DeepMed IO Ltd: Utilising AI to deliver tomorrow’s personalised medicine, today!
Efficient management and utilisation of histopathology, molecular profiling and clinical data is key for increasing the success rate of clinical trials, especially in oncology and immunology. Effective data integration and data mining with machine-learning technologies will enable the implementation of real precision medicine schemes and the design and validation of next-generation companion diagnostics.
DeepMed IO, is an AI company with deep roots in life sciences. Operating within an ISO 13485 Quality Management System, DeepMed has delivered a top-performing (1) AI-powered metastasis detection solution (DeepPath™-LYDIA) ready for decision support in primary diagnosis.
The DeepMed team has published a proof-of-concept study in CELL-REPORTS (2) utilising deep-learning to predict short-term drug response as well as Patient Overall Survival. In the same study drug-specific information was extracted, by mining the deep-learning black-box, and was turned into novel Mechanism-of- Action (MoA) knowledge.
The company is currently working on deep tissue annotation and knowledge extraction in a way that can be effectively integrated with ‘omics’ and clinical data. We offer this know-how to pharmaceutical and biotech companies for developing custom-made solutions to increase clinical trials ROI and probability of success.
To address large scale clinical trial integration, DeepMed has teamed up with Inspirata and Fujifilm to deliver AI-augmented clinical trial data management. Working with those world-class partners DeepMed can offer full digital pathology functionality, hardware integration and complete, fully customised end-to-end solutions.
Below, we present three case studies where DeepMed IO has utilised deep neural networks to deliver proofs of concept for personalised medicine pharmaceutical applications.
- Predicting Patient Overall Survival (OS) from Cell-line Data
Figure legend: Predicting short-term drug efficacy and overall survival (OS) from omics data using deep learning
The aim of precision oncology is to predict the optimal therapy for each patient based on case-specific diagnostic information. We were the first team in the world to demonstrate through a high impact publication in CELL-REPORTS (2) that deep-learning networks trained on cell-line baseline gene expression (RNAseq) to predict drug response can be used to predict short-term response to therapy as well as overall survival in large clinical cohorts. Specifically we trained various deep-learning architectures along with other state of the art models through robust nested cross-validation schemes on the GDSC pharmacogenomic database (2) comprising at the time of publication of 1,001 cell lines with 251 drugs. The best models of each framework were then validated on large clinical cohorts for various sources including the MD Anderson Cancer Center and the TCGA.
The deep-learning models presented the best performance with high statistical significance in predicting drug response from gene expression, capturing the intricate biological interactions more effectively than current state-of-the-art machine learning frameworks. Surprisingly, deep-learning networks also predicted Overall Survival (OS) with statistical significance in most of the studies, while all other frameworks failed to do so. This finding was unexpected because it is a well-known fact that short-term response to therapy poorly correlated with the long-term patient outcome which is captured by OS and which is a Key Performance Indicator determining the success or failure of a clinical trial.
This methodology can be utilised (a) to hedge failure-risk through AI-model driven patient stratification and (b) to drive the development of next generation AI-powered companion diagnostics.
- Predicting Novel Modes-of-Action (MOA) by Opening the AI Black Box
Figure legend: Mining the deep-learning black-box to extract novel drug-specific Mechanism-of-Action (MoA) knowledge
Deep Learning Networks, trained to predict drug-response from baseline gene-expression generated in the aforementioned work (2) were interrogated to reveal drug-specific biologically meaningful concepts, such as regulatory pathways involved in the response to a specific drug, by extracting the weights connecting the input layer (gene expression) to each of the nodes of the first hidden layer. These weights, that did not correlate with gene-expression, were then used to perform pathway enrichment analysis.
Detailed literature search revealed that around 80% of predictions following this methodology agreed with existing knowledge. However a significant percentage pointed towards novel pathways that are yet to be examined.
This methodology can point towards yet unknown MOAs from in-vitro data and reveal novel targets.
- Developing CDxs by performing AI-driven full slide annotation and knowledge extraction from H&E stained histopathology images using DNNs
Figure legend: Developing CDxs by performing AI-driven full slide annotation and knowledge extraction from H&E stained histopathology images using DNNs
A proof of concept work carried out by DeepMed’s collaborators (3) revealed that H&E stained biopsy tissues contain adequate information to predict patient outcomes such as recurrence. In particular, 5-year recurrence risk in breast cancer patients with ductal carcinoma in situ was accurately predicted (86% accuracy) by applying AI-assisted detailed tissue annotation, manual feature extraction and machine learning to digitised tumor biopsy images.
DeepMed is perfecting the methodology by applying advanced DNN-powered gigapixel-compression and semi-supervised multiple-Instance learning methodologies that boost performance and remove the need for detailed manual tissue annotations from expert pathologists, manual feature extraction and selection for machine learning training.
Such machine learning models could replace expensive genetic-based tests like OncotypeDXTM. Furthermore, they could form the core of AI-powered next-generation companion diagnostics for targeted therapies with zero overhead, low cost and extremely fast turn-around time.
To request further information and to discuss your specific needs, please email email@example.com or call +447391044143.
1. https://camelyon17.grand-challenge.org/evaluation/results/ (5th best solution at the time of submission)
2. Sakellaropoulos T, Vougas K, Narang S, et al. A Deep Learning Framework for Predicting Response to Therapy in Cancer. Cell Rep. 2019;29(11):3367‐3373.e4. doi:10.1016/j.celrep.2019.11.017
3. Klimov S, Miligy IM, Gertych A, et al. A whole slide image-based machine learning approach to predict ductal carcinoma in situ (DCIS) recurrence risk. Breast Cancer Res. 2019;21(1):83. Published 2019 Jul 29. doi:10.1186/s13058-019-1165-5