Five ground-breaking artificial intelligence biology research abstracts led by AI biotech Owkin have been published by ESMO ahead of its 2022 Congress.
Highlighting research at the cutting edge of machine learning and biology, each abstract furthers Owkin’s mission to discover new mechanisms of diseases and derive better treatments for unmet medical needs. Owkin’s research and partnership teams will be at booth #252a to share our latest innovations, including recently launched AI diagnostics solutions and upcoming projects at the intersection of spatial omics and AI.
Eric Durand, SVP Data Science at Owkin, said:
Owkin’s expertise in histogenomics enables us to discover clinically-relevant subtypes, predict patient outcomes, such as survival and relapse, and identify genomic alterations from routinely obtained H&E images. Owkin’s research demonstrates the strength of our multi-modal approach to precision medicine.
Predicting KRAS G12C subtype from non-small cell lung cancer H&E slides using deep learning
Poster. Presenter: Isabelle Soubeyran MD, Molecular Pathology Laboratory Head at Institut Bergonié, Bordeaux
This abstract is the fruit of collaboration between Owkin, Amgen and Institute Bergonié. The study aims to show that AI techniques applied to digital pathology could offer a fast and cheap mass patient screening solution to complement DNA testing by predicting the KRAS mutation status and its G12C subtype from digital pathology slides used in clinical workflow.
Data from 1,076 non-small cell carcinoma patients from Institut Bergonie and partner centers was used to train a DeepMIL deep learning model. For each patient, H&E diagnostic slides from resections or biopsies, as well as the KRAS mutation status and G12C subtype obtained with NGS, have been used to run analyses and train models.
Our deep learning models succeed in discriminating KRAS G12C mutated patients from wild type or other subtype by using routine H&E histology data with encouraging performance of 65% ROC AUC obtained on a local cohort. It is the first demonstration that KRAS G12C mutated patient identification would be possible through morphological patterns analysis from annotation-free whole slide imaging.
These innovative approaches could complete existing testing techniques and pave the way for a more systematic, early and potentially cost-effective mass screening approach of KRAS G12C mutations in clinical routine.
Machine Learning-based prediction of Germinal Center, MYC/BCL2 Double Protein Expressor status, and MYC rearrangement from Whole Slide Images in DLBCL patients
Oral presentation. Presenter: Charlotte Syrykh MD, IUCT Oncopole, Toulouse
This abstract, a collaboration between Carnot CALYM and Owkin, used deep learning to predict subtypes of patients with Diffuse Large B-Cells Lymphoma (DLBCL), the most common type of non-Hodgkin lymphoma in adults (30-40%). In this project – named ILIA (LYmphoma Study – Artificial Intelligence) and led by Pr Camille Laurent and Pr Christiane Copie-Bergmann – a deep learning model was trained on 565 whole-slide images (WSI) from LYSA – The Lymphoma Study Association studies. The model was trained to predict cell of origin (COO) and double-protein expressors (DPE) status, the presence of MYC rearrangements (no MYC rearrangement/MYC-Single Hit or HGBL-Double Hit/Triple Hit) and expression of BCL6, CD10 and MUM1 proteins from WSI. Performance was evaluated using several repetitions of stratified five-fold cross-validation.
The DL model achieved a ROC AUC of 0.624 for GC, 0.687 for DPE, and 0.675 for MYC rearrangement. Using Cox proportional hazard model, predictions of DPE status (HR=0.38, P=.016), and MYC rearrangements (HR=5.23, P<.001) and MUM1 expression (HR=2.80, P=.027) were associated with worse overall survival.
The study demonstrates the predictive power of deep learning applied to WSI to predict DLBCL subtypes. Such predictive models could be used to augment pathologists analysis capacities, especially when IHC staining or FISH are not available.
Deep learning predicts patients’ outcome and mutations from H&E slides in gastrointestinal stromal tumor (GIST)
Proffered paper. Presenter: Raul Perret, MD, MSc, Bone and Soft Tissue Pathologist, Institut Bergonié and Centre Hospitalier Universitaire de Bordeaux
This abstract, a collaboration between Owkin, Institut Bergonié and Center Léon Bérard, used deep learning to predict the outcome and mutations of patients with gastrointestinal stromal tumor (GIST) from H&E slides. Existing risk evaluating system depends heavily on mitotic counts, which is laborious and sometimes inconsistent between pathologists. Molecular testing is costly and time consuming, therefore not systematically performed in all countries. New methods to improve risk and molecular predictions are hence crucial to improve the tailoring of adjuvant therapy.
We have built deep learning (DL) models on digitized H&E-stained whole slide images (WSI) to predict patients’ outcome and mutations. Models were trained with a cohort of 1233 GIST and validated on an independent cohort of 286 GIST. DL models outperformed the Miettinen classification for relapse-free-survival prediction in localized GIST without adjuvant Imatinib (C-index=0.83 in testing and 0.72 for independent validation). DL splitted Miettinen intermediate risk GIST into high/low risk groups(p value=2.1e-03). DL models achieved an area under the receiver operating characteristic curve (AUC) of 0.81, 0.91 and 0.71 for predicting mutations in KIT, PDGFRA and wild type respectively in testing and 0.84, 0.93 and 0.56 in independent validation. Notably, PDGFRA exon18 D842V mutation which is resistant to Imatinib, was predicted with an AUC of 0.87 and 0.90 in testing and independent validation, respectively. Additionally, novel histological criteria predictive of patients’ outcome and mutations were identified by reviewing the tiles selected by the models.
Our results strongly suggest that implementing DL with digitized WSI may represent a reproducible way to improve tailoring therapy and precision medicine for patients with GIST.
Blind validation of an AI-based tool for predicting distant relapse from breast cancer HES stained slides
Poster. Presenter: Ingrid Garberis, Institut Gustave Roussy.
This abstract is the first fully-blind validation of RlapsRisk, an AI-based tool for assessing the risk of distant relapse, which received CE-IVD approval for diagnostic use in Europe this month. RlapsRisk is an AI-based tool for assessing the risk of distant relapse at five years of ER+HER2- early invasive breast cancer (eiBC) patients from HES (hematoxylin-eosin-safran) whole slide images (WSI).
HES WSI of 676 ER+/HER2- eiBC diagnosed at Gustave Roussy from 2012 to 2017 included in the CANTO cohort, constituted the validation dataset (19 patients relapsed at 5 years). We compared RACE performance to the two most relevant clinical scores: Predict Breast and CTS0.
The obtained results showed the ability of RlapsRisk to generalize on independent data and thus endorse the soundness of the method. Furthermore, additional analysis brings to light the clinical value of RlapsRisk and that it could be used for therapeutic de-escalation purposes. This validation will be extended to multi-site and multi-scanner eiBC WSI from the CANTO cohort under completion.
Blind validation of MSIntuit, an AI-based pre-screening tool for MSI detection from colorectal cancer H&E slides
Poster. Presenter: Professor Magali Svrcek, MD, PhD, Sorbonne Université, AP-HP, Saint-Antoine Hospital, Department of Pathology, Paris
This abstract is the first fully-blind validation of MSIntuit, an AI-based tool that pre-screens for microsatellite instability (MSI) in colorectal cancer (CRC) tumours, which received CE-IVD approval for diagnostic use in Europe this month.
H&E WSI of 600 consecutive resected CRC (including n=123 dMMR/MSI cases) diagnosed at Medipath pathology laboratories in 2017/2018 were studied. dMMR status was assessed using IHC for the 4 MMR proteins, and confirmed by PCR for doubtful cases. To assess performance in the most robust way, inference was done on the remaining 570 patients blinded to their status. Automated quality check (QC) discarded WSI that did not meet the tool requirements (large blurry/artifact regions, too few tumor tissue).
MSIntuit reaches sensitivity comparable to gold standard methods (92-95%) while reducing almost by 40% the number of patients to screen with standard techniques, paving the way for MSIntuit use in clinical practice.