Abstract 2788
Background
Deep learning (DL) is one of the best approaches to predict nonlinear behaviors from high dimensional data. Nevertheless predicting the outcome of patients affected by cancers from transcriptomic data has shown limited performance, even with DL (C-index usually <0.65). Transfer learning is a DL two-step method where a model is pre-trained for a basic task on large amount of data, and then fine-tuned on the aimed task. We hypothesized that using TL with RNAseq may improve the performances of cancer patients’ outcome estimation.
Methods
The model was a Multi-Mayer Perceptron (MLP) with 22913 inputs corresponding to genes bulk tumor whole genome RNAseq expression analysis. An important restriction was applied to the number of units at second layer (N = 100), with further linear decrease across subsequent layers. Architecture of the model (number of layers, skip connections), L1 normalization value and learning rate were optimized by grid search on 30 parallel models. Training was performed using Keras package in R. Data were split into 70% training, 15% cross validation, 15% validation for each step, without contamination between the 2 transfer learning steps. The pre-training step consisted in predicting the organs of sample origin using 17.487 public RNAseq data of normal & cancer tissues (GTEX from gtexportal.org & TCGA from cBioportal.org). Fine-tuning on patients survival used 6401 training tumors. The model’s performance on survival prediction was evaluated by C-index and the area under the survival receiver-operating characteristic curve (AUROC).
Results
The pre-training using GTEx and TCGA reached very high performance with validation accuracy of 0.96 to predict organ of origins for the best model (all models had validation accuracy > 0.9). Fine-tuning on survival, the prognostic performance of the best model on the validation cohort was C-index=0.74 and AUROC= 0.81 (80% of models had a C-index > 0.6). The best model had 8 hidden layers and a small penalization value.
Conclusions
Thanks to this original transfer learning method, we achieved a high performance to estimate cancer patients’ prognostic from whole genome expression, a classically challenging task. Learning on public databases is a valuable method of DL for personalized cancer care.
Clinical trial identification
Legal entity responsible for the study
The authors.
Funding
Has not received any funding.
Disclosure
E. Angevin: Advisory / Consultancy: Amgen; Advisory / Consultancy: Astellas; Advisory / Consultancy: AstraZeneca; Advisory / Consultancy: Bayer; Advisory / Consultancy: BeiGene; Advisory / Consultancy: BMS; Advisory / Consultancy: Celgene; Advisory / Consultancy: DebioPharma; Advisory / Consultancy: Genentech; Advisory / Consultancy: Ipsen; Advisory / Consultancy: Janssen; Advisory / Consultancy: Lilly; Advisory / Consultancy: MedImmune; Advisory / Consultancy: Novartis; Advisory / Consultancy: Pfizer; Advisory / Consultancy: Roche; Advisory / Consultancy: Sanofi; Advisory / Consultancy: Orion. A. Hollebecque: Advisory / Consultancy: Amgen; Advisory / Consultancy: Spectrum Pharmaceuticals; Advisory / Consultancy: Lilly; Advisory / Consultancy: Debiopharm; Travel / Accommodation / Expenses: Servier; Travel / Accommodation / Expenses: Amgen; Travel / Accommodation / Expenses: Lilly; Travel / Accommodation / Expenses: Incyte; Travel / Accommodation / Expenses: Debiopharm. E. Deutsch: Advisory / Consultancy: Boehringer; Advisory / Consultancy: Medimune; Advisory / Consultancy: Amgen; Research grant / Funding (self): AstraZeneca; Research grant / Funding (self): biotrachea; Research grant / Funding (institution): BristolMyersSquidd; Research grant / Funding (self): Clevelex; Research grant / Funding (self): EDF; Research grant / Funding (self): Lilly; Research grant / Funding (self): GlaxoSmisthKline; Research grant / Funding (self): Merk; Research grant / Funding (self): Nanobiotix; Research grant / Funding (self): Oseo; Research grant / Funding (self): Ray Search Laboratory; Research grant / Funding (self): Roche; Research grant / Funding (self): Ipsen; Research grant / Funding (self): Servier; Research grant / Funding (self): Takeda. C. Massard: Advisory / Consultancy: Amgen; Advisory / Consultancy: Astellas; Advisory / Consultancy: AstraZeneca; Advisory / Consultancy: Bayer; Advisory / Consultancy: BeiGene; Advisory / Consultancy: BMS; Advisory / Consultancy: Celgene; Advisory / Consultancy: DebioPharma; Advisory / Consultancy: Genentech; Advisory / Consultancy: Ipsen; Advisory / Consultancy: Janssen; Advisory / Consultancy: Lilly; Advisory / Consultancy: MedImmune; Advisory / Consultancy: Novartis; Advisory / Consultancy: Pfizer; Advisory / Consultancy: Roche; Advisory / Consultancy: Sanofi; Advisory / Consultancy: Orion. L. Verlingue: Research grant / Funding (self): Bristol-Myers Squibb; Advisory / Consultancy: Pierre Fabre; Advisory / Consultancy: Adaptherapy. All other authors have declared no conflicts of interest.
Resources from the same session
4543 - Long-term real-world (RW) outcomes in patients with advanced melanoma (MEL) treated with ipilimumab (IPI) and non-IPI therapies: IMAGE study
Presenter: Stéphane Dalle
Session: Poster Display session 3
Resources:
Abstract
4523 - Prognostic Factors for efficacy of Ipilimumab used after AntiPD1 and/or BRAF+MEK inhibitors in Melanoma Patients: an Italian Melanoma Intergroup study
Presenter: Riccardo Marconcini
Session: Poster Display session 3
Resources:
Abstract
3632 - Rechallenge with combination ipilimumab and anti-PD-1 (IPI+PD1) in metastatic melanoma after acquired resistance to IPI+PD1 immunotherapy
Presenter: Adriana Hepner
Session: Poster Display session 3
Resources:
Abstract
3732 - Clinicopathologic characteristics of immune colitis in melanoma patients treated with combination ipilimumab and anti-PD1 (IPI+PD1) and PD1 monotherapy.
Presenter: Kazi Nahar
Session: Poster Display session 3
Resources:
Abstract
5005 - Real-world outcomes of ipilimumab plus nivolumab for advanced melanoma in the Netherlands
Presenter: Michiel van Zeijl
Session: Poster Display session 3
Resources:
Abstract
5524 - Utilization of Real-World Data to Assess the Effectiveness of Immune Checkpoint Inhibitors (ICIs) in Elderly Patients with Metastatic Melanoma
Presenter: D Scott Ernst
Session: Poster Display session 3
Resources:
Abstract
5884 - Tumor mutational burden and response to PD-1 inhibitors: an analysis of 89 cases of metastatic melanoma.
Presenter: Léa Dousset
Session: Poster Display session 3
Resources:
Abstract
3120 - Increase in S100B and LDH as early outcome predictors for non-responsiveness to anti-PD-1 monotherapy in advanced melanoma.
Presenter: Elisa Rozeman
Session: Poster Display session 3
Resources:
Abstract
2157 - Immune status defined by molecular information layers predicts response to pembrolizumab treatment in advanced melanoma
Presenter: Guillermo Prado-Vázquez
Session: Poster Display session 3
Resources:
Abstract
2553 - Interim analysis of a phase Ib study of cobimetinib plus atezolizumab in patients with advanced BRAFV600 wild type melanoma progressing on prior anti-PD-L1 therapy
Presenter: Shahneen Sandhu
Session: Poster Display session 3
Resources:
Abstract