MoESurv: a zero-sample and transferable survival prediction framework for rare cancers using mixture of experts Authors: Shuping Fang, Yuhang Wang, Mengyan Zhou, Yanting Shao, Zhenghao Jiang, Yitong Guo, Tao Jiang, Feng Wang, Jingya Fang, Hong Tian Published: 2026 ABSTRACT AbstractMotivationAccurate survival prediction is crucial for personalized cancer treatment but remains challenging for rare cancers due to limited data. Most deep learning models require large training datasets, which are unavailable for rare cancer types, creating a significant clini-cal bottleneck.ResultsWe propose MoESurv, a zero-sample survival prediction framework that leverages a mix-ture-of-experts architecture to extract generalizable prognostic patterns from pan-cancer data. The model integrates shared experts, cancer-specific experts, and routing experts within an autoencoder to disentangle common and type-specific survival features. Evaluated on seven rare TCGA cancer types, MoESurv achieved state-of-the-art performance, improving the average C-index by 4 percentage points over the best baseline. Further-more, external validation across diverse populations and independent cohorts—including a Chinese glioma cohort (CGGA mRNAseq_693, C-index = 0.7433), a rare GBM IDH-mutant subtype (C-index = 0.8064), and the pan-cancer PCAWG cohort (C-index = 0.7090)—demonstrated that MoESurv possesses the most robust predictive per-formance, highlighting its generalizability. MoESurv also effectively stratified high- and low-risk patient groups and identified potential survival-associated genes, demonstrating both clinical utility and biological interpretability.AvailabilityThe code is freely available at https://github.com/HuaYC666/MoESurv and https://zenodo.org/records/20785500. 1 Introduction Accurate prognosis assessment is critical for personalizing treatment strategies and improving clinical outcomes in cancer patients (Cooperberg et al. 2023). In recent years, deep learning-based survival prediction models leveraging high-throughput transcriptomic data have made significant progress (Morganti et al. 2019). These models can capture complex molecular patterns closely associated with patient survival, enabling more precise risk stratification and outcome prediction, which in turn assists clinicians in optimizing treatment plans (Moon et al. 2023, Grist et al. 2025). Moreover, by linking gene expression with clinical endpoints, researchers can identify risk and protective genes with independent prognostic value (Gao et al. 2013, Akcakanat et al. 2021), offering important clues for uncovering disease mechanisms and developing targeted therapies. However, the development of most existing survival prediction models heavily relies on large-scale, high-quality training datasets. While feasible for common cancers such as lung or breast cancer, this requirement poses a major challenge for rare cancers—including adrenocortical carcinoma (ACC) and pheochromocytoma/paraganglioma (PCPG)—due to their very low incidence (Payabyab et al. 2016, Carrasquillo et al. 2021). The limited availability of samples hinders the construction of robust and accurate predictive models, leaving clinicians without reliable decision support for risk stratification and management. This represents a key obstacle in the field of precision medicine. An intuitive hypothesis is that, despite etiological and pathological heterogeneity across cancer types, they may share conserved biological processes—such as cell proliferation, immune evasion, or metabolic reprogramming—that are relevant to disease progression and prognosis. If such generalizable “shared prognostic information” could be effectively extracted from data-rich common cancers and transferred to rare cancers, it might be possible to overcome sample size limitations and achieve accurate survival risk prediction in a zero-sample setting. Motivated by this, we propose a cross-cancer zero-sample survival prediction framework named MoESurv (Mixture of Experts for Survival Analysis). The model deeply integrates a mixture-of-experts architecture with an autoencoder, incorporating shared experts, routing-experts, cancer-specific experts, and a routing network that work cooperatively. Shared experts extract generalizable prognostic features across cancers, specific experts handle cancer-type heterogeneity, routing experts are designed to extract heterogeneity not attributable to cancer type (such as intra-cancer heterogeneity), and the routing network dynamically assigns samples to appropriate routing experts. This design enables the identification of multiple types of heterogeneity while more efficiently capturing features shared across cancer types. In evaluations across seven rare or small-sample cancer types, MoESurv achieved the best overall zero-sample survival prediction performance, improving the average C-index by 4 percentage points over the best baseline, with a 6-point lead in KICH. Furthermore, the model helped identify potential survival-associated genes in rare cancers such as KICH and MESO, offering clues for mechanistic studies and targeted therapy development. MoESurv provides a effective solution to the problem of missing prognostic models for rare cancers, with both clinical and scientific relevance. Furthermore, external validation across diverse populations and independent cohorts—including a Chinese glioma cohort (CGGA mRNAseq_693, C-index = 0.7433), a rare GBM IDH-mutant subtype (C-index = 0.8064), and the pan-cancer PCAWG cohort (C-index = 0.7090)—demonstrated that MoESurv possesses the most robust predictive performance, highlighting its generalizability. 2 Related work Survival analysis methods can be broadly categorized into three groups: parametric models, semi parametric models represented by the Cox proportional hazards (CoxPH) model, and discrete time interval based models. Parametric models require assuming that patient survival times follow a specific distribution (Martinsson 2016), such as exponential or Weibull. However, this assumption can constrain model learning, especially when the actual data distribution deviates from the presumed form, potentially leading to significant performance degradation. To overcome this limitation, the CoxPH model assumes only proportional hazards for covariates while leaving the baseline hazard unspecified. By maximizing the partial likelihood to estimate covariate effects without imposing a parametric form on the baseline hazard, the CoxPH model maintains flexibility and enables robust analysis of survival data. Discrete time models divide the time axis into intervals and estimate survival probabilities within each interval (Lee et al. 2018, Zhao and Feng 2020), which allows them to operate without relying on the proportional hazards assumption. Although both parametric and discrete time models possess certain advantages, their implementation is often more cumbersome, making the CoxPH model the most widely adopted approach in practice. Deep learning methods, with their adaptive fitting capacity, have been increasingly applied to survival analysis using transcriptomic data, often within the CoxPH framework. For instance, Cox nnet (Ching et al. 2018) uses neural networks to extract hidden features for Cox based modeling. Similarly, DeepSurv (Katzman et al. 2018) employs neural networks to predict patient prognosis and recommend personalized treatment strategies via hazard ratios. To incorporate prior biological knowledge, DeepOmics (Zhao et al. 2021) introduces gene pathway information through masked input layers, enabling pathway aware feature extraction. Another approach transforms gene expression profiles into image like representations and applies convolutional neural networks for feature learning and risk prediction (Yan et al. 2025). Graph based methods such as AGGSurv (Ling et al. 2022) leverage multi view sparse graphs to capture comprehensive relationships in gene expression data. Most supervised learning methods rely on large training datasets to achieve good performance. However, biological data are often limited in sample size, making small sample or zero sample survival prediction an active research direction. Denoising autoencoders (Yousefi et al. 2017) have been used to extract pan cancer representations from transcriptomic data for survival prediction. Building on this, VAECox (Kim et al. 2020) employs a variational autoencoder to capture the latent distribution of gene expression and performs CoxPH modeling in that space, improving prediction in small sample settings. Methods like AUTOSurv (Jiang et al. 2024) integrate pathway information with variational autoencoders to mitigate overfitting in high dimensional, samll sample scenarios, achieving strong performance in ovarian and breast cancer prognosis. Federated learning techniques, illustrated by MOSAHit (Wen and Li 2025), enable knowledge transfer across institutions under data scarce conditions. Other small sample strategies include SVM based models tailored for predicting 5 year overall survival in hepatocellular carcinoma (Jiang et al. 2025), and meta learning frameworks (Qiu et al. 2020) that train a transferable feature extractor on pan cancer data, which can be quickly adapted to new cancer types with minimal fine tuning. Extensions of meta-learning method into multi omics data have also been proposed, such as MMOSurv (Wen and Li 2025). While small sample learning approaches partially address the challenge of rare cancers, they still require at least some target domain samples and therefore do not constitute a strict zero sample solution. At the same time, the selection of the small-size examples provided to the model for fine-tuning often significantly influences the model’s predictive performance. For rare cancers where even small sample data are unavailable or hard to choose, developing genuine zero sample prediction methods remains an important unmet need. 3 Methods 3.1 Datasets and preprocessing Gene expression data used in this study were obtained from The Cancer Genome Atlas (TCGA) database (Tomczak et al. 2015). To evaluate the zero-sample prediction performance of MoESurv on rare cancer types, gene expression data from 18 common cancer types (Table 1) were used for model training, while data from seven relatively rare or small-sample-size cancer types in TCGA (Table 2)—such as READ—were reserved for testing. All datasets were normalized during preprocessing using transcripts per million (TPM) transformation to mitigate biases arising from gene length and sequencing depth. Following TPM normalization, a log(x + 1) transformation was applied to reduce the influence of extreme values and approximate a more Gaussian distribution of the data. Table 1Summary of datasets used in the unsupervised training of MoESurv.Cancer typesFull namePatientsProp.CensoredBRCABreast Cancer12100.8364KIRCKidney Clear Cell Carcinoma6030.6683LUADLung Adenocarcinoma5650.6301HNSCHead and Neck Cancer5630.5506LGGLower Grade Glioma5210.7505SKCMMelanoma4550.5297STADStomach Cancer4430.6230OVOvarian Cancer4250.3765BLCABladder Cancer4250.5576LIHCLiver Cancer4200.6143SARCSarcoma2640.6250KIRPKidney Papillary Cell Carcinoma3200.8406CESCCervical Cancer3090.7638UCECEndometrioid Cancer1930.8135PAADPancreatic Cancer1830.4809LAMLAcute Myeloid Leukemia1610.3540GBMGlioblastoma1650.2000ESCAEsophageal Cancer1950.5846 Summary of datasets used in the unsupervised training of MoESurv. Table 2Summary of dataset from rare cancer types or small-sample-size cancer types pending survival prediction.Cancer typesFull namePatientsProp.CensoredACCAdrenocortical Cancer770.6494CHOLBile Duct Cancer450.4889READRectal Cancer1020.8039KICHKidney Chromophobe890.8652MESOMesothelioma860.1512PCPGPheochromocytoma & Paraganglioma1850.9568UCSUterine Carcinosarcoma570.3860 Summary of dataset from rare cancer types or small-sample-size cancer types pending survival prediction. For external validation, we used data from the PCAWG (Underwood 2020) and CGGA (Zhao et al. 2021) databases. The PCAWG pan-cancer cohort was used to assess the generalizability of MoESurv across different patient cohorts. The CGGA mRNAseq_693 glioma cohort, including its rare GBM IDH-mutant subset, was employed to evaluate the model’s performance across ethnic groups and its applicability to rare cancer types. We first identified the overlapping genes between the TCGA dataset and each external cohort. Gene expression matrices were then log-transformed using log(TPM + 1). The preprocessed TCGA data, encompassing 18 cancer types, was used to train the model, which was subsequently tested directly on the held-out external cohorts. 3.2 Cancer-type-specific mixture-of-experts system The Mixture of Experts (MoE) is an advanced machine learning framework that uses a routing mechanism to selectively activate specialized expert models based on input characteristics. While MoE has been widely applied in natural language processing (Liu et al. 2024, Zhu et al. 2025), its potential remains largely unexplored in biological data analysis. Cancer is highly heterogeneous, with substantial differences in gene expression patterns not only between cancer types but also within the same cancer type (Dentro et al. 2021). To build a generalizable survival prediction model, it is essential to capture common prognostic features across cancers while appropriately filtering out inter- and intra-type heterogeneity. We therefore introduce MoESurv, which employs a mixture of three expert modules with distinct roles. Shared experts extract universal gene-expression features across cancers, and cancer-type experts model type-specific characteristics. Crucially, the routing experts act as a specialized filtering system. A gating network evaluates each input sample and dynamically routes it to the most appropriate routing expert, which has become specialized in processing a particular type of genomic pattern during training. This cooperative mechanism ensures a balanced and comprehensive analysis of both shared and idiosyncratic signals. Each expert is implemented as a feed-forward network with a hidden dimension of l=1024. The network first projects the input gene expression matrix from its original dimension to an intermediate 2048-dimensional space. After applying the ReLU non-linearity (Agarap 2018), the features are mapped to the 1024-dimensional latent space, which serves as the expert’s output representation. Suppose there are ns shared experts, nr routing experts, and for each cancer-type, there are nc cancer experts. Let xi∈Rg denote the gene expression vector of the i-th sample, where g is the number of genes. MoESurv performs initial feature extraction from gene expression data using its expert modules: Where h(s),h(c),h(r)∈Rl. FFNk(s) corresponds to the k-th shared expert, FFNk,j(c) denotes the k-th cancer expert for the j-th cancer type, and FFNk(r) refers to the k-th routing expert. The weight gk,i assigned to the k-th routing expert for the i-th sample is computed to adaptively select the most suitable routing experts for each sample, as follows: Where sk,i denotes the score assigned by the router to the k-th expert for the i-th sample. The router acts as a gating mechanism that performs binary logistic regression on the input data to determine the suitability of each routing expert. ek∈Rg represents the regression coefficient vector used to compute the gating score for the k-th routing expert. A value of sk,i closer to 1 indicates that the corresponding expert is more suitable for feature extraction from the current sample. The function Topk(·,Kr) selects the top Kr (defaults to 1) experts with the highest scores for each sample, while setting the scores of all other experts to zero. To ensure balanced utilization of all routing experts and prevent the model from repeatedly activating only a fixed subset, a load-balancing loss Lb is introduced following prior work (Liu et al. 2024). This loss encourages all experts to be adequately trained during the learning process: Where g¯batch,i is the average gating weight of the i-th routing expert over the batch of samples that input the router, and Pbatch,i denotes the empirical probability of the expert being selected within the same batch of samples. 3.3 Mixture autoencoder Unlike traditional mixture-of-experts systems, MoESurv does not directly combine features from different expert modules through summation and then excute supervised training. Instead, it further employs these features to construct a mixture autoencoder. An autoencoder is a deep learning-based model that maps high-dimensional data into a low-dimensional space, learning latent representations of the data. It has been widely applied for dimensionality reduction and feature extraction in gene expression data analysis (Li et al. 2023, Qiu et al. 2025). MoESurv introduces the concept of a mixture autoencoder to extract generalizable features across multiple cancer types. The model first computes latent feature representations from each expert module independently, then combines these feature vectors via element-wise summation: Where hi(s), hi(c) and hi(r) denote the feature vectors extracted by the shared expert, cancer expert, and routing expert modules, respectively. To ensure that the mixture latent features are compact and effectively represent the gene expression data, MoESurv adopts an asymmetric decoder structure. During the reconstruction phase, the decoder performs two linear transformations. First, the mixture latent features of dimension l are projected to a 2048-dimensional latent space; subsequently, they are mapped back to the original input dimension of gene expression. This process is formulated as: Where hi(m)∈Rl denotes the vector of mixture latent features. W1 and W2 are the linear-transform matrices, respectively, and b1, b2 are the corresponding biases. The model is trained by minimizing the mean squared error reconstruction loss: This loss function guides the network to learn low-dimensional embeddings that summarize the input data effectively, encouraging the expert modules to collaboratively extract features of different cancer types. 3.4 Generalizable supervised training paradigm Although the mixture autoencoder enables expert modules to focus on learning their respective feature types, thereby helping the shared expert module capture common expression patterns in cancer and extract generalizable features, this approach alone is insufficient. The commonly shared features may not necessarily be strongly associated with patient survival, potentially introducing significant noise and compromising the generalization capability of the shared representations. We designed a generalizable supervised training paradigm to enhance the model’s ability to extract not only features useful for reconstruction but, more importantly, survival-related features that generalize across different cancer types (Fig. 1). Let Ii denote the time until the event of interest (death) for the i-th sample, and Ci the censoring time. The observed time is Oi=min⁡(Ii,Ci), and δi=I(Ii