DeepOS: pan-cancer prognosis estimation from RNA-sequencing data

DeepOS: pan-cancer prognosis estimation from RNA-sequencing data | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article DeepOS: pan-cancer prognosis estimation from RNA-sequencing data Marie Pavageau, Louis Rebaud, Charles Tanguy, Daphné Morel, Eric Deutsch, and 3 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7556679/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract RNA-sequencing (RNA-seq) analysis offers a tumor-centered approach of interest for personalizing cancer care. However, existing methods – including deep learning models – struggle to reach satisfying performances on survival prediction based upon pan-cancer RNA-seq data. Here, we present DeepOS, a deep learning model that predicts overall survival (OS) from pan-cancer RNA-seq with a concordance-index of 0.714 and a survival AUC of 0.749 across 33 TCGA tumor types whilst tested on an unseen test cohort. DeepOS notably uses (i) prior biological knowledge to condense inputs dimensionality, (ii) mean squared error adapted to survival loss function and (iii) transfer learning to enlarge its training capacity through pre-training on organ prediction to improve the model performances (factors sorted by contributions). Interpretation showed that DeepOS learned biologically-relevant prognosis biomarkers. Altogether, DeepOS achieved competitive and consistent performances on pan-cancer prognosis estimation from individual RNA-seq data. Health sciences/Biomarkers Biological sciences/Cancer Biological sciences/Computational biology and bioinformatics Health sciences/Oncology neoplasm prognosis deep learning transfer learning RNA-sequencing Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Introduction Among patients diagnosed with cancer, prognosis estimation is often required to draw a risk profile and adapt treatment accordingly. Currently recommended prognostic and predictive biomarkers that drive cancer care management usually combine several items, such as: individual characteristics (e.g. age, gender, ECOG status), tumor characteristics (e.g. tumor stage, localization and number of metastasis), serum markers (e.g. albumin, LDH, CRP) and eventually tumor molecular features (e.g. PD-L1 expression, BRCA loss-of-function, ERBB2, EGFR , BRAF , ALK mutations, NTRK fusion) 1 – 6 . Such scoring systems mainly stratify patients into low- or high-risk groups, defining therapeutic procedures to be followed. More recently, the growing interest in high-dimensional multi-omics data in assisting clinicians on treatment decision has brought forward the high potential of tumor RNA-sequencing (RNA-seq) on studying the link between tumor gene expression and patient survival outcome in a personalized way 7 . RNA-seq provides gene expression quantifications of the whole transcriptome (transcripts of more than 20,000 protein-encoding genes) or of preselected transcripts of interest (targeted sequencing), bearing the underlying hypothesis that each tumor gene expression profile mirrors the tumor aggressiveness and potential behavior in response to a particular treatment and therefore, should correlate with overall survival (OS). Since each tumor is unique in its complexity, an almost-infinite number of gene expression combinations could be expected; which drastically overcomplicates any prediction task based on RNA-seq data. Several teams recently intended to predict individual OS from RNA-seq analyses of multiple cancer types obtained from the Cancer Genome Atlas (TCGA) dataset 8 , using machine and deep learning 9 – 18 (summarized in Supplementary Table 1). Model architectures included Random Forest, Cox regression with Lasso penalization, Multilayer Perceptron (MLP), Convolutional Neural Networks, Auto-Encoder with Cox loss function, among others. Those models prediction performances in validation or test cohorts were often limited, close to 0.60, and rarely exceeded 0.62 of median concordance-index (C-index) on pan-cancer predictions. C-index is a popular metric that evaluates the ability of a model to rank survival predictions within a particular cohort rather than the difference between the predicted and the observed values (with C-indexes of 0.50 and 1.00 respectively corresponding to random and perfect order of predictions) 19 , 20 . Yet, prognosis estimation from pan-cancer RNA-seq data should be feasible since each tissue and each tumor type express their own transcriptomic signatures. With machine and deep learning, over-fitting can arise when the training dataset has greater number of dimensions (variables) than number of samples available (the size of the training set). Over-fitted models generally fail to generalize decently 21 . In the case of RNA-seq, the large dimensionality (e.g. >20,000 gene expressions) requires massive amounts of training data, which is tricky to obtain when applied to cancer and which presumably lacked to the above-mentioned models (total number of samples comprised between 953 and 11,854). We hypothesized that OS prediction using supervised deep learning on pan-cancer RNA-seq data as inputs would benefit from (i) starting with reducing input dimensions using prior biological knowledge and (ii) increasing the size of the training set, using a transfer learning strategy. Transfer learning consists in pre-training a model on a task that is related, but not strictly identical to the final question, and for which a larger number of samples is available 22 . We therefore (i) filtered whole transcriptome expressions to reduce inputs dimensionality and (ii) designed an MLP neural network that we pre-trained to predict the organ of origin from large healthy and cancer RNA-seq data and fine-tuned to predict OS from pan-cancer RNA-seq data (Fig. 1 ). We additionally characterized the parameters that influenced the model performances, including the gene list selection, the size of the training set, the type of survival loss function for training, the duration of OS, the cancer type, and the genes most implicated in the model predictions. Results Gene selection We first intended to reduce the dimensions of our input dataset by selecting genes of interest, which are known to be implicated in cancer initiation, progression, dissemination or response to treatment. We merged gene lists obtained from the Molecular Signatures Database (MSigDB, relative to hallmarks of cancer) 23 , and the LM22 immune gene signatures 24 . After removal of duplicates and genes associated with no expression values within our dataset, we obtained 4,499 genes (Supplementary Table 2). Pilot overall survival prediction task: without pre-training We designed a pilot experiment of survival estimation starting with only tumor RNA-seq data. We retrieved all The Cancer Genome Atlas (TCGA) pan-cancer RNA-seq raw data publicly-available on February, 11, 2019, via recount2 25 and selected samples associated with annotated survival outcomes (Fig. 2 a). We excluded uninformative patients who were censored during the first half of the total duration of the follow-up and the top 5% of patients with the longest OS, considering them cured by surgery. Altogether, we collected 6,529 RNA-seq samples from 33 tumor types fulfilling the criteria, among which 54.8% were censored during the second half of the follow-up. Based on this dataset, the pan-cancer median OS was 67.2 months (95% confidence interval 95%CI [64.8;72.0]) (Fig. 2 b), and highly dependent on the tumor type (Fig. 2 c). Glioblastoma, esophageal cancer, mesothelioma and pancreatic cancers were associated with the worst prognosis (median OS of 12.0 months for glioblastoma and 16.8 months for the three others); on the other hand, median OS was not reached after a 120-months follow-up for five tumor types (chromophobe and papillary renal carcinoma, pheochromocytoma, testicular cancer and thyroid cancer) (Supplementary Table 3). We randomly assigned each sample to either a training set (N = 5,529), a validation set (N = 500) or a test set for final evaluation (N = 500). Splitting was well-balanced considering the fraction of censored patients (0.546, 0.536 and 0.546, respectively within training, validation and test cohorts), median OS (67.2 months, 72.0 months and 62.4 months, respectively) and diversity of cancer types (Supplementary Table 4 and Supplementary Fig. 1). To train our models, we transformed survival data into survival probabilities per time interval and thus, could implement the classical mean squared error (MSE) loss function. Survival probabilities were set to 1 for intervals during which a patient is alive, and to 0 when a patient is deceased. Censored intervals were ignored to calculate the loss. Ignoring censored intervals in MSE allowed the model to be trained only on observed time intervals for each patient. The model learned a probability of survival for each patient individually, using classical methods in deep learning for multiclass classification. To decipher whether this approach could be competitive, we compared the performances of DeepOS to state-of-the-art deep learning model based on Cox-loss (i.e. DeepSurv and SAVAE-Cox) to train on survival data 18 , 26 . We used a Tree-structured Parzen Estimator (TPE) 27 algorithm to explore hyper-parameters and to select the best model upon the highest C-index obtained on the validation set. After training on 5,529 pan-cancer RNA-seq samples, the best model reached a C-index of 0.73 on the validation set, with cross-validation C-index mean of 0.63 and standard deviation of 0.09. The best model had five hidden layers, each of them having a dropout of 0.013, and L1 and L2 penalization of 0.007 and 0.0012, and trained with a learning rate of 0.00003. On a final and previously unseen test cohort, this model achieved a C-index of 0.714 on predicting patient survival from their pan-cancer RNA-seq data. This model surpassed DeepSurv performances on the same split data (C-index of 0.606 on validation and test sets) and SAVAE-Cox model (test C-index of 0.696) Learning curves of survival prediction To study how the number of samples within the training set influenced the model performances, we repeated the survival training task with an escalating number of samples composing the training cohort, without modifying the validation set. Learning curves indicated that C-indexes reach a steady state for training cohorts containing at least 2,000 RNA-seq samples (Fig. 3 a). The best training set C-index (0.93) was achieved with 500 samples; although the difference between training- and validation-related C-indexes indicated that the model was subject to overfitting (Fig. 3 b). Overfitting defines a model that learns too perfectly from a training set so that it fails to generalize adequately on unseen additional data. According to our results and consistently with what was previously described 28 , 29 , overfitting tends to be reduced by increasing the number of samples within the training set (Fig. 3 b). We therefore hypothesized that a transfer learning strategy could benefit our model since it could indirectly expand the training dataset through pre-training on a similar task. Transfer learning: data collection for the pre-training task Since OS distribution was related to tumor type (Fig. 2 b), we assumed that learning to predict the organ of origin from a larger cohort could improve the overall estimation of survival duration. We therefore chose to pre-train our model on the prediction of the organ of origin from the RNA-seq expression data of the 4,499 selected genes using either healthy or tumor tissue (Fig. 4 a). Healthy organs data were obtained through recount2 from the Genotype-Tissue Expression (GTEx) project 30 . We additionally retrieved all the TCGA gene expression data of tumor samples (including those not associated with survival data). Each tumor type was aligned with its organ of origin (for example, kidney chromophobe, clear cell carcinoma and papillary cell carcinoma were all considered as kidney tissue). To avoid data leakage we removed from the organ dataset the samples corresponding to the patients in the validation and test survival dataset. After removing 1086 samples from the pooled dataset of 18,571 corresponding to the samples used as validations and tests in the survival dataset, the organ dataset consisted of 17,485 RNA-seq samples from 38 distinct human tissues (Fig. 4 b). The most represented organs were brain (8.5% of samples), lung (8.2%) and breast tissue (7.9%). We randomly divided the organ dataset samples into two distinct sets for training (15,485) and validation (2,000). Splitting was well balanced and both sets harbored samples belonging to the 38 types of tissue. Pre-training on organ prediction We thus pre-trained DeepOS on organ prediction. DeepOS is an MLP neural network that takes gene expression values in transcripts per million (TPM) as inputs, and outputs either organ classification or survival probabilities (Supplementary Fig. 2). DeepOS architecture comprises hidden layers composed of stacked units of dense layers, Rectified Linear Unit (ReLU) activations, dropout effect penalization, L1 and L2 regularization, and batch normalization. Most of the models tested reached very high validation performances to predict the organ of origin (mean hyper-parameter search accuracy = 0.764 and standard deviation = 0.263), with best model reaching and validation accuracy of 0.9720, precision of 0.9698, recall of 0.9667 and F1-score of 0.9676 (Supplementary Fig. 3). The best organ-specific model had four layers, with a dropout rate of 0.014, a L1 and L2 regularization parameters respectively of 0.0010 and 0.0018 and was trained with a learning rate of 0.00019. We did not perform a test set evaluation as this step was only used to select the best pre-trained model to fine-tune on survival. Fine-tuning on survival prediction To implement our transfer learning strategy, we then fine-tuned DeepOS on survival prediction based on the pan-cancer cohort described above. The part of DeepOS pretrained on organ prediction was frozen during the first fine-tuning step and unfrozen during the second fine-tuning step (Supplementary Fig. 2). Similarly to the pilot task performed without pre-training, we selected the best version of the fully trained (comprising pre-training and fine-tuning) model based on the C-index and Kaplan-Meier (KM) statistics obtained on the validation set. The best validation C-index was 0.751 (mean cross-validation C-index = 0.728, standard deviation = 0.007). On the unseen test cohort, DeepOS achieved a C-index of 0.714 and an area under the survival ROC curve (AUC) of 0.749 (Fig. 5 a and 5 b). Transfer learning slightly improved survival prediction of + 7% for the C-index (without pre-training 0.707) and + 0.7% for the mean AUC (without pre-training 0.742). In addition to the organ-specific features, the best model architecture had six layers, with a dropout rate of 0.013 and L1 and L2 regulation parameters respectively of 0.0011 and 0.0003. We have then evaluated if these learned hyper-parameters from DeepOS full training changed the performance of a model trained only on survival (without pre-training on the organ classification problem as in the pilot survival prediction task). We found that the predicted KM curve was closer to the ground truth but the C-index didn’t improve (Supplementary Fig. 4a). It suggests that the overfitting of the pilot survival prediction tasks was mostly related to the hyperparameters learnt. DeepOS according to patient OS To interpret how DeepOS predicts pan-tumor survival, we generated a predicted Kaplan-Meier survival curve and compared it to the true survival curve of the test cohort (Fig. 5 c). We noticed that there was no significant difference between DeepOS prediction over time and the ground truth (log-rank p-value: 0.08, superior to 0.05). When applied to the training and validation cohorts, we similarly observed apparent proximity between predicted and observed curves for OS comprised between 35 and 90 months (Supplementary Fig. 4), although we repeatedly noticed divergences of the curve slopes for shorter OS, with DeepOS behaving over-optimistic as compared to reality. To further evaluate DeepOS performances according to survival duration, we generated ten subgroups of 50 patients ranked by OS on the test cohort, and computed the C-indexes of each subgroup. We could indeed observe that DeepOS performed modestly for survival predictions of patients deceased between 1 and 20 months (C-index < 0.58; <0.55 without pre-training) (Supplementary Fig. 5a). This observation was not associated with underrepresentation of such population within the training set (Supplementary Fig. 5b). DeepOS according to tumor type Among the 29 cancer types that contained at least 3 uncensored samples, 23(79.3%) displayed a C-index > 0.50 which corresponds to better than random prediction and 8(27.6%) had a C-index > 0.7145 on the test set, including three that reached the perfect score of 1 (Adrenocortical carcinoma, Testicular Germ Cell Tumors and Uveal Melanoma), although these were composed of only 5, 6, and 4 patients within the test set, respectively (Fig. 6 a). Despite our previous observation, DeepOS was able to perform reasonably well on four out the five tumor types displaying a median OS < 20 months, with C-indexes comprised between 0.55 and 0.74 (ESCA, MESO, PAAD, STAD), but the glioblastoma group was ill-predicted, (Supplementary Table 5). Transfer learning unchanged or improved survival predictions for 20 (69.0%) tumor types, up to + 28,5%, as assessed by C-index calculation, but achieved worst performance that the pilot survival model for 9 tumor types including frequent ones (COAD, ESCA, GBM, KICH, LUSC, PAAD, READ, SARC, SKCM) (Supplementary Fig. 6a, Supplementary Table 5). DeepOS few-shot learning per tumor type We have evaluated if few-shot learning could further improve DeepOs performance on frequent tumor types with low C-index such as BRCA, LUAD, LUSC. Few-shot learning of DeepOs improved the test performance for LUSC and BRCA, but decreased the performance for LUAD (Supplementary Fig. 6b). In addition, it improved the proximity to the ground-truth survival curve of patients affected by these frequent tumor types (log-rank p-value 0.61, c-index 0.67, Supplementary Fig. 6c). DeepOS according to gene selection We then evaluated if the gene set selected on prior biological knowledge helped DeepOS performance. We first built a model without gene selection, and followed the same training as with DeepOS, i.e. pre-training on organ prediction and fine-tuning on survival prediction with hyperparameter search. We found good ranking performances but the predicted Kaplan-Meier curve showed very poor relation to the real survival curve (Supplementary Fig. 7) We then trained models with dimension reduction using 100 random selections of 4,499 genes (examples of gene sets in Supplementary Table 2). Using random genes as inputs impaired the model performances, achieving an average C-index of 0.69 and a standard deviation of 0.009 on the test set (Fig. 5 a). Predicted Kaplan-Meier survival curves were significantly different to the ground truth with mean p-values of 2.6*10 − 5 and mean standard deviation of 2.5*10 − 4. Gene selection based on prior-knowledge thus contributed significantly to DeepOS generalization performance. DeepOS according to gene expressions To finally better characterize which genes were the most important for DeepOS predictions, we estimated the mean SHapley Additive exPlanations (SHAP) values. SHAP provides an interpretation of the importance attributed by the algorithm to each input feature 31 . Among the 4,499 genes, 496 had a mean absolute SHAP value > 0.001 and 57 genes had a mean absolute SHAP value > 0.005, which we considered as the most important for DeepOS predictions (Fig. 6 b, Supplementary Fig. 8a). We compared the effect of gene expressions importance and direction for the model decision with known biological findings. Gene expression of FOX2A was the most important feature, with high expression values correlated with poor OS. This is also consistent with the known role of FOX2A, a transcription factor promoting proliferation and epithelial-mesenchymal transition in multiple cancer types 32 – 35 . The second most important gene was CLDN18 , encoding for the Claudin-18 protein involved in tight junctions between epithelial cells and altered in epithelial-mesenchymal transition. Claudin-18.2 is a therapeutic target in gastric cancer and under development in pancreatic cancers. The prognostic role of Claudin-18 is debated and may depend on the tumor type 36 , 37 . Consistently DeepOS used low expression to relate to prognosis and its importance was different in pancreatic and gastric cancers compared to other tumor types (supplementary Fig. 8b). An example of a similar relation between expression and prognosis in DeepOS is the SST gene encoding for somatostatin, which also is relevant compared to the literature showing that low expression relates to poor survival in colorectal cancers 38 . Conclusions, discussion We have developed a deep learning model, DeepOS, to estimate OS from pan-cancer RNA-seq data with several methodological improvement in the field including (i) prior biological knowledge to condense inputs dimensionality, (ii) transfer learning to enlarge its training capacity through pre-training on organ prediction, and (iii) mean squared error adapted to survival loss function. Although we first observed a relation between the performance of survival prediction with the number of observations to learn from, transfer learning only slightly improved the prediction performances. The pre-training task consisted in predicting the organ of origin, motivated by an important relation between the organ of origin and cancer prognosis. We obtained very high accuracy, precision and recall performances (0.9720, 0.970, 0.967 each on the validation set). For survival prediction, DeepOS reached a median pan-cancer C-index of 0.714 on a previously unseen test set and a survival AUC of 0.75. DeepOS can output a discrete estimation of individual survival, which enables us to plot Kaplan-Meier survival curves from individual predictions. Doing so, we could confirm that DeepOS survival predictions over time were not different from the ground truth in the test set (log-rank p-value 0.08). We also found a way to use the mean square error, a classical loss function in deep learning, to train DeepOS on survival data. We used Cox-loss based models (DeepSurv and SAVAE-Cox) as an internal comparator using the same training, validation and testing data split. DeepSurv is a deep neural network trained with a Cox proportional hazards loss function, which is considered a state-of-the-art method for survival prediction 26 . DeepOS significantly outperformed DeepSurv model performances (DeepSurv test C-indexes were between 0.60 and 0.61, and AUC of 0.63 on the test set, and SAVAE-Cox test c-index of 0.69, similarly to other published models for such task; Supplementary Table 1). Finally, we observed that input dimension reduction from prior-knowledge improved the model performance compared to random gene set selection and limited overfitting compared to the largest input possible. This is relevant with other studies using prior-knowledge dimension reduction for other tasks such as treatment effect prediction in cell lines or systems biology in cancer 39 , 40 . Our model has limitations. Firstly, as compared to the survival prediction task ran without pre-training, one could argue that the transfer learning strategy only slightly improved the model performances (+ 0.7% on the test C-index). We found a contrast between the metrics (C-index, survival AUC, log rank test) and the visual interpretation of the Kaplan-Meier curves, supporting the importance of this way to clinically interpret the prediction of models in a cohort of patients. The model without pretraining had a significant decrease in performance for long term predictions (Supplementary Fig. 4) compared to DeepOS. The overfitting observed in the pilot survival prediction task without pre-training was in large part due to overfitting of the hyperparameters. Pre-training on organ prediction is a pragmatic choice given the availability of public molecular data with relatively controlled batch effect, and could have helped on long term survival estimation and hyperparameter search. Research on pre-trained fondation survival models including molecular data could help future improvements. Besides, even minor upgrading in performance is challenging to obtain for C-index values above 0.70 for such task. Our study supports a benefice of using large tumor RNA-seq datasets with survival observation. Secondly, we noted over-optimism in DeepOS predictions for short survival durations (mainly < 20 months). However, it did not negatively impact the model performances within tumor types of unfavorable prognosis. For example, the C-index within the pancreatic adenocarcinoma, esophageal carcinoma, and the stomach adenocarcinoma cohorts reached above 0.55, while median OS was below 20 months (16.8, 16.8, and 19.2 months respectively). A possible explanation is that patients with short survival have poor prognosis factors such as tumor location, poor general condition or comorbidities that are missed by using only RNA-seq data from a microscopic tumor sample. We further explored this hypothesis in another work and indeed demonstrated that short survival is better predicted by clinical variables enclosed in consultations reports 41 (submitted to the same collection). Analyzing the influence of time on models performance should be generalized to better define the application framework of deep learning models for such tasks. We also explored few-shot learning to refine the performance on frequent tumor types and showed that improvements were not always possible. Finally, the pan-cancer TCGA RNA-seq dataset that we used for our study was built mostly upon primary tumor samples obtained from surgical resection of localized or locally-advanced diseases. This is highlighted by the observed median OS for several cancer types longer than expected for metastatic stages at diagnosis. It is nevertheless possible that DeepOS learned from the metastatic potential of the tumor samples. This is supported by the detail of the genes with the highest importance for the algorithm, which are mainly related to cancer progression and epithelial-mesenchymal transition and thus, to cancer dissemination. However, further refinement and validation studies are warranted to statue on the generalizability of the model in metastatic cancers. Other approaches have proposed clustering analysis from RNA-seq to identify groups of patients with similar prognosis 12 42 . Thorsson et al. could identify six immune subtype features from TCGA pan-cancer data comprising RNA-seq, miRNA-seq and exome sequencing data 42 . They rigorously characterized immune subtypes associated with good and poor prognosis, although pan-cancer performances were modest with a median pan-cancer C-index visually lower than 0.60. In addition to transfer learning, DeepOS comprised methodological adaptations that we believe have permitted this upgrading. First, we transformed survival data into time interval survival probabilities, so that censored time intervals did not influence the loss function calculation. Thus, we could train on the mean squared error. We also reduced input dimensions by applying prior knowledge on the biology of cancer and immunity to limit overfitting due to irrelevant genes for our task, which contributed to improve the model predictions. Overall, our study demonstrated and/or validated that (i) predicting survival outcomes from pan-cancer RNA-seq data is feasible and can achieve decent performances, (ii) dimension reduction based on prior knowledge improve the performance, and (iii) partially censored survival data can be used to train supervised deep learning models with standard loss functions. DeepOS offers a promising proof-of-concept that prognosis estimation among patients affected with various types of cancer can be personalized beyond classical score calculations. It provides a more tumor-centered way to estimate the disease aggressiveness and perhaps, to estimate its sensitivity to multiple therapeutic options. Methods Objectives - This study aimed at predicting the survival of patients affected by various tumor types from their gene expression analysis. This is a classical task with gold-standard datasets that we used to evaluate methodological improvements (Supplementary Table 1). We have developed a new format of survival data to train deep learning models, a prior-knowledge based dimension reduction and a transfer learning strategy. We hypothesized that these methods should help model performance and interpretability. Labels – Survival - We used the publicly available survival data of the TCGA database from Liu et al. 43 . The top 5% of patients with the highest overall survival were removed because they were considered cured (by surgery, as their overall survival was higher than nine years). Patients with no follow-up were also removed (i.e. 0 days or survival status not known). Early censored patients had poor relevance for the training; we thus removed patients censored before the median follow-up of the cohort. We then performed a random split of the data (80%, 10%, 10%). Labels – Organ - For the pre-training on organ prediction, we have pooled GTEx and TCGA data 8 , 30 . GTEx concerned the analysis of normal organs and TCGA the analysis of primary tumors classified by organs of origin. A random split (90%, 10%) was performed on the organ data set (no test cohort was required as the organ data were used for the pre-training task). To avoid data leakage from the test set into the training set, we made sure to exclude from the organ dataset any patient that would be present in the validation set and test set from the survival dataset. Input – RNA sequencing - Inputs used to feed DeepOS were gene expression values estimated from RNA-seq. RNA-seq was the most frequent analysis commonly performed in both GTEx and TCGA and allowed to gather a maximum of examples matched with the labels described above. RNA-seq is a multistep process. RNA is first extracted from the tissue sample and sequenced. For TCGA, a vast majority of primary tumor samples came from surgical interventions while for GTEx, it came from non-diseased tissue samples from human donors. Gene expression is then estimated by the number of RNA fragments corresponding to a genome locus from a sequenced sample. TCGA and GTEx gene expressions were analyzed with the same bioinformatic pipeline from raw sequencing data and available in Recount2 25 . Gene expression was estimated in TPM (transcripts per million) with the Rail-RNA pipeline 44 . TPM followed a Poisson distribution, so we log-transformed and scaled the data matrix using natural logarithm. Input – Dimension reduction on prior-knowledge - RNA-seq gene expression data is usually highly dimensional (~ 23k protein coding genes plus non-coding regions) which can be a source of overfitting during the learning step of deep neural networks 28 . To reduce the dimensions of input data, we selected important cancer-related genes based on prior-knowledge. MSigDB database 23 provided gene lists related to cancer hallmark and LM22 provided immune cell line specific gene lists 24 that are important mechanisms for cancer evolution. These two sources comprised a total of 4,499 genes also found in GTEx and TCGA RNA-seq data. Input – study of the gene selection - For comparison, we trained models with random selections of 4,499 input genes, excluding the ones found from cancer hallmarks and LM22. We trained those models on the same RNA-seq data, using the same workflow (hyper-parameter search and selection of the best model on validation C-index). We replicated the experiment twice, each time with different selections of random genes (#1 and #2). Models’ architecture - DeepOS model is a multilayers perceptron (MLP), which consists of at least three types of layers: the input layer, hidden layers and the output layer. Except for the input data, each unit uses a linear function using parameters W and b, activated by a nonlinear function such as ReLU used here for the hidden layers. Training was supervised using the backpropagation of the gradients of the error to improve model predictions, step by step, by correcting the parameters W and b. The last layer of our model was composed of linear functions. Loss – Survival loss - Patient survival in TCGA was calculated by the number of days to death (the event of interest) since the date of sampling. Censored patients were patients that were still alive (have not presented the event of interest) at the time of end of follow-up. Patients with good outcomes are thus more prone to be censored. Removing censored patients would influence the model to be over-pessimistic and would decrease the number of examples for training. Keeping censored patients leads to challenges in the design of a loss function to minimize. We have developed and implemented an approach to train deep learning models on survival data. With this approach, follow up was divided into a vector of B time-bins (or time intervals). In the raw data, each day of the follow up was associated with one value: 1 if the patient is alive, 0 if he is deceased and − 1 if he is censored. The value of a bin was the mean of the values of each day included in this bin. The bin value ranged from − 1 to 1. For example, the bin values corresponding to a time interval of 5 days for a patient deceased at day 4 are the following: In days: [1, 1, 1, 0, 0, 0, 0, 0, 0, 0] ⇒ bin values: [0.6, 0] Concerning a patient censored at day 3 the bin values are: In days: [1, 1, -1, -1, -1, -1, -1, -1, -1, -1] ⇒ bin values: [-0.2, -1] We used the MSE, a classical loss function used to backpropagate the error of deep learning models (examples in Supplementary Table 6). The MSE is given by: $$\:MSE=\:\frac{1}{m}{\sum\:}_{i=1}^{n}({y\_true}_{i}-{{y}_{pred}}_{i})²$$ with m the number of non-censored patients and n the number of patients. Consequently, the model output layer was designed as a vector of survival probabilities over time with the number of neurons corresponding to the number of bins. Censored values were ignored in the computation of MSE and doing so, the model was trained only on the observed follow up. The cutoff probability value for the model to predict time to death was set to 0.5 and first bin with value less than 0.5 was considered. Loss – Cox loss - Most of the previous studies predicting survival from RNA-seq used Cox proportional hazard model to handle censored survival data. As a control, we trained a model with DeepSurv, a MLP with a Cox log-likelihood loss function 26 . Loss – For organ prediction - Categorical cross entropy was used for the organ prediction task which consisted in 38 classes. Training – Penalization and learning - Penalization comprises a set of classical methods to prevent overfitting during training, such as L1 and L2 regularization and dropout. Another method proposed to limit overfitting consists in adding Gaussian noise to the input data for each epoch during the training step 45 . Adam optimizer and batch normalization were also used to accelerate convergence 46 . Training – Hyper-parameters optimization - Hyper-parameters are parameters controlling the MLP architecture, learning strategies and/or penalization of the learning. We have optimized the following hyper-parameters: - The number of layers in the MLP; - The number of nodes of the first hidden layer; - The decrease rate of the number of unit per layer (rate by which the number of nodes of the previous layer is multiplied to determine the number of nodes of the current layer); - The learning rates lr1 (for organ prediction task) and/or lr2 (for survival prediction task); - The regularization parameters: - The standard deviation of the gaussian noise added to input data; - The dropout rate (continuous values within [0, 0.8]); - Lambda values for L1 and L2 normalization; - The batch size; - The number of epochs of learning. Considering the two training tasks (organ and survival), the hyper-parameters search space had 24 dimensions. We used the Tree-structured Parzen Estimator (TPE) algorithm to train DeepOS hyper-parameters 27 . TPE is a Bayesian approach that outperformed the traditional grid search and random search on hyper-parameters search. For each new set of hyper-parameters a new random model was fully trained. Performance metrics were calculated on the validation set(s) (for organ and/or survival). New hyper-parameters were inferred from the validation performance by the TPE algorithm. We performed 500 trials for hyper-parameters search, based on previous studies 47 . The model with the best performance on the validation set was finally evaluated on the test set. Transfer learning strategy - The transfer learning strategy for DeepOS was composed of pre-training on organ prediction and fine-tuning on survival prediction, each of these steps with independent hyper-parameters search. We used validation accuracy to select the best model on organ prediction. We then added new layers (number defined by the hyper-parameter search) and an output layer to this model. We froze the organ layers for the first fine-tuning step on survival (including hyper-parameter search), considering it as a low abstraction representation of gene expression. A second fine-tuning step (including hyper-parameter search) was performed on the same MLP with all layers unfrozen. The final survival model selection was based on the validation cohort. Evaluation of the model - Metric for organ prediction - To evaluate the performance of the model on the organ task, we used classification metrics: accuracy, precision and F1 score. Evaluation of the model - Metric for survival prediction - We used the concordance correlation coefficient (concordance index, or C-index) to evaluate survival models with censored data 19 , 20 . C-index represents the proportion of concordant pairs divided by the total number of possible evaluation pairs. For example, if a patient A has died at time tA and a patient B has been censored at time tB, they can still be compared if tA < tB. If the model gives a prediction pA for patient A and pB for patient B, the pair can be qualified as concordant if pA tB then it is not possible to evaluate this pair and it will not count as a possible evaluation pair. We also computed the survival AUROC using sklearn (sksurv.metrics.cumulative_dynamic_auc), which is a cumulative area under the ROC curve adapted to censored data 48 . Finally, we used the p-value of the log-rank test to compare the predicted Kaplan-Meier survival curve to the ground truth. The log-rank test determines if two survival curves are statistically equivalent (null hypothesis) with a chi2 test. The p-value gives indication on whether we should reject the null hypothesis: the smaller it is the more two survival curves are different. Conversely, neural networks trained with Cox loss predict a risk and are barely used to predict individual survival in time; therefore log rank has not been used to date in this setting, to our knowledge. Evaluation of the model - Performances by survival time – To further evaluate DeepOS predictions, we have assessed the performance depending on survival time. We have sorted the test cohort by OS and divided the cohort into 10 subgroups, each group composed of 50 patients. We have then computed the C-index of each subgroup. Evaluation of the model - Learning curves - In order to evaluate the effect of the training set size on the model performances for survival prediction, we have generated learning curves. We used the validation cohort of 500 patients, given by the data split described previously. For the training set, data were iteratively and randomly added, from 500 to 5,529 samples, with steps of 500 samples. Every time a new training and hyper-parameters search was launched with several sampling procedures. The C-indexes were computed for the training and validation sets with plots for the median, the first and the third quartile. Evaluation of the model – comparison to Cox-loss models DeepSurv model We trained, validated and tested DeepSurv model architecture on the same survival data split and its performances were compared to DeepOS 26 . We used the same hyper-parameter search strategy based on the TPE algorithm. We did not perform pre-training on organ prediction with DeepSurv because of incompatibility with the Cox-loss, and because the objective was to compare our model to the existing literature. SAEVAE-Cox model We trained SAVAE-Cox model (Self Attention Variational AutoEncoder - Cox) in two steps (Supplementary Fig. 9). A pre-training step used an auto-encoder with self-attention. The auto-encoder architecture has a bottleneck shape for latent space representation with an encoder part that reduces dimension and a decoder part. The training consisted in outputting the input RNA-seq using a discriminator model, for 300 epochs on the train set. We fine-tuned the pre-trained encoder part with the Cox-loss on the survival data for 30 epochs. Model interpretability – While still an active research field, some techniques allow interpreting MLP training. SHAP values were used, a model agnostic technique that quantifies the influence of each input on the model’s predictions 31 . SHAP values give an input-output correlation mixed with feature importance. Code and libraries To load and process the GTEx and TCGA data we have used the R package recount2 21 . We have used python 3 with Keras 2.2.5 and Tensorflow 1.14, to build and train the model. Hyper-parameter search with Tree-structured Parzen Estimator (TPE) was performed with the Optuna library 49 . Declarations Code, model and data availability The code to load and preprocess the data, together with the code to build, train and test the model is publically available on www.github.com/DITEP/DeepOS. The preprocessed data, ready to be inputted in the model, is also publically available for maximum transparency. We provided Jupiter notebooks to navigate intuitively through the steps of the analysis with results and figures included. The user that would want to run the analysis may have slightly different results as few steps are randomized (weight initialization and hyper-parameter search for example). DeepOS model trained and presented in this paper is also provided under Keras hd5 format. Acknowledgments The authors are grateful to their colleagues and collaborators for their advices and support for this work and specifically: Rebecca Clodion, Roger Sun, Eric Angevin, Antoine Hollebecque, Daniel Gautheret, Stefan Michiels, Fabrice André, Andrei Zinovyev, Laurence Calzone, Emmanuel Barillot, Jean-Yves Blay, Jean-Charles Soria, Pierre Saintigny. The results shown here are based upon data generated by the TCGA Research Network: https://www.cancer.gov/tcga. Author contributions Conception and design: MP, LR, HV, LV Development of methodology: MP, LR, LV, CT Acquisition, analysis and/or interpretation: MP, LR, DM, LV, CT Writing, review and/or revision of the manuscript: all authors Supervision: LV Competing Interests statement LV reports personal fees from Adaptherapy, non-personal fees from Pierre-Fabre and Servier, grants from Bristol-Myers Squibb, all outside the submitted work. As part of the Drug Development Department (DITEP) of Gustave Roussy (France), LV, ED and CM report being: Principal/sub-Investigator of Clinical Trials for Abbvie, Adaptimmune, Aduro Biotech, Agios Pharmaceuticals, Amgen, Argen-X Bvba, Arno Therapeutics, Astex Pharmaceuticals, Astra Zeneca Ab, Aveo, Basilea Pharmaceutica International Ltd, Bayer Healthcare Ag, Bbb Technologies Bv, Beigene, Blueprint Medicines, Boehringer Ingelheim, Boston Pharmaceuticals, Bristol Myers Squibb, Ca, Celgene Corporation, Chugai Pharmaceutical Co, Clovis Oncology, Cullinan-Apollo, Daiichi Sankyo, Debiopharm, Eisai, Eisai Limited, Eli Lilly, Exelixis, Faron Pharmaceuticals Ltd, Forma Tharapeutics, Gamamabs, Genentech, Glaxosmithkline, H3 Biomedicine, Hoffmann La Roche Ag, Imcheck Therapeutics, Innate Pharma, Institut De Recherche Pierre Fabre, Iris Servier, Janssen Cilag, Janssen Research Foundation, Kura Oncology, Kyowa Kirin Pharm. Dev, Lilly France, Loxo Oncology, Lytix Biopharma As, Medimmune, Menarini Ricerche, Merck Sharp & Dohme Chibret, Merrimack Pharmaceuticals, Merus, Millennium Pharmaceuticals, Molecular Partners Ag, Nanobiotix, Nektar Therapeutics, Novartis Pharma, Octimet Oncology Nv, Oncoethix, Oncopeptides, Orion Pharma, Ose Pharma, Pfizer, Pharma Mar, Pierre Fabre, Medicament, Roche, Sanofi Aventis, Seattle Genetics, Sotio A.S, Syros Pharmaceuticals, Taiho Pharma, Tesaro, Xencor. Research Grants from Astrazeneca, BMS, Boehringer Ingelheim, Janssen Cilag, Merck, Novartis, Onxeo, Pfizer, Roche, Sanofi. Non-financial support (drug supplied) from Astrazeneca, Bayer, BMS, Boringher Ingelheim, Medimmune, Merck, NH TherAGuiX, Onxeo, Pfizer. ED reports grants and personal fees from Roche Genentech, grants from Boehringer, grants from Astrazeneca, grants and personal fees from Merck Serono, grants from BMS, and grants from MSD Roche. Other authors have no conflict of interest to declare. Funding This project has been funded in part and realized in partnership with ARC foundation for cancer research: Fondation ARC pour la recherche clinique – 9 rue Guy Môquet 94803 Villejuif – France. Grant number SIGNIT201901302. Previous presentation An intermediate version of this work has been presented at ESMO 2019 congress under the reference: Abstract 165P - Enhanced performance of prognostic estimation from TCGA RNAseq data using transfer learning. H Vanacker, E Angevin, A Hollebecque, R Sun, E Deutsch, A Zinovyev, L Calzone, E Barillot, C Massard, L Verlingue. Annals of Oncology, Volume 30, Issue Supplement_5, October 2019 References Cardoso, F. et al. Early breast cancer: ESMO Clinical Practice Guidelines for diagnosis, treatment and follow-up. Annals of Oncology 30 , 1194–1220 (2019). Escudier, B. et al. Renal cell carcinoma: ESMO Clinical Practice Guidelines for diagnosis, treatment and follow-up†. Annals of Oncology 30 , 706–720 (2019). Glynne-Jones, R. et al. Rectal cancer: ESMO Clinical Practice Guidelines for diagnosis, treatment and follow-up†. Annals of Oncology 28 , iv22–iv40 (2017). Michielin, O., van Akkooi, A. C. J., Ascierto, P. A., Dummer, R. & Keilholz, U. Cutaneous melanoma: ESMO Clinical Practice Guidelines for diagnosis, treatment and follow-up. Annals of Oncology 30 , 1884–1901 (2019). Colorectal Cancer Screening - NCCN Clinical Practice Guidelines in Oncology V2.2020. (2020). Lung Cancer Screening - NCCN Clinical Practice Guidelines in Oncology V1.2021. (2020). Kuksin, M. et al. Applications of single-cell and bulk RNA sequencing in onco-immunology. European Journal of Cancer 149 , 193–210 (2021). https://www.cancer.gov/tcga. Yuan, Y. et al. Assessing the clinical utility of cancer genomic and proteomic data across tumor types. Nature Biotechnology 32 , 644–652 (2014). Zheng, X., Amos, C. I. & Frost, H. R. Comparison of pathway and gene-level models for cancer prognosis prediction. BMC Bioinformatics 21 , 76 (2020). Ching, T., Zhu, X. & Garmire, L. Cox-nnet: An artificial neural network method for prognosis prediction of high-throughput omics data. PLoS Comput Biol 14 , e1006076–e1006076 (2018). Ramazzotti, D., Lal, A., Wang, B., Batzoglou, S. & Sidow, A. Multi-omic tumor data reveal diversity of molecular mechanisms that correlate with survival. Nature Communications 9 , 4453 (2018). Cheerla, A. & Gevaert, O. Deep learning with multimodal representation for pancancer prognosis prediction. Bioinformatics 35 , i446–i454 (2019). Huang, Z. et al. Deep learning-based cancer survival prognosis from RNA-seq data: approaches and evaluations. BMC Medical Genomics 13 , 41 (2020). Wang, Z. & Sun, J. SurvTRACE: Transformers for Survival Analysis with Competing Events. in 1–9 (2022). doi:10.1145/3535508.3545521. Zhao, Y. et al. BERTSurv: BERT-Based Survival Models for Predicting Outcomes of Trauma Patients. arXiv.org https://arxiv.org/abs/2103.10928v1 (2021). Hu, S., Fridgeirsson, E., Wingen, G. van & Welling, M. Transformer-Based Deep Survival Analysis. in Proceedings of AAAI Spring Symposium on Survival Prediction - Algorithms, Challenges, and Applications 2021 132–148 (PMLR, 2021). Meng, X. et al. A Novel Attention-Mechanism Based Cox Survival Model by Exploiting Pan-Cancer Empirical Genomic Information. Cells 11 , 1421 (2022). Harrell, F. E., Califf, R. M., Pryor, D. B., Lee, K. L. & Rosati, R. A. Evaluating the yield of medical tests. JAMA 247 , 2543–2546 (1982). Uno, H., Cai, T., Pencina, M. J., D’Agostino, R. B. & Wei, L. J. On the C-statistics for evaluating overall adequacy of risk prediction procedures with censored survival data. Stat Med 30 , 1105–1117 (2011). Teschendorff, A. E. Avoiding common pitfalls in machine learning omic data science. Nat. Mater. 18 , 422–427 (2019). Raffel, C. et al. Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer. arXiv:1910.10683 [cs, stat] (2020). Liberzon, A. et al. The Molecular Signatures Database (MSigDB) hallmark gene set collection. Cell Syst 1 , 417–425 (2015). Newman, A. M. et al. Robust enumeration of cell subsets from tissue expression profiles. Nature methods 12 , 453–7 (2015). Collado-Torres, L. et al. Reproducible RNA-seq analysis using recount2. Nat Biotechnol 35 , 319–321 (2017). Katzman, J. et al. DeepSurv: Personalized Treatment Recommender System Using A Cox Proportional Hazards Deep Neural Network. BMC Med Res Methodol 18 , 24 (2018). Bergstra, J., Bardenet, R., Bengio, Y. & Kégl, B. Algorithms for Hyper-Parameter Optimization. Advances in Neural Information Processing Systems 24 , 2546–2554 (2011). Everitt, B. S. & Skrondal, A. The Cambridge Dictionary of Statistics, Fourth Edition . (2011). Banko, M. & Brill, E. Scaling to Very Very Large Corpora for Natural Language Disambiguation. in Proceedings of the 39th Annual Meeting of the Association for Computational Linguistics 26–33 (Association for Computational Linguistics, Toulouse, France, 2001). doi:10.3115/1073012.1073017. https://gtexportal.org/home/. Lundberg, S. M. & Lee, S.-I. A Unified Approach to Interpreting Model Predictions. Advances in Neural Information Processing Systems 30 , 4765–4774 (2017). Bhat, A. A. et al. Tight Junction Proteins and Signaling Pathways in Cancer and Inflammation: A Functional Crosstalk. Front. Physiol. 9 , (2019). Gröne, J. et al. Differential expression of genes encoding tight junction proteins in colorectal cancer: frequent dysregulation of claudin-1, -8 and -12. Int J Colorectal Dis 22 , 651–659 (2007). Kim, S. S. et al. Immunohistochemical stain for cytokeratin 7, S100A1 and claudin 8 is valuable in differential diagnosis of chromophobe renal cell carcinoma from renal oncocytoma. Histopathology 54 , 633–635 (2009). https://www.proteinatlas.org/ENSG00000156284-CLDN8/pathology/renal+cancer. Hong, J. Y. et al. Claudin 18.2 expression in various tumor types and its role as a potential target in advanced gastric cancer. Transl Cancer Res 9 , 3367–3374 (2020). Wang, C. et al. CLDN18.2 expression and its impact on prognosis and the immune microenvironment in gastric cancer. BMC Gastroenterol 23 , 283 (2023). Yu, X. et al. Immunological role and prognostic value of somatostatin receptor family members in colon adenocarcinoma. Front Pharmacol 14 , 1255809 (2023). Manica, M. et al. Toward Explainable Anticancer Compound Sensitivity Prediction via Multimodal Attention-Based Convolutional Encoders. Mol. Pharm. 16 , 4797–4806 (2019). Bauer, E. & Thiele, I. From Network Analysis to Functional Metabolic Modeling of the Human Gut Microbiota. mSystems 3 , (2018). Piat, C. et al. A Validated and Explainable Deep Learning Model Instantly Predicts Survival from Consultation Reports. SSRN Scholarly Paper at https://doi.org/10.2139/ssrn.4410792 (2023). Thorsson, V. et al. The Immune Landscape of Cancer. Immunity 48 , 812-830.e14 (2018). Liu, J. et al. An Integrated TCGA Pan-Cancer Clinical Data Resource to Drive High-Quality Survival Outcome Analytics. Cell 173 , 400-416.e11 (2018). Nellore, A. et al. Rail-RNA: scalable analysis of RNA-seq splicing and coverage. Bioinformatics 33 , 4033–4040 (2017). An, G. The Effects of Adding Noise During Backpropagation Training on a Generalization Performance. Neural Computation 8 , 643–674 (1996). Heaton, J. Ian Goodfellow, Yoshua Bengio, and Aaron Courville: Deep learning: The MIT Press, 2016, 800 pp, ISBN: 0262035618. Genetic Programming and Evolvable Machines 19 , (2017). Bergstra, J., Yamins, D. & Cox, D. D. Making a science of model search: hyperparameter optimization in hundreds of dimensions for vision architectures. in Proceedings of the 30th International Conference on International Conference on Machine Learning - Volume 28 I-115-I–123 (JMLR.org, Atlanta, GA, USA, 2013). deCastro, B. R. Cumulative ROC curves for discriminating three or more ordinal outcomes with cutpoints on a shared continuous measurement scale. PLOS ONE 14 , e0221433 (2019). Akiba, T., Sano, S., Yanase, T., Ohta, T. & Koyama, M. Optuna: A Next-generation Hyperparameter Optimization Framework. arXiv:1907.10902 [cs, stat] (2019). Additional Declarations No competing interests reported. Supplementary Files DeepOSFiguresSuppl.pptx DeepOSSuppLegend.docx DeepOSTablesSuppl.xlsx Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-7556679","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":538570695,"identity":"becd848d-5ff1-47d2-9e1b-e778a0671a00","order_by":0,"name":"Marie Pavageau","email":"","orcid":"","institution":"Gustave Roussy - Cancer Campus","correspondingAuthor":false,"prefix":"","firstName":"Marie","middleName":"","lastName":"Pavageau","suffix":""},{"id":538570696,"identity":"9e027983-e247-47d6-81a9-3a64a26446da","order_by":1,"name":"Louis Rebaud","email":"","orcid":"","institution":"Gustave Roussy - Cancer Campus","correspondingAuthor":false,"prefix":"","firstName":"Louis","middleName":"","lastName":"Rebaud","suffix":""},{"id":538570697,"identity":"73eb67f4-b4cc-4b0b-ab1f-8237e858f35c","order_by":2,"name":"Charles Tanguy","email":"","orcid":"","institution":"","correspondingAuthor":false,"prefix":"","firstName":"Charles","middleName":"","lastName":"Tanguy","suffix":""},{"id":538570698,"identity":"00633eec-b6bd-45cf-ab92-83ddf35c7c65","order_by":3,"name":"Daphné Morel","email":"","orcid":"","institution":"Gustave Roussy - Cancer Campus","correspondingAuthor":false,"prefix":"","firstName":"Daphné","middleName":"","lastName":"Morel","suffix":""},{"id":538570699,"identity":"78593af4-4cb3-41cd-ad47-24e6289c3ecc","order_by":4,"name":"Eric Deutsch","email":"","orcid":"","institution":"Gustave Roussy - Cancer Campus","correspondingAuthor":false,"prefix":"","firstName":"Eric","middleName":"","lastName":"Deutsch","suffix":""},{"id":538570700,"identity":"4ffab364-4364-44cf-8dd7-6707da62423c","order_by":5,"name":"Christophe Massard","email":"","orcid":"","institution":"Gustave Roussy - Cancer Campus","correspondingAuthor":false,"prefix":"","firstName":"Christophe","middleName":"","lastName":"Massard","suffix":""},{"id":538570701,"identity":"c97be7fc-b221-4f75-b325-870c84d084ad","order_by":6,"name":"Hélène Vanacker","email":"","orcid":"","institution":"Centre Léon Bérard","correspondingAuthor":false,"prefix":"","firstName":"Hélène","middleName":"","lastName":"Vanacker","suffix":""},{"id":538570702,"identity":"aede4b8c-e5c9-4f16-9999-90a1a1a4d74d","order_by":7,"name":"Loic Verlingue","email":"data:image/png;base64,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","orcid":"","institution":"Centre Léon Bérard","correspondingAuthor":true,"prefix":"","firstName":"Loic","middleName":"","lastName":"Verlingue","suffix":""}],"badges":[],"createdAt":"2025-09-07 13:53:17","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-7556679/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7556679/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":95041296,"identity":"65608b75-3919-4aa9-91de-b7cec0f5e625","added_by":"auto","created_at":"2025-11-03 16:11:07","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":3628658,"visible":true,"origin":"","legend":"","description":"","filename":"DeepOSManuscript.docx","url":"https://assets-eu.researchsquare.com/files/rs-7556679/v1/b7d05a741d3b1a3ac4b65259.docx"},{"id":95041293,"identity":"d2e2f581-ad89-48eb-90cc-288c59ac58b6","added_by":"auto","created_at":"2025-11-03 16:11:07","extension":"json","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":9327,"visible":true,"origin":"","legend":"","description":"","filename":"3d63addc421649458d6b34ade7be8021.json","url":"https://assets-eu.researchsquare.com/files/rs-7556679/v1/ae9ac016fe6d9805d13d9846.json"},{"id":95041340,"identity":"c75cf4eb-a08e-4521-86d8-413fec668939","added_by":"auto","created_at":"2025-11-03 16:11:10","extension":"pptx","order_by":3,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":14237852,"visible":true,"origin":"","legend":"","description":"","filename":"DeepOSFiguresSuppl.pptx","url":"https://assets-eu.researchsquare.com/files/rs-7556679/v1/a622bbe19691a05857c5b9a2.pptx"},{"id":95223001,"identity":"701c152f-e98c-40f0-b43d-0471ec8f104f","added_by":"auto","created_at":"2025-11-05 16:21:30","extension":"docx","order_by":4,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":39632,"visible":true,"origin":"","legend":"","description":"","filename":"DeepOSSuppLegend.docx","url":"https://assets-eu.researchsquare.com/files/rs-7556679/v1/d262c4f94ed7015689656ffd.docx"},{"id":95041300,"identity":"82ec9a26-ca20-4dd6-a537-7b51569ac569","added_by":"auto","created_at":"2025-11-03 16:11:08","extension":"xlsx","order_by":5,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":328586,"visible":true,"origin":"","legend":"","description":"","filename":"DeepOSTablesSuppl.xlsx","url":"https://assets-eu.researchsquare.com/files/rs-7556679/v1/b296b0cfd5415952484108b3.xlsx"},{"id":95041336,"identity":"a53189b0-e741-4aaa-be34-d3b85495e33b","added_by":"auto","created_at":"2025-11-03 16:11:10","extension":"xml","order_by":6,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":123067,"visible":true,"origin":"","legend":"","description":"","filename":"3d63addc421649458d6b34ade7be80211enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-7556679/v1/8ce3dd3dd2197b0ec42185a6.xml"},{"id":95041287,"identity":"95e719c8-34ed-462d-9c6e-2f84e523fd7f","added_by":"auto","created_at":"2025-11-03 16:11:06","extension":"pptx","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":8528432,"visible":true,"origin":"","legend":"","description":"","filename":"DeepOSFiguresMain.pptx","url":"https://assets-eu.researchsquare.com/files/rs-7556679/v1/46198e89f19452b60c6367da.pptx"},{"id":95221740,"identity":"8c3a998b-99dd-4615-b03e-6fc5ab06a0de","added_by":"auto","created_at":"2025-11-05 16:19:38","extension":"xml","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":118905,"visible":true,"origin":"","legend":"","description":"","filename":"3d63addc421649458d6b34ade7be80211structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-7556679/v1/7a82edb7629ba71184066425.xml"},{"id":95041329,"identity":"1f33920a-b086-428e-a2e1-3c74a4d2e7f8","added_by":"auto","created_at":"2025-11-03 16:11:08","extension":"html","order_by":9,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":135197,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-7556679/v1/d9e91fe9d79bc7a41f96513b.html"},{"id":95041299,"identity":"7e91fc98-8fa6-46f3-8532-b2dced2d9b75","added_by":"auto","created_at":"2025-11-03 16:11:07","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":135806,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eGraphical abstract: Pipeline description of DeepOS\u003c/strong\u003e – First, the model is pretrained to predict the organ of origin of both healthy and tumor tissues. Then, the model is fine-tuned on survival of the pan-cancer RNA-seq cohort. DeepOS is a multilayer perceptron neural network model, that uses the RNA-seq expression of 4,499 cancer and immune genes as inputs. DeepOS outputs a probability of survival per time intervals (in the example, one interval represents 72 days). This allows training DeepOS on censored survival data. Survival is estimated by the first interval meeting the probability = 0.5. On the example, a 50% risk of death is predicted to occur at the 14\u003csup\u003eth\u003c/sup\u003e interval, which corresponds to 33 months.\u003c/p\u003e","description":"","filename":"Slide1.png","url":"https://assets-eu.researchsquare.com/files/rs-7556679/v1/42b65d8f8f9d2539c0845871.png"},{"id":95041280,"identity":"b63c358a-3b14-4844-b995-d7b7ebf13bd7","added_by":"auto","created_at":"2025-11-03 16:11:05","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":99851,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003ePan-cancer survival data description\u003c/strong\u003e –\u003cstrong\u003e a\u003c/strong\u003e, Flow-chart of the survival cohort. Pan-cancer RNA-seq and survival and clinical data were retrieved from the TCGA dataset. After selection of the 6,529 samples fulfilling the selection criteria, we used an 80%/10%/10% random split rule to create the training, validation and testing datasets. \u003cstrong\u003eb,c, \u003c/strong\u003eKaplan-Meier survival curves of the whole cohort (\u003cstrong\u003eb\u003c/strong\u003e) and per tumor types (\u003cstrong\u003ec\u003c/strong\u003e). TCGA study abbreviations, median overall survival per tumor type and 95% confidence intervals are detailed in Supplementary Table 3.\u003c/p\u003e","description":"","filename":"Slide2.png","url":"https://assets-eu.researchsquare.com/files/rs-7556679/v1/565e9b837ca070685e816d8e.png"},{"id":95041334,"identity":"70deea23-9c4d-46f4-9dde-4aebdaf66938","added_by":"auto","created_at":"2025-11-03 16:11:10","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":146089,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eSurvival learning curves\u003c/strong\u003e – \u003cstrong\u003ea,b\u003c/strong\u003e, Learning curves for the survival cohort represented by line charts of the median, 1\u003csup\u003est\u003c/sup\u003e and 3\u003csup\u003erd\u003c/sup\u003e quartiles of the C-indexes on the training (blue) and the validation (pink) datasets (\u003cstrong\u003ea\u003c/strong\u003e) and the resulting difference between training and validation median C-indexes (\u003cstrong\u003eb\u003c/strong\u003e) according to the size on the training set (from 500 to 5,529 samples, with steps of 500 samples). C-index = concordance index.\u003c/p\u003e","description":"","filename":"Slide3.png","url":"https://assets-eu.researchsquare.com/files/rs-7556679/v1/d1f0fb0ebf1f6d8f61808a2b.png"},{"id":95221970,"identity":"ab2afeea-08a7-4c9a-8617-3b4f28500fed","added_by":"auto","created_at":"2025-11-05 16:19:59","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":111440,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cem\u003e\u003cstrong\u003eThe organ cohort\u003c/strong\u003e\u003c/em\u003e\u003cstrong\u003e \u003c/strong\u003e– \u003cstrong\u003ea\u003c/strong\u003e, Flow-chart of the organ cohort used to pre-train DeepOS. Healthy organs RNA-seq were obtained from the GTEx project, while pan-cancer RNA-seq data were obtained from the TCGA dataset. Samples were randomly assigned to either the training or the validation cohort with a 90%/10% split. \u003cstrong\u003eb\u003c/strong\u003e, Distribution of organ types across pooled RNA-seq data used for pre-training. The number alongside each feature refers to the number of patients.\u003c/p\u003e","description":"","filename":"Slide4.png","url":"https://assets-eu.researchsquare.com/files/rs-7556679/v1/72029add5976054a31440ec2.png"},{"id":95041281,"identity":"364b2ade-1251-4b72-bb11-a5fc1073a622","added_by":"auto","created_at":"2025-11-03 16:11:05","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":108869,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eDeepOS results\u003c/strong\u003e – \u003cstrong\u003ea\u003c/strong\u003e, Summary of the performances obtained with DeepOS with and without pre-training on organ prediction, DeepSurv and DeepOS with pre-training while using random gene selection. Performances on the training, validation and unseen test set are depicted, based on the same data split. The gene set lists used for DeepOS predictions are detailed in Supplementary Table 2. \u003cstrong\u003eb\u003c/strong\u003e, Line chart of the survival Area Under the ROC Curve (AUC) according to time for DeepOS predictions on the test set. The grey vertical line refers to the mean of all AUC = 0.749. \u003cstrong\u003ec\u003c/strong\u003e, Kaplan-Meier survival curves of OS probability over time, either predicted from DeepOS (green) or observed (blue) within the test cohort. The predicted curve stops shortly after 100 months of OS, which corresponds to the longest OS prediction by DeepOS when analyzing the test cohort. Log-rank p-value indicates the absence of statistical difference between the two Kaplan-Meier curves. C-index: Concordance index; OS: overall survival.\u003c/p\u003e","description":"","filename":"Slide5.png","url":"https://assets-eu.researchsquare.com/files/rs-7556679/v1/47c1c8d9afaf21d02cc5e620.png"},{"id":95041343,"identity":"2163e063-0f1e-4f61-a048-d3d9474e5e4a","added_by":"auto","created_at":"2025-11-03 16:11:11","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":198432,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eDeepOS interpretation\u003c/strong\u003e – \u003cstrong\u003ea\u003c/strong\u003e, Bar chart of the C-indexes of DeepOS according to the tumor types. We considered tumor types within the test cohort with at least three uncensored samples. The red dotted line indicates a C-index of 0.50 (random prediction). The black line indicates a C-index of 0.715, which refers to the median C-index of DeepOS pan-cancer predictions. Patient numbers for each cohort are represented above the bars. \u003cstrong\u003eb\u003c/strong\u003e, The importance of DeepOS input gene is represented by a mirror bar chart of the SHapley Additive exPlanations (SHAP) values. SHAP values for individual predictions are plotted on the left panel. Genes are ranked by mean SHAP values as reported on the right panel. A high positive feature value (pink to the right) means that an increased expression of the gene is related to a reduced OS prediction, whereas a low positive feature value (blue to the right) means that a decreased gene expression is liked to a reduced OS prediction. Each dot represents an individual prediction.\u003c/p\u003e","description":"","filename":"Slide6.png","url":"https://assets-eu.researchsquare.com/files/rs-7556679/v1/40fd077cc92f44be1b773f3c.png"},{"id":105564007,"identity":"7f91502a-8bc2-4f34-b9f1-7a7fcfed8531","added_by":"auto","created_at":"2026-03-27 12:48:26","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1927068,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7556679/v1/97047c43-0dc3-4614-bf5d-1561355e6914.pdf"},{"id":95041327,"identity":"d6c468b4-1b91-4c13-8fbc-fe118c19345f","added_by":"auto","created_at":"2025-11-03 16:11:08","extension":"pptx","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":14237852,"visible":true,"origin":"","legend":"","description":"","filename":"DeepOSFiguresSuppl.pptx","url":"https://assets-eu.researchsquare.com/files/rs-7556679/v1/3d2b672dcaa4109b5adc67f2.pptx"},{"id":95041332,"identity":"c5acd724-ce12-404c-bb32-6d786eb3c877","added_by":"auto","created_at":"2025-11-03 16:11:09","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":39632,"visible":true,"origin":"","legend":"","description":"","filename":"DeepOSSuppLegend.docx","url":"https://assets-eu.researchsquare.com/files/rs-7556679/v1/da0b10fd44036d3847c26a33.docx"},{"id":95041344,"identity":"81269a46-405a-4af9-b21f-7db5e944c075","added_by":"auto","created_at":"2025-11-03 16:11:11","extension":"xlsx","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":328586,"visible":true,"origin":"","legend":"","description":"","filename":"DeepOSTablesSuppl.xlsx","url":"https://assets-eu.researchsquare.com/files/rs-7556679/v1/21b71bb99ba850637273232a.xlsx"}],"financialInterests":"No competing interests reported.","formattedTitle":"DeepOS: pan-cancer prognosis estimation from RNA-sequencing data","fulltext":[{"header":"Introduction","content":"\u003cp\u003eAmong patients diagnosed with cancer, prognosis estimation is often required to draw a risk profile and adapt treatment accordingly. Currently recommended prognostic and predictive biomarkers that drive cancer care management usually combine several items, such as: individual characteristics (e.g. age, gender, ECOG status), tumor characteristics (e.g. tumor stage, localization and number of metastasis), serum markers (e.g. albumin, LDH, CRP) and eventually tumor molecular features (e.g. PD-L1 expression, \u003cem\u003eBRCA\u003c/em\u003e loss-of-function, \u003cem\u003eERBB2, EGFR\u003c/em\u003e, \u003cem\u003eBRAF\u003c/em\u003e, \u003cem\u003eALK\u003c/em\u003e mutations, \u003cem\u003eNTRK\u003c/em\u003e fusion) \u003csup\u003e\u003cspan additionalcitationids=\"CR2 CR3 CR4 CR5\" citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u003c/sup\u003e. Such scoring systems mainly stratify patients into low- or high-risk groups, defining therapeutic procedures to be followed. More recently, the growing interest in high-dimensional multi-omics data in assisting clinicians on treatment decision has brought forward the high potential of tumor RNA-sequencing (RNA-seq) on studying the link between tumor gene expression and patient survival outcome in a personalized way \u003csup\u003e\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e\u003c/sup\u003e. RNA-seq provides gene expression quantifications of the whole transcriptome (transcripts of more than 20,000 protein-encoding genes) or of preselected transcripts of interest (targeted sequencing), bearing the underlying hypothesis that each tumor gene expression profile mirrors the tumor aggressiveness and potential behavior in response to a particular treatment and therefore, should correlate with overall survival (OS).\u003c/p\u003e\u003cp\u003eSince each tumor is unique in its complexity, an almost-infinite number of gene expression combinations could be expected; which drastically overcomplicates any prediction task based on RNA-seq data. Several teams recently intended to predict individual OS from RNA-seq analyses of multiple cancer types obtained from the Cancer Genome Atlas (TCGA) dataset \u003csup\u003e\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e\u003c/sup\u003e, using machine and deep learning \u003csup\u003e\u003cspan additionalcitationids=\"CR10 CR11 CR12 CR13 CR14 CR15 CR16 CR17\" citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e\u003c/sup\u003e (summarized in Supplementary Table\u0026nbsp;1). Model architectures included Random Forest, Cox regression with Lasso penalization, Multilayer Perceptron (MLP), Convolutional Neural Networks, Auto-Encoder with Cox loss function, among others. Those models prediction performances in validation or test cohorts were often limited, close to 0.60, and rarely exceeded 0.62 of median concordance-index (C-index) on pan-cancer predictions. C-index is a popular metric that evaluates the ability of a model to rank survival predictions within a particular cohort rather than the difference between the predicted and the observed values (with C-indexes of 0.50 and 1.00 respectively corresponding to random and perfect order of predictions) \u003csup\u003e\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e,\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e\u003c/sup\u003e. Yet, prognosis estimation from pan-cancer RNA-seq data should be feasible since each tissue and each tumor type express their own transcriptomic signatures.\u003c/p\u003e\u003cp\u003eWith machine and deep learning, over-fitting can arise when the training dataset has greater number of dimensions (variables) than number of samples available (the size of the training set). Over-fitted models generally fail to generalize decently \u003csup\u003e\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e\u003c/sup\u003e. In the case of RNA-seq, the large dimensionality (e.g. \u0026gt;20,000 gene expressions) requires massive amounts of training data, which is tricky to obtain when applied to cancer and which presumably lacked to the above-mentioned models (total number of samples comprised between 953 and 11,854). We hypothesized that OS prediction using supervised deep learning on pan-cancer RNA-seq data as inputs would benefit from (i) starting with reducing input dimensions using prior biological knowledge and (ii) increasing the size of the training set, using a transfer learning strategy. Transfer learning consists in pre-training a model on a task that is related, but not strictly identical to the final question, and for which a larger number of samples is available \u003csup\u003e\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e\u003c/sup\u003e. We therefore (i) filtered whole transcriptome expressions to reduce inputs dimensionality and (ii) designed an MLP neural network that we pre-trained to predict the organ of origin from large healthy and cancer RNA-seq data and fine-tuned to predict OS from pan-cancer RNA-seq data (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). We additionally characterized the parameters that influenced the model performances, including the gene list selection, the size of the training set, the type of survival loss function for training, the duration of OS, the cancer type, and the genes most implicated in the model predictions.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e"},{"header":"Results","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e\u003ch2\u003eGene selection\u003c/h2\u003e\u003cp\u003eWe first intended to reduce the dimensions of our input dataset by selecting genes of interest, which are known to be implicated in cancer initiation, progression, dissemination or response to treatment. We merged gene lists obtained from the Molecular Signatures Database (MSigDB, relative to hallmarks of cancer) \u003csup\u003e\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e\u003c/sup\u003e, and the LM22 immune gene signatures \u003csup\u003e\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e\u003c/sup\u003e. After removal of duplicates and genes associated with no expression values within our dataset, we obtained 4,499 genes (Supplementary Table\u0026nbsp;2).\u003c/p\u003e\u003c/div\u003e\n\u003ch3\u003ePilot overall survival prediction task: without pre-training\u003c/h3\u003e\n\u003cp\u003eWe designed a pilot experiment of survival estimation starting with only tumor RNA-seq data. We retrieved all The Cancer Genome Atlas (TCGA) pan-cancer RNA-seq raw data publicly-available on February, 11, 2019, via recount2 \u003csup\u003e25\u003c/sup\u003e and selected samples associated with annotated survival outcomes (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ea). We excluded uninformative patients who were censored during the first half of the total duration of the follow-up and the top 5% of patients with the longest OS, considering them cured by surgery. Altogether, we collected 6,529 RNA-seq samples from 33 tumor types fulfilling the criteria, among which 54.8% were censored during the second half of the follow-up.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eBased on this dataset, the pan-cancer median OS was 67.2 months (95% confidence interval 95%CI [64.8;72.0]) (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eb), and highly dependent on the tumor type (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ec). Glioblastoma, esophageal cancer, mesothelioma and pancreatic cancers were associated with the worst prognosis (median OS of 12.0 months for glioblastoma and 16.8 months for the three others); on the other hand, median OS was not reached after a 120-months follow-up for five tumor types (chromophobe and papillary renal carcinoma, pheochromocytoma, testicular cancer and thyroid cancer) (Supplementary Table\u0026nbsp;3).\u003c/p\u003e\u003cp\u003eWe randomly assigned each sample to either a training set (N = 5,529), a validation set (N = 500) or a test set for final evaluation (N = 500). Splitting was well-balanced considering the fraction of censored patients (0.546, 0.536 and 0.546, respectively within training, validation and test cohorts), median OS (67.2 months, 72.0 months and 62.4 months, respectively) and diversity of cancer types (Supplementary Table\u0026nbsp;4 and Supplementary Fig.\u0026nbsp;1).\u003c/p\u003e\u003cp\u003eTo train our models, we transformed survival data into survival probabilities per time interval and thus, could implement the classical mean squared error (MSE) loss function. Survival probabilities were set to 1 for intervals during which a patient is alive, and to 0 when a patient is deceased. Censored intervals were ignored to calculate the loss. Ignoring censored intervals in MSE allowed the model to be trained only on observed time intervals for each patient. The model learned a probability of survival for each patient individually, using classical methods in deep learning for multiclass classification. To decipher whether this approach could be competitive, we compared the performances of DeepOS to state-of-the-art deep learning model based on Cox-loss (i.e. DeepSurv and SAVAE-Cox) to train on survival data \u003csup\u003e\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e,\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e\u003cp\u003eWe used a Tree-structured Parzen Estimator (TPE) \u003csup\u003e\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e\u003c/sup\u003e algorithm to explore hyper-parameters and to select the best model upon the highest C-index obtained on the validation set. After training on 5,529 pan-cancer RNA-seq samples, the best model reached a C-index of 0.73 on the validation set, with cross-validation C-index mean of 0.63 and standard deviation of 0.09. The best model had five hidden layers, each of them having a dropout of 0.013, and L1 and L2 penalization of 0.007 and 0.0012, and trained with a learning rate of 0.00003.\u003c/p\u003e\u003cp\u003eOn a final and previously unseen test cohort, this model achieved a C-index of 0.714 on predicting patient survival from their pan-cancer RNA-seq data. This model surpassed DeepSurv performances on the same split data (C-index of 0.606 on validation and test sets) and SAVAE-Cox model (test C-index of 0.696)\u003c/p\u003e\n\u003ch3\u003eLearning curves of survival prediction\u003c/h3\u003e\n\u003cp\u003eTo study how the number of samples within the training set influenced the model performances, we repeated the survival training task with an escalating number of samples composing the training cohort, without modifying the validation set.\u003c/p\u003e\u003cp\u003eLearning curves indicated that C-indexes reach a steady state for training cohorts containing at least 2,000 RNA-seq samples (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003ea). The best training set C-index (0.93) was achieved with 500 samples; although the difference between training- and validation-related C-indexes indicated that the model was subject to overfitting (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eb). Overfitting defines a model that learns too perfectly from a training set so that it fails to generalize adequately on unseen additional data. According to our results and consistently with what was previously described \u003csup\u003e\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e,\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e\u003c/sup\u003e, overfitting tends to be reduced by increasing the number of samples within the training set (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eb). We therefore hypothesized that a transfer learning strategy could benefit our model since it could indirectly expand the training dataset through pre-training on a similar task.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\n\u003ch3\u003eTransfer learning: data collection for the pre-training task\u003c/h3\u003e\n\u003cp\u003eSince OS distribution was related to tumor type (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eb), we assumed that learning to predict the organ of origin from a larger cohort could improve the overall estimation of survival duration. We therefore chose to pre-train our model on the prediction of the organ of origin from the RNA-seq expression data of the 4,499 selected genes using either healthy or tumor tissue (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003ea). Healthy organs data were obtained through recount2 from the Genotype-Tissue Expression (GTEx) project \u003csup\u003e\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e\u003c/sup\u003e. We additionally retrieved all the TCGA gene expression data of tumor samples (including those not associated with survival data). Each tumor type was aligned with its organ of origin (for example, kidney chromophobe, clear cell carcinoma and papillary cell carcinoma were all considered as kidney tissue). To avoid data leakage we removed from the organ dataset the samples corresponding to the patients in the validation and test survival dataset.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eAfter removing 1086 samples from the pooled dataset of 18,571 corresponding to the samples used as validations and tests in the survival dataset, the organ dataset consisted of 17,485 RNA-seq samples from 38 distinct human tissues (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003eb). The most represented organs were brain (8.5% of samples), lung (8.2%) and breast tissue (7.9%). We randomly divided the organ dataset samples into two distinct sets for training (15,485) and validation (2,000). Splitting was well balanced and both sets harbored samples belonging to the 38 types of tissue.\u003c/p\u003e\n\u003ch3\u003ePre-training on organ prediction\u003c/h3\u003e\n\u003cp\u003eWe thus pre-trained DeepOS on organ prediction. DeepOS is an MLP neural network that takes gene expression values in transcripts per million (TPM) as inputs, and outputs either organ classification or survival probabilities (Supplementary Fig.\u0026nbsp;2). DeepOS architecture comprises hidden layers composed of stacked units of dense layers, Rectified Linear Unit (ReLU) activations, dropout effect penalization, L1 and L2 regularization, and batch normalization.\u003c/p\u003e\u003cp\u003eMost of the models tested reached very high validation performances to predict the organ of origin (mean hyper-parameter search accuracy = 0.764 and standard deviation = 0.263), with best model reaching and validation accuracy of 0.9720, precision of 0.9698, recall of 0.9667 and F1-score of 0.9676 (Supplementary Fig.\u0026nbsp;3). The best organ-specific model had four layers, with a dropout rate of 0.014, a L1 and L2 regularization parameters respectively of 0.0010 and 0.0018 and was trained with a learning rate of 0.00019. We did not perform a test set evaluation as this step was only used to select the best pre-trained model to fine-tune on survival.\u003c/p\u003e\u003cdiv id=\"Sec8\" class=\"Section2\"\u003e\u003ch2\u003eFine-tuning on survival prediction\u003c/h2\u003e\u003cp\u003eTo implement our transfer learning strategy, we then fine-tuned DeepOS on survival prediction based on the pan-cancer cohort described above. The part of DeepOS pretrained on organ prediction was frozen during the first fine-tuning step and unfrozen during the second fine-tuning step (Supplementary Fig.\u0026nbsp;2). Similarly to the pilot task performed without pre-training, we selected the best version of the fully trained (comprising pre-training and fine-tuning) model based on the C-index and Kaplan-Meier (KM) statistics obtained on the validation set. The best validation C-index was 0.751 (mean cross-validation C-index = 0.728, standard deviation = 0.007).\u003c/p\u003e\u003cp\u003eOn the unseen test cohort, DeepOS achieved a C-index of 0.714 and an area under the survival ROC curve (AUC) of 0.749 (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003ea and \u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003eb). Transfer learning slightly improved survival prediction of + 7% for the C-index (without pre-training 0.707) and + 0.7% for the mean AUC (without pre-training 0.742). In addition to the organ-specific features, the best model architecture had six layers, with a dropout rate of 0.013 and L1 and L2 regulation parameters respectively of 0.0011 and 0.0003. We have then evaluated if these learned hyper-parameters from DeepOS full training changed the performance of a model trained only on survival (without pre-training on the organ classification problem as in the pilot survival prediction task). We found that the predicted KM curve was closer to the ground truth but the C-index didn’t improve (Supplementary Fig.\u0026nbsp;4a). It suggests that the overfitting of the pilot survival prediction tasks was mostly related to the hyperparameters learnt.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e\n\u003ch3\u003eDeepOS according to patient OS\u003c/h3\u003e\n\u003cp\u003eTo interpret how DeepOS predicts pan-tumor survival, we generated a predicted Kaplan-Meier survival curve and compared it to the true survival curve of the test cohort (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003ec). We noticed that there was no significant difference between DeepOS prediction over time and the ground truth (log-rank p-value: 0.08, superior to 0.05).\u003c/p\u003e\u003cp\u003eWhen applied to the training and validation cohorts, we similarly observed apparent proximity between predicted and observed curves for OS comprised between 35 and 90 months (Supplementary Fig.\u0026nbsp;4), although we repeatedly noticed divergences of the curve slopes for shorter OS, with DeepOS behaving over-optimistic as compared to reality.\u003c/p\u003e\u003cp\u003eTo further evaluate DeepOS performances according to survival duration, we generated ten subgroups of 50 patients ranked by OS on the test cohort, and computed the C-indexes of each subgroup. We could indeed observe that DeepOS performed modestly for survival predictions of patients deceased between 1 and 20 months (C-index \u0026lt; 0.58; \u0026lt;0.55 without pre-training) (Supplementary Fig.\u0026nbsp;5a). This observation was not associated with underrepresentation of such population within the training set (Supplementary Fig.\u0026nbsp;5b).\u003c/p\u003e\n\u003ch3\u003eDeepOS according to tumor type\u003c/h3\u003e\n\u003cp\u003eAmong the 29 cancer types that contained at least 3 uncensored samples, 23(79.3%) displayed a C-index \u0026gt; 0.50 which corresponds to better than random prediction and 8(27.6%) had a C-index \u0026gt; 0.7145 on the test set, including three that reached the perfect score of 1 (Adrenocortical carcinoma, Testicular Germ Cell Tumors and Uveal Melanoma), although these were composed of only 5, 6, and 4 patients within the test set, respectively (Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003ea). Despite our previous observation, DeepOS was able to perform reasonably well on four out the five tumor types displaying a median OS \u0026lt; 20 months, with C-indexes comprised between 0.55 and 0.74 (ESCA, MESO, PAAD, STAD), but the glioblastoma group was ill-predicted, (Supplementary Table\u0026nbsp;5). Transfer learning unchanged or improved survival predictions for 20 (69.0%) tumor types, up to + 28,5%, as assessed by C-index calculation, but achieved worst performance that the pilot survival model for 9 tumor types including frequent ones (COAD, ESCA, GBM, KICH, LUSC, PAAD, READ, SARC, SKCM) (Supplementary Fig.\u0026nbsp;6a, Supplementary Table\u0026nbsp;5).\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cdiv id=\"Sec11\" class=\"Section2\"\u003e\u003ch2\u003eDeepOS few-shot learning per tumor type\u003c/h2\u003e\u003cp\u003eWe have evaluated if few-shot learning could further improve DeepOs performance on frequent tumor types with low C-index such as BRCA, LUAD, LUSC. Few-shot learning of DeepOs improved the test performance for LUSC and BRCA, but decreased the performance for LUAD (Supplementary Fig.\u0026nbsp;6b). In addition, it improved the proximity to the ground-truth survival curve of patients affected by these frequent tumor types (log-rank p-value 0.61, c-index 0.67, Supplementary Fig.\u0026nbsp;6c).\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec12\" class=\"Section2\"\u003e\u003ch2\u003eDeepOS according to gene selection\u003c/h2\u003e\u003cp\u003eWe then evaluated if the gene set selected on prior biological knowledge helped DeepOS performance. We first built a model without gene selection, and followed the same training as with DeepOS, i.e. pre-training on organ prediction and fine-tuning on survival prediction with hyperparameter search. We found good ranking performances but the predicted Kaplan-Meier curve showed very poor relation to the real survival curve (Supplementary Fig.\u0026nbsp;7)\u003c/p\u003e\u003cp\u003eWe then trained models with dimension reduction using 100 random selections of 4,499 genes (examples of gene sets in Supplementary Table\u0026nbsp;2). Using random genes as inputs impaired the model performances, achieving an average C-index of 0.69 and a standard deviation of 0.009 on the test set (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003ea). Predicted Kaplan-Meier survival curves were significantly different to the ground truth with mean p-values of 2.6*10 − 5 and mean standard deviation of 2.5*10 − 4. Gene selection based on prior-knowledge thus contributed significantly to DeepOS generalization performance.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec13\" class=\"Section2\"\u003e\u003ch2\u003eDeepOS according to gene expressions\u003c/h2\u003e\u003cp\u003eTo finally better characterize which genes were the most important for DeepOS predictions, we estimated the mean SHapley Additive exPlanations (SHAP) values. SHAP provides an interpretation of the importance attributed by the algorithm to each input feature \u003csup\u003e\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e\u003cp\u003eAmong the 4,499 genes, 496 had a mean absolute SHAP value \u0026gt; 0.001 and 57 genes had a mean absolute SHAP value \u0026gt; 0.005, which we considered as the most important for DeepOS predictions (Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003eb, Supplementary Fig.\u0026nbsp;8a). We compared the effect of gene expressions importance and direction for the model decision with known biological findings. Gene expression of \u003cem\u003eFOX2A\u003c/em\u003e was the most important feature, with high expression values correlated with poor OS. This is also consistent with the known role of FOX2A, a transcription factor promoting proliferation and epithelial-mesenchymal transition in multiple cancer types \u003csup\u003e\u003cspan additionalcitationids=\"CR33 CR34\" citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e–\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e\u003c/sup\u003e. The second most important gene was \u003cem\u003eCLDN18\u003c/em\u003e, encoding for the Claudin-18 protein involved in tight junctions between epithelial cells and altered in epithelial-mesenchymal transition. Claudin-18.2 is a therapeutic target in gastric cancer and under development in pancreatic cancers. The prognostic role of Claudin-18 is debated and may depend on the tumor type \u003csup\u003e\u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e,\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e\u003c/sup\u003e. Consistently DeepOS used low expression to relate to prognosis and its importance was different in pancreatic and gastric cancers compared to other tumor types (supplementary Fig.\u0026nbsp;8b). An example of a similar relation between expression and prognosis in DeepOS is the SST gene encoding for somatostatin, which also is relevant compared to the literature showing that low expression relates to poor survival in colorectal cancers \u003csup\u003e\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec15\" class=\"Section2\"\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec16\" class=\"Section2\"\u003e\u003cdiv id=\"Sec17\" class=\"Section3\"\u003e\u003c/div\u003e\u003c/div\u003e"},{"header":"Conclusions, discussion","content":"\u003cp\u003eWe have developed a deep learning model, DeepOS, to estimate OS from pan-cancer RNA-seq data with several methodological improvement in the field including (i) prior biological knowledge to condense inputs dimensionality, (ii) transfer learning to enlarge its training capacity through pre-training on organ prediction, and (iii) mean squared error adapted to survival loss function. Although we first observed a relation between the performance of survival prediction with the number of observations to learn from, transfer learning only slightly improved the prediction performances. The pre-training task consisted in predicting the organ of origin, motivated by an important relation between the organ of origin and cancer prognosis. We obtained very high accuracy, precision and recall performances (0.9720, 0.970, 0.967 each on the validation set). For survival prediction, DeepOS reached a median pan-cancer C-index of 0.714 on a previously unseen test set and a survival AUC of 0.75.\u003c/p\u003e\u003cp\u003eDeepOS can output a discrete estimation of individual survival, which enables us to plot Kaplan-Meier survival curves from individual predictions. Doing so, we could confirm that DeepOS survival predictions over time were not different from the ground truth in the test set (log-rank p-value 0.08).\u003c/p\u003e\u003cp\u003eWe also found a way to use the mean square error, a classical loss function in deep learning, to train DeepOS on survival data. We used Cox-loss based models (DeepSurv and SAVAE-Cox) as an internal comparator using the same training, validation and testing data split. DeepSurv is a deep neural network trained with a Cox proportional hazards loss function, which is considered a state-of-the-art method for survival prediction \u003csup\u003e\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e\u003c/sup\u003e. DeepOS significantly outperformed DeepSurv model performances (DeepSurv test C-indexes were between 0.60 and 0.61, and AUC of 0.63 on the test set, and SAVAE-Cox test c-index of 0.69, similarly to other published models for such task; Supplementary Table\u0026nbsp;1).\u003c/p\u003e\u003cp\u003eFinally, we observed that input dimension reduction from prior-knowledge improved the model performance compared to random gene set selection and limited overfitting compared to the largest input possible. This is relevant with other studies using prior-knowledge dimension reduction for other tasks such as treatment effect prediction in cell lines or systems biology in cancer \u003csup\u003e\u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e39\u003c/span\u003e,\u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e\u003cp\u003eOur model has limitations. Firstly, as compared to the survival prediction task ran without pre-training, one could argue that the transfer learning strategy only slightly improved the model performances (+ 0.7% on the test C-index). We found a contrast between the metrics (C-index, survival AUC, log rank test) and the visual interpretation of the Kaplan-Meier curves, supporting the importance of this way to clinically interpret the prediction of models in a cohort of patients. The model without pretraining had a significant decrease in performance for long term predictions (Supplementary Fig.\u0026nbsp;4) compared to DeepOS. The overfitting observed in the pilot survival prediction task without pre-training was in large part due to overfitting of the hyperparameters. Pre-training on organ prediction is a pragmatic choice given the availability of public molecular data with relatively controlled batch effect, and could have helped on long term survival estimation and hyperparameter search. Research on pre-trained fondation survival models including molecular data could help future improvements. Besides, even minor upgrading in performance is challenging to obtain for C-index values above 0.70 for such task. Our study supports a benefice of using large tumor RNA-seq datasets with survival observation.\u003c/p\u003e\u003cp\u003eSecondly, we noted over-optimism in DeepOS predictions for short survival durations (mainly \u0026lt; 20 months). However, it did not negatively impact the model performances within tumor types of unfavorable prognosis. For example, the C-index within the pancreatic adenocarcinoma, esophageal carcinoma, and the stomach adenocarcinoma cohorts reached above 0.55, while median OS was below 20 months (16.8, 16.8, and 19.2 months respectively). A possible explanation is that patients with short survival have poor prognosis factors such as tumor location, poor general condition or comorbidities that are missed by using only RNA-seq data from a microscopic tumor sample. We further explored this hypothesis in another work and indeed demonstrated that short survival is better predicted by clinical variables enclosed in consultations reports \u003csup\u003e\u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e41\u003c/span\u003e\u003c/sup\u003e(submitted to the same collection). Analyzing the influence of time on models performance should be generalized to better define the application framework of deep learning models for such tasks. We also explored few-shot learning to refine the performance on frequent tumor types and showed that improvements were not always possible.\u003c/p\u003e\u003cp\u003e Finally, the pan-cancer TCGA RNA-seq dataset that we used for our study was built mostly upon primary tumor samples obtained from surgical resection of localized or locally-advanced diseases. This is highlighted by the observed median OS for several cancer types longer than expected for metastatic stages at diagnosis. It is nevertheless possible that DeepOS learned from the metastatic potential of the tumor samples. This is supported by the detail of the genes with the highest importance for the algorithm, which are mainly related to cancer progression and epithelial-mesenchymal transition and thus, to cancer dissemination. However, further refinement and validation studies are warranted to statue on the generalizability of the model in metastatic cancers.\u003c/p\u003e\u003cp\u003eOther approaches have proposed clustering analysis from RNA-seq to identify groups of patients with similar prognosis \u003csup\u003e12 42\u003c/sup\u003e. \u003cem\u003eThorsson et al.\u003c/em\u003e could identify six immune subtype features from TCGA pan-cancer data comprising RNA-seq, miRNA-seq and exome sequencing data \u003csup\u003e\u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e42\u003c/span\u003e\u003c/sup\u003e. They rigorously characterized immune subtypes associated with good and poor prognosis, although pan-cancer performances were modest with a median pan-cancer C-index visually lower than 0.60. In addition to transfer learning, DeepOS comprised methodological adaptations that we believe have permitted this upgrading. First, we transformed survival data into time interval survival probabilities, so that censored time intervals did not influence the loss function calculation. Thus, we could train on the mean squared error. We also reduced input dimensions by applying prior knowledge on the biology of cancer and immunity to limit overfitting due to irrelevant genes for our task, which contributed to improve the model predictions.\u003c/p\u003e\u003cp\u003eOverall, our study demonstrated and/or validated that (i) predicting survival outcomes from pan-cancer RNA-seq data is feasible and can achieve decent performances, (ii) dimension reduction based on prior knowledge improve the performance, and (iii) partially censored survival data can be used to train supervised deep learning models with standard loss functions. DeepOS offers a promising proof-of-concept that prognosis estimation among patients affected with various types of cancer can be personalized beyond classical score calculations. It provides a more tumor-centered way to estimate the disease aggressiveness and perhaps, to estimate its sensitivity to multiple therapeutic options.\u003c/p\u003e"},{"header":"Methods","content":"\u003cp\u003e\u003cstrong\u003eObjectives -\u003c/strong\u003e This study aimed at predicting the survival of patients affected by various tumor types from their gene expression analysis. This is a classical task with gold-standard datasets that we used to evaluate methodological improvements (Supplementary Table 1). We have developed a new format of survival data to train deep learning models, a prior-knowledge based dimension reduction and a transfer learning strategy. We hypothesized that these methods should help model performance and interpretability.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eLabels \u0026ndash; Survival\u003c/strong\u003e - We used the publicly available survival data of the TCGA database from \u003cem\u003eLiu et al.\u003c/em\u003e \u003csup\u003e\u003cspan class=\"CitationRef\"\u003e43\u003c/span\u003e\u003c/sup\u003e. The top 5% of patients with the highest overall survival were removed because they were considered cured (by surgery, as their overall survival was higher than nine years). Patients with no follow-up were also removed (i.e. 0 days or survival status not known). Early censored patients had poor relevance for the training; we thus removed patients censored before the median follow-up of the cohort. We then performed a random split of the data (80%, 10%, 10%).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eLabels \u0026ndash; Organ -\u003c/strong\u003e For the pre-training on organ prediction, we have pooled GTEx and TCGA data \u003csup\u003e\u003cspan class=\"CitationRef\"\u003e8\u003c/span\u003e,\u003cspan class=\"CitationRef\"\u003e30\u003c/span\u003e\u003c/sup\u003e. GTEx concerned the analysis of normal organs and TCGA the analysis of primary tumors classified by organs of origin. A random split (90%, 10%) was performed on the organ data set (no test cohort was required as the organ data were used for the pre-training task). To avoid data leakage from the test set into the training set, we made sure to exclude from the organ dataset any patient that would be present in the validation set and test set from the survival dataset.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eInput \u0026ndash; RNA sequencing -\u003c/strong\u003e Inputs used to feed DeepOS were gene expression values estimated from RNA-seq. RNA-seq was the most frequent analysis commonly performed in both GTEx and TCGA and allowed to gather a maximum of examples matched with the labels described above. RNA-seq is a multistep process. RNA is first extracted from the tissue sample and sequenced. For TCGA, a vast majority of primary tumor samples came from surgical interventions while for GTEx, it came from non-diseased tissue samples from human donors. Gene expression is then estimated by the number of RNA fragments corresponding to a genome locus from a sequenced sample. TCGA and GTEx gene expressions were analyzed with the same bioinformatic pipeline from raw sequencing data and available in Recount2 \u003csup\u003e25\u003c/sup\u003e. Gene expression was estimated in TPM (transcripts per million) with the Rail-RNA pipeline \u003csup\u003e\u003cspan class=\"CitationRef\"\u003e44\u003c/span\u003e\u003c/sup\u003e. TPM followed a Poisson distribution, so we log-transformed and scaled the data matrix using natural logarithm.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eInput \u0026ndash; Dimension reduction on prior-knowledge\u003c/strong\u003e \u003cstrong\u003e-\u003c/strong\u003e RNA-seq gene expression data is usually highly dimensional (~\u0026thinsp;23k protein coding genes plus non-coding regions) which can be a source of overfitting during the learning step of deep neural networks \u003csup\u003e\u003cspan class=\"CitationRef\"\u003e28\u003c/span\u003e\u003c/sup\u003e. To reduce the dimensions of input data, we selected important cancer-related genes based on prior-knowledge. MSigDB database \u003csup\u003e\u003cspan class=\"CitationRef\"\u003e23\u003c/span\u003e\u003c/sup\u003e provided gene lists related to cancer hallmark and LM22 provided immune cell line specific gene lists \u003csup\u003e\u003cspan class=\"CitationRef\"\u003e24\u003c/span\u003e\u003c/sup\u003e that are important mechanisms for cancer evolution. These two sources comprised a total of 4,499 genes also found in GTEx and TCGA RNA-seq data.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eInput \u0026ndash; study of the gene selection\u003c/strong\u003e - For comparison, we trained models with random selections of 4,499 input genes, excluding the ones found from cancer hallmarks and LM22. We trained those models on the same RNA-seq data, using the same workflow (hyper-parameter search and selection of the best model on validation C-index). We replicated the experiment twice, each time with different selections of random genes (#1 and #2).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eModels\u0026rsquo; architecture -\u003c/strong\u003e DeepOS model is a multilayers perceptron (MLP), which consists of at least three types of layers: the input layer, hidden layers and the output layer. Except for the input data, each unit uses a linear function using parameters W and b, activated by a nonlinear function such as ReLU used here for the hidden layers. Training was supervised using the backpropagation of the gradients of the error to improve model predictions, step by step, by correcting the parameters W and b. The last layer of our model was composed of linear functions.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eLoss \u0026ndash; Survival loss -\u003c/strong\u003e Patient survival in TCGA was calculated by the number of days to death (the event of interest) since the date of sampling. Censored patients were patients that were still alive (have not presented the event of interest) at the time of end of follow-up. Patients with good outcomes are thus more prone to be censored. Removing censored patients would influence the model to be over-pessimistic and would decrease the number of examples for training. Keeping censored patients leads to challenges in the design of a loss function to minimize. We have developed and implemented an approach to train deep learning models on survival data. With this approach, follow up was divided into a vector of B time-bins (or time intervals). In the raw data, each day of the follow up was associated with one value: 1 if the patient is alive, 0 if he is deceased and \u0026minus;\u0026thinsp;1 if he is censored. The value of a bin was the mean of the values of each day included in this bin. The bin value ranged from \u0026minus;\u0026thinsp;1 to 1.\u003c/p\u003e\n\u003cp\u003eFor example, the bin values corresponding to a time interval of 5 days for a patient deceased at day 4 are the following:\u003c/p\u003e\n\u003cp\u003eIn days: [1, 1, 1, 0, 0, 0, 0, 0, 0, 0] \u0026rArr; bin values: [0.6, 0]\u003c/p\u003e\n\u003cp\u003eConcerning a patient censored at day 3 the bin values are:\u003c/p\u003e\n\u003cp\u003eIn days: [1, 1, -1, -1, -1, -1, -1, -1, -1, -1] \u0026rArr; bin values: [-0.2, -1]\u003c/p\u003e\n\u003cp\u003eWe used the MSE, a classical loss function used to backpropagate the error of deep learning models (examples in Supplementary Table\u0026nbsp;6). The MSE is given by:\u003c/p\u003e\n\u003cdiv id=\"Equa\" class=\"Equation\"\u003e\n \u003cdiv class=\"mathdisplay\" id=\"FileID_Equa\" name=\"EquationSource\"\u003e$$\\:MSE=\\:\\frac{1}{m}{\\sum\\:}_{i=1}^{n}({y\\_true}_{i}-{{y}_{pred}}_{i})\u0026sup2;$$\u003c/div\u003e\n\u003c/div\u003e\n\u003cp\u003ewith m the number of non-censored patients and n the number of patients.\u003c/p\u003e\n\u003cp\u003eConsequently, the model output layer was designed as a vector of survival probabilities over time with the number of neurons corresponding to the number of bins. Censored values were ignored in the computation of MSE and doing so, the model was trained only on the observed follow up. The cutoff probability value for the model to predict time to death was set to 0.5 and first bin with value less than 0.5 was considered.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eLoss \u0026ndash; Cox loss -\u003c/strong\u003e Most of the previous studies predicting survival from RNA-seq used Cox proportional hazard model to handle censored survival data. As a control, we trained a model with DeepSurv, a MLP with a Cox log-likelihood loss function \u003csup\u003e\u003cspan class=\"CitationRef\"\u003e26\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eLoss \u0026ndash; For organ prediction\u003c/strong\u003e - Categorical cross entropy was used for the organ prediction task which consisted in 38 classes.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTraining \u0026ndash; Penalization and learning -\u003c/strong\u003e Penalization comprises a set of classical methods to prevent overfitting during training, such as L1 and L2 regularization and dropout. Another method proposed to limit overfitting consists in adding Gaussian noise to the input data for each epoch during the training step \u003csup\u003e\u003cspan class=\"CitationRef\"\u003e45\u003c/span\u003e\u003c/sup\u003e. Adam optimizer and batch normalization were also used to accelerate convergence \u003csup\u003e\u003cspan class=\"CitationRef\"\u003e46\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTraining \u0026ndash; Hyper-parameters optimization -\u003c/strong\u003e Hyper-parameters are parameters controlling the MLP architecture, learning strategies and/or penalization of the learning. We have optimized the following hyper-parameters:\u003c/p\u003e\n\u003cp\u003e- The number of layers in the MLP;\u003c/p\u003e\n\u003cp\u003e- The number of nodes of the first hidden layer;\u003c/p\u003e\n\u003cp\u003e- The decrease rate of the number of unit per layer (rate by which the number of nodes of the previous layer is multiplied to determine the number of nodes of the current layer);\u003c/p\u003e\n\u003cp\u003e- The learning rates lr1 (for organ prediction task) and/or lr2 (for survival prediction task);\u003c/p\u003e\n\u003cp\u003e- The regularization parameters:\u003c/p\u003e\n\u003cp\u003e- The standard deviation of the gaussian noise added to input data;\u003c/p\u003e\n\u003cp\u003e- The dropout rate (continuous values within [0, 0.8]);\u003c/p\u003e\n\u003cp\u003e- Lambda values for L1 and L2 normalization;\u003c/p\u003e\n\u003cp\u003e- The batch size;\u003c/p\u003e\n\u003cp\u003e- The number of epochs of learning.\u003c/p\u003e\n\u003cp\u003eConsidering the two training tasks (organ and survival), the hyper-parameters search space had 24 dimensions. We used the Tree-structured Parzen Estimator (TPE) algorithm to train DeepOS hyper-parameters \u003csup\u003e\u003cspan class=\"CitationRef\"\u003e27\u003c/span\u003e\u003c/sup\u003e. TPE is a Bayesian approach that outperformed the traditional grid search and random search on hyper-parameters search. For each new set of hyper-parameters a new random model was fully trained. Performance metrics were calculated on the validation set(s) (for organ and/or survival). New hyper-parameters were inferred from the validation performance by the TPE algorithm. We performed 500 trials for hyper-parameters search, based on previous studies \u003csup\u003e\u003cspan class=\"CitationRef\"\u003e47\u003c/span\u003e\u003c/sup\u003e. The model with the best performance on the validation set was finally evaluated on the test set.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTransfer learning strategy -\u003c/strong\u003e The transfer learning strategy for DeepOS was composed of pre-training on organ prediction and fine-tuning on survival prediction, each of these steps with independent hyper-parameters search. We used validation accuracy to select the best model on organ prediction. We then added new layers (number defined by the hyper-parameter search) and an output layer to this model. We froze the organ layers for the first fine-tuning step on survival (including hyper-parameter search), considering it as a low abstraction representation of gene expression. A second fine-tuning step (including hyper-parameter search) was performed on the same MLP with all layers unfrozen. The final survival model selection was based on the validation cohort.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEvaluation of the model - Metric for organ prediction -\u003c/strong\u003e To evaluate the performance of the model on the organ task, we used classification metrics: accuracy, precision and F1 score.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEvaluation of the model - Metric for survival prediction -\u003c/strong\u003e We used the concordance correlation coefficient (concordance index, or C-index) to evaluate survival models with censored data \u003csup\u003e\u003cspan class=\"CitationRef\"\u003e19\u003c/span\u003e,\u003cspan class=\"CitationRef\"\u003e20\u003c/span\u003e\u003c/sup\u003e. C-index represents the proportion of concordant pairs divided by the total number of possible evaluation pairs. For example, if a patient A has died at time tA and a patient B has been censored at time tB, they can still be compared if tA\u0026thinsp;\u0026lt;\u0026thinsp;tB. If the model gives a prediction pA for patient A and pB for patient B, the pair can be qualified as concordant if pA\u0026thinsp;\u0026lt;\u0026thinsp;pB and non-concordant otherwise. If tA\u0026thinsp;\u0026gt;\u0026thinsp;tB then it is not possible to evaluate this pair and it will not count as a possible evaluation pair.\u003c/p\u003e\n\u003cp\u003eWe also computed the survival AUROC using sklearn (sksurv.metrics.cumulative_dynamic_auc), which is a cumulative area under the ROC curve adapted to censored data \u003csup\u003e\u003cspan class=\"CitationRef\"\u003e48\u003c/span\u003e\u003c/sup\u003e. Finally, we used the p-value of the log-rank test to compare the predicted Kaplan-Meier survival curve to the ground truth. The log-rank test determines if two survival curves are statistically equivalent (null hypothesis) with a chi2 test. The p-value gives indication on whether we should reject the null hypothesis: the smaller it is the more two survival curves are different. Conversely, neural networks trained with Cox loss predict a risk and are barely used to predict individual survival in time; therefore log rank has not been used to date in this setting, to our knowledge.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEvaluation of the model - Performances by survival time \u0026ndash;\u003c/strong\u003e To further evaluate DeepOS predictions, we have assessed the performance depending on survival time. We have sorted the test cohort by OS and divided the cohort into 10 subgroups, each group composed of 50 patients. We have then computed the C-index of each subgroup.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEvaluation of the model - Learning curves -\u003c/strong\u003e In order to evaluate the effect of the training set size on the model performances for survival prediction, we have generated learning curves. We used the validation cohort of 500 patients, given by the data split described previously. For the training set, data were iteratively and randomly added, from 500 to 5,529 samples, with steps of 500 samples. Every time a new training and hyper-parameters search was launched with several sampling procedures. The C-indexes were computed for the training and validation sets with plots for the median, the first and the third quartile.\u003c/p\u003e\n\u003ch2\u003eEvaluation of the model \u0026ndash; comparison to Cox-loss models\u003c/h2\u003e\n\u003ch2\u003eDeepSurv model\u003c/h2\u003e\n\u003cp\u003eWe trained, validated and tested DeepSurv model architecture on the same survival data split and its performances were compared to DeepOS \u003csup\u003e\u003cspan class=\"CitationRef\"\u003e26\u003c/span\u003e\u003c/sup\u003e. We used the same hyper-parameter search strategy based on the TPE algorithm. We did not perform pre-training on organ prediction with DeepSurv because of incompatibility with the Cox-loss, and because the objective was to compare our model to the existing literature.\u003c/p\u003e\n\u003ch2\u003eSAEVAE-Cox model\u003c/h2\u003e\n\u003cp\u003eWe trained SAVAE-Cox model (Self Attention Variational AutoEncoder - Cox) in two steps (Supplementary Fig.\u0026nbsp;9). A pre-training step used an auto-encoder with self-attention. The auto-encoder architecture has a bottleneck shape for latent space representation with an encoder part that reduces dimension and a decoder part. The training consisted in outputting the input RNA-seq using a discriminator model, for 300 epochs on the train set. We fine-tuned the pre-trained encoder part with the Cox-loss on the survival data for 30 epochs.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eModel interpretability \u0026ndash;\u003c/strong\u003e While still an active research field, some techniques allow interpreting MLP training. SHAP values were used, a model agnostic technique that quantifies the influence of each input on the model\u0026rsquo;s predictions \u003csup\u003e\u003cspan class=\"CitationRef\"\u003e31\u003c/span\u003e\u003c/sup\u003e. SHAP values give an input-output correlation mixed with feature importance.\u003c/p\u003e\n\u003ch2\u003eCode and libraries\u003c/h2\u003e\n\u003cp\u003eTo load and process the GTEx and TCGA data we have used the R package recount2 \u003csup\u003e21\u003c/sup\u003e. We have used python 3 with Keras 2.2.5 and Tensorflow 1.14, to build and train the model. Hyper-parameter search with Tree-structured Parzen Estimator (TPE) was performed with the Optuna library \u003csup\u003e\u003cspan class=\"CitationRef\"\u003e49\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eCode, model and data availability\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe code to load and preprocess the data, together with the code to build, train and test the model is publically available on www.github.com/DITEP/DeepOS. The preprocessed data, ready to be inputted in the model, is also publically available for maximum transparency. We provided Jupiter notebooks to navigate intuitively through the steps of the analysis with results and figures included. The user that would want to run the analysis may have slightly different results as few steps are randomized (weight initialization and hyper-parameter search for example). DeepOS model trained and presented in this paper is also provided under Keras hd5 format.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgments\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors are grateful to their colleagues and collaborators for their advices and support for this work and specifically: Rebecca Clodion, Roger Sun, Eric Angevin, Antoine Hollebecque, Daniel Gautheret, Stefan Michiels, Fabrice Andr\u0026eacute;, Andrei Zinovyev, Laurence Calzone, Emmanuel Barillot, Jean-Yves Blay, Jean-Charles Soria, Pierre Saintigny. The results shown here are based upon data generated by the TCGA Research Network: https://www.cancer.gov/tcga.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eConception and design: MP, LR, HV, LV\u003c/p\u003e\n\u003cp\u003eDevelopment of methodology: MP, LR, LV, CT\u003c/p\u003e\n\u003cp\u003eAcquisition, analysis and/or interpretation: MP, LR, DM, LV, CT\u003c/p\u003e\n\u003cp\u003eWriting, review and/or revision of the manuscript: all authors\u003c/p\u003e\n\u003cp\u003eSupervision: LV\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting Interests statement\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eLV reports personal fees from Adaptherapy, non-personal fees from Pierre-Fabre and Servier, grants from Bristol-Myers Squibb, all outside the submitted work. As part of the Drug Development Department (DITEP) of Gustave Roussy (France), LV, ED and CM report being: Principal/sub-Investigator of Clinical Trials for \u0026nbsp;Abbvie, Adaptimmune, Aduro Biotech, Agios Pharmaceuticals, Amgen, Argen-X Bvba, Arno Therapeutics, Astex Pharmaceuticals, Astra Zeneca Ab, Aveo, Basilea Pharmaceutica International Ltd, Bayer Healthcare Ag, Bbb Technologies Bv, Beigene, Blueprint Medicines, Boehringer Ingelheim, Boston Pharmaceuticals, Bristol Myers Squibb, Ca, Celgene Corporation, Chugai Pharmaceutical Co, Clovis Oncology, Cullinan-Apollo, Daiichi Sankyo, Debiopharm, Eisai, Eisai Limited, Eli Lilly, Exelixis, Faron Pharmaceuticals Ltd, Forma Tharapeutics, Gamamabs, Genentech, Glaxosmithkline, H3 Biomedicine, Hoffmann La Roche Ag, Imcheck Therapeutics, Innate Pharma, Institut De Recherche Pierre Fabre, Iris Servier, Janssen Cilag, Janssen Research Foundation, Kura Oncology, Kyowa Kirin Pharm. Dev, Lilly France, Loxo Oncology, Lytix Biopharma As, Medimmune, Menarini Ricerche, Merck Sharp \u0026amp; Dohme Chibret, Merrimack Pharmaceuticals, Merus, Millennium Pharmaceuticals, Molecular Partners Ag, Nanobiotix, Nektar Therapeutics, Novartis Pharma, Octimet Oncology Nv, Oncoethix, Oncopeptides, Orion Pharma, Ose Pharma, Pfizer, Pharma Mar, Pierre Fabre, Medicament, Roche, Sanofi Aventis, Seattle Genetics, Sotio A.S, Syros Pharmaceuticals, Taiho Pharma, Tesaro, Xencor. Research Grants from Astrazeneca, BMS, Boehringer Ingelheim, Janssen Cilag, Merck, Novartis, Onxeo, Pfizer, Roche, Sanofi. Non-financial support (drug supplied) from Astrazeneca, Bayer, BMS, Boringher Ingelheim, Medimmune, Merck, NH TherAGuiX, Onxeo, Pfizer. ED reports grants and personal fees from Roche Genentech, grants from Boehringer, grants from Astrazeneca, grants and personal fees from Merck Serono, grants from BMS, and grants from MSD Roche.\u003c/p\u003e\n\u003cp\u003eOther authors have no conflict of interest to declare.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis project has been funded in part and realized in partnership with ARC foundation for cancer research: Fondation ARC pour la recherche clinique \u0026ndash; 9 rue Guy M\u0026ocirc;quet 94803 Villejuif \u0026ndash; France. Grant number SIGNIT201901302.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003ePrevious presentation\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAn intermediate version of this work has been presented at ESMO 2019 congress under the reference: Abstract 165P - Enhanced performance of prognostic estimation from TCGA RNAseq data using transfer learning. H Vanacker, E Angevin, A Hollebecque, R Sun, E Deutsch, A Zinovyev, L Calzone, E Barillot, C Massard, L Verlingue. Annals of Oncology, Volume 30, Issue Supplement_5, October 2019\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eCardoso, F. \u003cem\u003eet al.\u003c/em\u003e Early breast cancer: ESMO Clinical Practice Guidelines for diagnosis, treatment and follow-up. \u003cem\u003eAnnals of Oncology\u003c/em\u003e \u003cstrong\u003e30\u003c/strong\u003e, 1194\u0026ndash;1220 (2019).\u003c/li\u003e\n\u003cli\u003eEscudier, B. \u003cem\u003eet al.\u003c/em\u003e Renal cell carcinoma: ESMO Clinical Practice Guidelines for diagnosis, treatment and follow-up\u0026dagger;. \u003cem\u003eAnnals of Oncology\u003c/em\u003e \u003cstrong\u003e30\u003c/strong\u003e, 706\u0026ndash;720 (2019).\u003c/li\u003e\n\u003cli\u003eGlynne-Jones, R. \u003cem\u003eet al.\u003c/em\u003e Rectal cancer: ESMO Clinical Practice Guidelines for diagnosis, treatment and follow-up\u0026dagger;. \u003cem\u003eAnnals of Oncology\u003c/em\u003e \u003cstrong\u003e28\u003c/strong\u003e, iv22\u0026ndash;iv40 (2017).\u003c/li\u003e\n\u003cli\u003eMichielin, O., van Akkooi, A. C. J., Ascierto, P. A., Dummer, R. \u0026amp; Keilholz, U. Cutaneous melanoma: ESMO Clinical Practice Guidelines for diagnosis, treatment and follow-up. \u003cem\u003eAnnals of Oncology\u003c/em\u003e \u003cstrong\u003e30\u003c/strong\u003e, 1884\u0026ndash;1901 (2019).\u003c/li\u003e\n\u003cli\u003eColorectal Cancer Screening - NCCN Clinical Practice Guidelines in Oncology V2.2020. (2020).\u003c/li\u003e\n\u003cli\u003eLung Cancer Screening - NCCN Clinical Practice Guidelines in Oncology V1.2021. (2020).\u003c/li\u003e\n\u003cli\u003eKuksin, M. \u003cem\u003eet al.\u003c/em\u003e Applications of single-cell and bulk RNA sequencing in onco-immunology. \u003cem\u003eEuropean Journal of Cancer\u003c/em\u003e \u003cstrong\u003e149\u003c/strong\u003e, 193\u0026ndash;210 (2021).\u003c/li\u003e\n\u003cli\u003ehttps://www.cancer.gov/tcga.\u003c/li\u003e\n\u003cli\u003eYuan, Y. \u003cem\u003eet al.\u003c/em\u003e Assessing the clinical utility of cancer genomic and proteomic data across tumor types. \u003cem\u003eNature Biotechnology\u003c/em\u003e \u003cstrong\u003e32\u003c/strong\u003e, 644\u0026ndash;652 (2014).\u003c/li\u003e\n\u003cli\u003eZheng, X., Amos, C. I. \u0026amp; Frost, H. R. Comparison of pathway and gene-level models for cancer prognosis prediction. \u003cem\u003eBMC Bioinformatics\u003c/em\u003e \u003cstrong\u003e21\u003c/strong\u003e, 76 (2020).\u003c/li\u003e\n\u003cli\u003eChing, T., Zhu, X. \u0026amp; Garmire, L. Cox-nnet: An artificial neural network method for prognosis prediction of high-throughput omics data. \u003cem\u003ePLoS Comput Biol\u003c/em\u003e \u003cstrong\u003e14\u003c/strong\u003e, e1006076\u0026ndash;e1006076 (2018).\u003c/li\u003e\n\u003cli\u003eRamazzotti, D., Lal, A., Wang, B., Batzoglou, S. \u0026amp; Sidow, A. Multi-omic tumor data reveal diversity of molecular mechanisms that correlate with survival. \u003cem\u003eNature Communications\u003c/em\u003e \u003cstrong\u003e9\u003c/strong\u003e, 4453 (2018).\u003c/li\u003e\n\u003cli\u003eCheerla, A. \u0026amp; Gevaert, O. Deep learning with multimodal representation for pancancer prognosis prediction. \u003cem\u003eBioinformatics\u003c/em\u003e \u003cstrong\u003e35\u003c/strong\u003e, i446\u0026ndash;i454 (2019).\u003c/li\u003e\n\u003cli\u003eHuang, Z. \u003cem\u003eet al.\u003c/em\u003e Deep learning-based cancer survival prognosis from RNA-seq data: approaches and evaluations. \u003cem\u003eBMC Medical Genomics\u003c/em\u003e \u003cstrong\u003e13\u003c/strong\u003e, 41 (2020).\u003c/li\u003e\n\u003cli\u003eWang, Z. \u0026amp; Sun, J. SurvTRACE: Transformers for Survival Analysis with Competing Events. in 1\u0026ndash;9 (2022). doi:10.1145/3535508.3545521.\u003c/li\u003e\n\u003cli\u003eZhao, Y. \u003cem\u003eet al.\u003c/em\u003e BERTSurv: BERT-Based Survival Models for Predicting Outcomes of Trauma Patients. \u003cem\u003earXiv.org\u003c/em\u003e https://arxiv.org/abs/2103.10928v1 (2021).\u003c/li\u003e\n\u003cli\u003eHu, S., Fridgeirsson, E., Wingen, G. van \u0026amp; Welling, M. Transformer-Based Deep Survival Analysis. in \u003cem\u003eProceedings of AAAI Spring Symposium on Survival Prediction - Algorithms, Challenges, and Applications 2021\u003c/em\u003e 132\u0026ndash;148 (PMLR, 2021).\u003c/li\u003e\n\u003cli\u003eMeng, X. \u003cem\u003eet al.\u003c/em\u003e A Novel Attention-Mechanism Based Cox Survival Model by Exploiting Pan-Cancer Empirical Genomic Information. \u003cem\u003eCells\u003c/em\u003e \u003cstrong\u003e11\u003c/strong\u003e, 1421 (2022).\u003c/li\u003e\n\u003cli\u003eHarrell, F. E., Califf, R. M., Pryor, D. B., Lee, K. L. \u0026amp; Rosati, R. A. Evaluating the yield of medical tests. \u003cem\u003eJAMA\u003c/em\u003e \u003cstrong\u003e247\u003c/strong\u003e, 2543\u0026ndash;2546 (1982).\u003c/li\u003e\n\u003cli\u003eUno, H., Cai, T., Pencina, M. J., D\u0026rsquo;Agostino, R. B. \u0026amp; Wei, L. J. On the C-statistics for evaluating overall adequacy of risk prediction procedures with censored survival data. \u003cem\u003eStat Med\u003c/em\u003e \u003cstrong\u003e30\u003c/strong\u003e, 1105\u0026ndash;1117 (2011).\u003c/li\u003e\n\u003cli\u003eTeschendorff, A. E. Avoiding common pitfalls in machine learning omic data science. \u003cem\u003eNat. Mater.\u003c/em\u003e \u003cstrong\u003e18\u003c/strong\u003e, 422\u0026ndash;427 (2019).\u003c/li\u003e\n\u003cli\u003eRaffel, C. \u003cem\u003eet al.\u003c/em\u003e Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer. \u003cem\u003earXiv:1910.10683 [cs, stat]\u003c/em\u003e (2020).\u003c/li\u003e\n\u003cli\u003eLiberzon, A. \u003cem\u003eet al.\u003c/em\u003e The Molecular Signatures Database (MSigDB) hallmark gene set collection. \u003cem\u003eCell Syst\u003c/em\u003e \u003cstrong\u003e1\u003c/strong\u003e, 417\u0026ndash;425 (2015).\u003c/li\u003e\n\u003cli\u003eNewman, A. M. \u003cem\u003eet al.\u003c/em\u003e Robust enumeration of cell subsets from tissue expression profiles. \u003cem\u003eNature methods\u003c/em\u003e \u003cstrong\u003e12\u003c/strong\u003e, 453\u0026ndash;7 (2015).\u003c/li\u003e\n\u003cli\u003eCollado-Torres, L. \u003cem\u003eet al.\u003c/em\u003e Reproducible RNA-seq analysis using recount2. \u003cem\u003eNat Biotechnol\u003c/em\u003e \u003cstrong\u003e35\u003c/strong\u003e, 319\u0026ndash;321 (2017).\u003c/li\u003e\n\u003cli\u003eKatzman, J. \u003cem\u003eet al.\u003c/em\u003e DeepSurv: Personalized Treatment Recommender System Using A Cox Proportional Hazards Deep Neural Network. \u003cem\u003eBMC Med Res Methodol\u003c/em\u003e \u003cstrong\u003e18\u003c/strong\u003e, 24 (2018).\u003c/li\u003e\n\u003cli\u003eBergstra, J., Bardenet, R., Bengio, Y. \u0026amp; K\u0026eacute;gl, B. Algorithms for Hyper-Parameter Optimization. \u003cem\u003eAdvances in Neural Information Processing Systems\u003c/em\u003e \u003cstrong\u003e24\u003c/strong\u003e, 2546\u0026ndash;2554 (2011).\u003c/li\u003e\n\u003cli\u003eEveritt, B. S. \u0026amp; Skrondal, A. \u003cem\u003eThe Cambridge Dictionary of Statistics, Fourth Edition\u003c/em\u003e. (2011).\u003c/li\u003e\n\u003cli\u003eBanko, M. \u0026amp; Brill, E. Scaling to Very Very Large Corpora for Natural Language Disambiguation. in \u003cem\u003eProceedings of the 39th Annual Meeting of the Association for Computational Linguistics\u003c/em\u003e 26\u0026ndash;33 (Association for Computational Linguistics, Toulouse, France, 2001). doi:10.3115/1073012.1073017.\u003c/li\u003e\n\u003cli\u003ehttps://gtexportal.org/home/.\u003c/li\u003e\n\u003cli\u003eLundberg, S. M. \u0026amp; Lee, S.-I. A Unified Approach to Interpreting Model Predictions. \u003cem\u003eAdvances in Neural Information Processing Systems\u003c/em\u003e \u003cstrong\u003e30\u003c/strong\u003e, 4765\u0026ndash;4774 (2017).\u003c/li\u003e\n\u003cli\u003eBhat, A. A. \u003cem\u003eet al.\u003c/em\u003e Tight Junction Proteins and Signaling Pathways in Cancer and Inflammation: A Functional Crosstalk. \u003cem\u003eFront. Physiol.\u003c/em\u003e \u003cstrong\u003e9\u003c/strong\u003e, (2019).\u003c/li\u003e\n\u003cli\u003eGr\u0026ouml;ne, J. \u003cem\u003eet al.\u003c/em\u003e Differential expression of genes encoding tight junction proteins in colorectal cancer: frequent dysregulation of claudin-1, -8 and -12. \u003cem\u003eInt J Colorectal Dis\u003c/em\u003e \u003cstrong\u003e22\u003c/strong\u003e, 651\u0026ndash;659 (2007).\u003c/li\u003e\n\u003cli\u003eKim, S. S. \u003cem\u003eet al.\u003c/em\u003e Immunohistochemical stain for cytokeratin 7, S100A1 and claudin 8 is valuable in differential diagnosis of chromophobe renal cell carcinoma from renal oncocytoma. \u003cem\u003eHistopathology\u003c/em\u003e \u003cstrong\u003e54\u003c/strong\u003e, 633\u0026ndash;635 (2009).\u003c/li\u003e\n\u003cli\u003ehttps://www.proteinatlas.org/ENSG00000156284-CLDN8/pathology/renal+cancer.\u003c/li\u003e\n\u003cli\u003eHong, J. Y. \u003cem\u003eet al.\u003c/em\u003e Claudin 18.2 expression in various tumor types and its role as a potential target in advanced gastric cancer. \u003cem\u003eTransl Cancer Res\u003c/em\u003e \u003cstrong\u003e9\u003c/strong\u003e, 3367\u0026ndash;3374 (2020).\u003c/li\u003e\n\u003cli\u003eWang, C. \u003cem\u003eet al.\u003c/em\u003e CLDN18.2 expression and its impact on prognosis and the immune microenvironment in gastric cancer. \u003cem\u003eBMC Gastroenterol\u003c/em\u003e \u003cstrong\u003e23\u003c/strong\u003e, 283 (2023).\u003c/li\u003e\n\u003cli\u003eYu, X. \u003cem\u003eet al.\u003c/em\u003e Immunological role and prognostic value of somatostatin receptor family members in colon adenocarcinoma. \u003cem\u003eFront Pharmacol\u003c/em\u003e \u003cstrong\u003e14\u003c/strong\u003e, 1255809 (2023).\u003c/li\u003e\n\u003cli\u003eManica, M. \u003cem\u003eet al.\u003c/em\u003e Toward Explainable Anticancer Compound Sensitivity Prediction via Multimodal Attention-Based Convolutional Encoders. \u003cem\u003eMol. Pharm.\u003c/em\u003e \u003cstrong\u003e16\u003c/strong\u003e, 4797\u0026ndash;4806 (2019).\u003c/li\u003e\n\u003cli\u003eBauer, E. \u0026amp; Thiele, I. From Network Analysis to Functional Metabolic Modeling of the Human Gut Microbiota. \u003cem\u003emSystems\u003c/em\u003e \u003cstrong\u003e3\u003c/strong\u003e, (2018).\u003c/li\u003e\n\u003cli\u003ePiat, C. \u003cem\u003eet al.\u003c/em\u003e A Validated and Explainable Deep Learning Model Instantly Predicts Survival from Consultation Reports. SSRN Scholarly Paper at https://doi.org/10.2139/ssrn.4410792 (2023).\u003c/li\u003e\n\u003cli\u003eThorsson, V. \u003cem\u003eet al.\u003c/em\u003e The Immune Landscape of Cancer. \u003cem\u003eImmunity\u003c/em\u003e \u003cstrong\u003e48\u003c/strong\u003e, 812-830.e14 (2018).\u003c/li\u003e\n\u003cli\u003eLiu, J. \u003cem\u003eet al.\u003c/em\u003e An Integrated TCGA Pan-Cancer Clinical Data Resource to Drive High-Quality Survival Outcome Analytics. \u003cem\u003eCell\u003c/em\u003e \u003cstrong\u003e173\u003c/strong\u003e, 400-416.e11 (2018).\u003c/li\u003e\n\u003cli\u003eNellore, A. \u003cem\u003eet al.\u003c/em\u003e Rail-RNA: scalable analysis of RNA-seq splicing and coverage. \u003cem\u003eBioinformatics\u003c/em\u003e \u003cstrong\u003e33\u003c/strong\u003e, 4033\u0026ndash;4040 (2017).\u003c/li\u003e\n\u003cli\u003eAn, G. The Effects of Adding Noise During Backpropagation Training on a Generalization Performance. \u003cem\u003eNeural Computation\u003c/em\u003e \u003cstrong\u003e8\u003c/strong\u003e, 643\u0026ndash;674 (1996).\u003c/li\u003e\n\u003cli\u003eHeaton, J. Ian Goodfellow, Yoshua Bengio, and Aaron Courville: Deep learning: The MIT Press, 2016, 800 pp, ISBN: 0262035618. \u003cem\u003eGenetic Programming and Evolvable Machines\u003c/em\u003e \u003cstrong\u003e19\u003c/strong\u003e, (2017).\u003c/li\u003e\n\u003cli\u003eBergstra, J., Yamins, D. \u0026amp; Cox, D. D. Making a science of model search: hyperparameter optimization in hundreds of dimensions for vision architectures. in \u003cem\u003eProceedings of the 30th International Conference on International Conference on Machine Learning - Volume 28\u003c/em\u003e I-115-I\u0026ndash;123 (JMLR.org, Atlanta, GA, USA, 2013).\u003c/li\u003e\n\u003cli\u003edeCastro, B. R. Cumulative ROC curves for discriminating three or more ordinal outcomes with cutpoints on a shared continuous measurement scale. \u003cem\u003ePLOS ONE\u003c/em\u003e \u003cstrong\u003e14\u003c/strong\u003e, e0221433 (2019).\u003c/li\u003e\n\u003cli\u003eAkiba, T., Sano, S., Yanase, T., Ohta, T. \u0026amp; Koyama, M. Optuna: A Next-generation Hyperparameter Optimization Framework. \u003cem\u003earXiv:1907.10902 [cs, stat]\u003c/em\u003e (2019).\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"neoplasm, prognosis, deep learning, transfer learning, RNA-sequencing","lastPublishedDoi":"10.21203/rs.3.rs-7556679/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7556679/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e RNA-sequencing (RNA-seq) analysis offers a tumor-centered approach of interest for personalizing cancer care. However, existing methods \u0026ndash; including deep learning models \u0026ndash; struggle to reach satisfying performances on survival prediction based upon pan-cancer RNA-seq data. Here, we present DeepOS, a deep learning model that predicts overall survival (OS) from pan-cancer RNA-seq with a concordance-index of 0.714 and a survival AUC of 0.749 across 33 TCGA tumor types whilst tested on an unseen test cohort. DeepOS notably uses (i) prior biological knowledge to condense inputs dimensionality, (ii) mean squared error adapted to survival loss function and (iii) transfer learning to enlarge its training capacity through pre-training on organ prediction to improve the model performances (factors sorted by contributions). Interpretation showed that DeepOS learned biologically-relevant prognosis biomarkers. Altogether, DeepOS achieved competitive and consistent performances on pan-cancer prognosis estimation from individual RNA-seq data.\u003c/p\u003e","manuscriptTitle":"DeepOS: pan-cancer prognosis estimation from RNA-sequencing data","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-11-03 16:06:04","doi":"10.21203/rs.3.rs-7556679/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"3301dc3d-7a54-4df0-827c-f8aa58215bbe","owner":[],"postedDate":"November 3rd, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":57281770,"name":"Health sciences/Biomarkers"},{"id":57281771,"name":"Biological sciences/Cancer"},{"id":57281772,"name":"Biological sciences/Computational biology and bioinformatics"},{"id":57281773,"name":"Health sciences/Oncology"}],"tags":[],"updatedAt":"2026-03-23T08:28:48+00:00","versionOfRecord":[],"versionCreatedAt":"2025-11-03 16:06:04","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7556679","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7556679","identity":"rs-7556679","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}