAGGIC cohort with fewer than 200 patients. However, it is restricted to tabular features, its weights come from the same scarce target data, and the file flags negative transfer when source and target differ in underlying mechanism. We replace the data-estimated weights with weights that exist before any target data arrive. PONET shows that biological structure helps multimodal survival models: it hard-wires Reactome pathways into a sparse genomic network, reaching a C-index of 0.829 on TCGA-KIRP with 2.8M parameters. But it uses one fixed pathway graph for every cancer, trains on full cohorts, and does no transfer. Our prior is target-specific and soft. GraphCare is the clearest evidence that external graphs can substitute for patients: with knowledge-graph augmentation, 36 MIMIC-III patients matched baselines trained on 720. GraphCare, though, works only on discrete EHR codes and uses the graph as model input. We use the graph to decide what to transfer into a model over slides and omics.
Research Program
The network follows MultiSourcDSim, which fused gene–disease, GO biological process–disease and symptom–disease networks from CTD MEDIC over 8,126 diseases. Its disease-similarity AUC was 0.905, against 0.771–0.797 for any single source. That result is our evidence that multi-source integration matters, and we reuse its method: random walk with restart (a walk that keeps jumping back to its start node, so visit frequencies measure closeness), followed by diffusion component analysis to compress the visit profiles into low-dimensional disease vectors. We add microRNA–disease edges from HMDD, because TCGA provides microRNA expression (Fan et al. used it for pan-cancer survival). For target diseases with few recorded microRNA links, we fill gaps with KATZ path-count scores in the style of KATZNCP, which reported AUC 0.9325 under leave-one-out on HMDD v2.0 and was designed for diseases with sparse links.
Two quantities come out of this network for a target cancer $t$. The first is a cohort weight. Let $z_s$ and $z_t$ be the diffusion vectors of source cancer $s$ and target $t$, and let $w_{s,t}=\mathrm{softmax}_s(\cos(z_s,z_t)/\tau)$, where the temperature $\tau$ is tuned on source-only folds. The second is a per-feature closeness $\pi_t(g)\in[0,1]$: the normalized restart-walk mass that the target disease places on gene or microRNA $g$. Training minimizes
$$ \mathcal{L}=\sum_{s} w_{s,t}\,\ell_{\text{Cox}}(\theta;D_s)+\ell_{\text{Cox}}(\theta;D_t)+\lambda\sum_{g}\bigl(1-\pi_t(g)\bigr)\,\lVert W_g\rVert_2 . $$
Here $\ell_{\text{Cox}}$ is the Cox partial-likelihood loss on a dataset, $D_s$ and $D_t$ are the source and target patients, $\theta$ are the model parameters, and $W_g$ is the column of first-layer omics weights attached to feature $g$. The last term is a group penalty. Features near the target disease are penalized less, so the model can use them first.
The backbone is held fixed and kept modest so that the comparison isolates the prior. Slides are encoded with a frozen self-supervised patch encoder and attention pooling. Fusion uses the latent-bottleneck design of HEALNet, chosen because it skips missing modalities rather than imputing them, which matters in small cohorts. The cohort weights $w_{s,t}$ act during pretraining on source cancers. The feature penalty acts throughout pretraining and during a short target fine-tune.
Baselines
The naive reference point has two parts. The first is a penalized Cox model on the target cohort alone, using clinical variables, principal components of omics and pooled slide features. PONET reports a plain Cox C-index of 0.544 on TCGA-KIRP with full data, and TSB reports 0.500 for target-only Cox on MAGGIC Euro, so we expect this baseline to be weak but not trivial. The second is HEALNet trained from scratch on the target cohort.
The strongest existing competitors are three. PONET is trained on the same subsampled target data with its own Reactome sparsity. HEALNet is pretrained on all other TCGA cohorts with uniform weights and then fine-tuned; this is the pan-cancer pooling approach that Fan et al. found effective across 19 of 20 cohorts. TSB is applied to the same tabular summaries, giving data-driven instance weighting.
For full-data reference points, the file gives PONET 0.829 and HFBSurv 0.769 on KIRP. It also gives five-cohort average C-index for KEMM (0.7123), CMTA (0.6836) and MCAT (0.6744). There is no reported result on Lacuna for any multimodal TCGA survival model at a target size of 100 or fewer. All small-N numbers will therefore be produced here, under one protocol.
Fairness rules apply to every method. All methods see identical target subsamples and identical source pools. All share the frozen slide encoder. Each gets the same hyperparameter search budget of 30 trials, scored by inner cross-validation on the target subsample only. The parameter count of the fusion head is matched within 10% across neural methods.
Two ablations isolate the key ingredient. In the first, the network is replaced by a degree-preserving random rewiring, which keeps the regularization strength but destroys the knowledge. In the second, the cohort weights and the feature penalty are switched off one at a time.
Beating the competitors means two things at target sizes of 25 and 50. Mean C-index must improve over the best competitor by at least 0.02, with the improvement holding in at least four of five held-out cancers across 20 resamples. Integrated Brier score must not get worse.
Why not just…?
Why not just pool all source cancers uniformly and fine-tune? This is the most serious threat. Pan-cancer pooling already works well at full scale (Fan et al.). If most cancers share prognostic signal, uniform weights may be close to optimal. The uniform-weight arm runs under the same budget and is the comparison that decides the cohort-weighting half of the proposal. If it ties, we keep only the feature prior and report the weighting result as negative.
Why not just weight source patients by empirical similarity, as TSB does? TSB's strong MAGGIC result came from targets of up to roughly 200 patients with tabular features. At 25 patients with high-dimensional omics, density estimates of the target distribution are noisy, which is exactly where we predict network weights win. The design allows either outcome. We also test a hybrid that starts from the network weights and moves toward empirical weights as the target sample grows. If TSB matches us even at 25 patients, the central claim is wrong.
Why not just borrow from cancers of the same organ? A skeptic could argue that "kidney borrows from kidney" captures most of what the network knows. We include an organ-of-origin weighting arm. The network weighting justifies its complexity only if it beats this arm, for example by linking cancers across organs through shared processes. The decisive correlation analysis below measures this directly.
Why not just use penalized Cox on pathway scores and frozen slide features? At 25 patients a low-capacity linear model may simply be the right answer. Bria et al. document how often medical cohorts fall below 100 subjects and how unconstrained networks fail there. We add network-derived pathway scores to the Cox baseline, giving it the same knowledge without the deep model. If it wins at the smallest sizes, we report that, and the useful deliverable becomes the network prior applied to a linear model.
Proposed Experiments
The main experiment is a leave-one-cancer-out study on TCGA. The held-out targets are KIRP, PAAD, BLCA, LUSC and UCEC; each target is in turn the small cohort, and every other TCGA cohort in the file is a source. Target training sets are subsampled to 25, 50, 100 and 200 patients, plus the full cohort, with 20 stratified draws per size. Each draw is evaluated on the remaining target patients. We report the number of events per draw, because C-index is unstable with few events. Metrics are Harrell's C-index, integrated Brier score with inverse-probability-of-censoring weighting (following TSB), and calibration slope. Results are plotted as learning curves against target size, which shows the predicted crossover point.
The decisive check comes before any of this. For each source–target pair we measure the actual transfer gain: the C-index improvement of a model pretrained on that single source, compared with no pretraining, evaluated with full target data. We then correlate this gain with network cosine similarity and with organ identity. If network similarity explains no more variance in transfer gain than organ identity, or than the randomly rewired network, the cohort-weighting idea is dead regardless of downstream numbers.
Two further checks probe the feature prior. First, if the prior encodes biology, the genes with high $\pi_t(g)$ should be the ones that receive large weights when the model is trained on the full target cohort. We test this with rank correlation, comparing small-N priors against full-N fitted weights. Second, we repeat the main comparison with RNA-seq withheld at test time for half the patients. This checks that the prior does not make the model brittle when omics is missing, a case HEALNet's skip mechanism should otherwise handle.
References
- Boosting Transfer Learning with Survival Data from Heterogeneous Domains
- Deep Biological Pathway Informed Pathology-Genomic Multimodal Survival Prediction
- GraphCare: Enhancing Healthcare Predictions with Personalized Knowledge Graphs
- MultiSourcDSim: an integrated approach for exploring disease similarity
- Pancancer survival prediction using a deep learning architecture with multimodal representation and integration
- KATZNCP: a miRNA-disease association prediction model integrating KATZ algorithm and network consistency projection
- HEALNet: Multimodal Fusion for Heterogeneous Biomedical Data
- LLM-driven Knowledge Enhancement for Multimodal Cancer Survival Prediction
- Towards AI-driven Next Generation Personalized Healthcare and Well-being