Precision medicine is driven by technologies such as rapid genome sequencing and artificial intelligence (AI), according to a presentation at the AACR virtual meeting II.
AI can be applied to the sequencing information derived from advanced cancers to make highly personalized treatment recommendations for patients, said, of Weill Cornell Medicine, New York.
Dr. Elemento described such work during theof the meeting.
Dr. Elemento advocated for whole-genome sequencing (WGS) of metastatic sites, as it can reveal “branched evolution” as tumors progress from localized to metastatic ().
The metastases share common mutations with the primaries from which they arise but also develop their own mutational profiles, which facilitate site-of-origin-agnostic, predictive treatment choices.
As examples, Dr. Elemento mentioned HER2 amplification found in a patient with urothelial cancer () and a patient with uterine serous carcinoma ( ), both of whom experienced long-lasting remissions to HER2-targeted therapy.
Dr. Elemento also noted that WGS can reveal complex structural variants in lung adenocarcinomas that lack alterations in the RTK/RAS/RAF pathway (unpublished data).
Application of machine learning
One study suggested that microRNA expression and machine learning can be used to identify malignant thyroid lesions (). The approach diagnosed malignant lesions with 90% accuracy, 100% sensitivity, and 86% specificity.
Dr. Elemento and colleagues used a similar approach to predict response to immunotherapy in melanoma (unpublished data).
The idea was to mine the cancer genome and transcriptome, allowing for identification of signals from neoantigens, immune gene expression, immune cell composition, and T-cell receptor repertoires, Dr. Elemento said. Integrating these signals with clinical outcome data via machine learning technology enabled the researchers to predict immunotherapy response in malignant melanoma with nearly 90% accuracy.
AI and image analysis
Studies have indicated that AI can be applied to medical images to improve diagnosis and treatment. The approach has been shown to:
- Facilitate correct diagnoses of malignant skin lesions (Nature. 2017 Feb 2;542:115-8).
- Distinguish lung adenocarcinoma from squamous cell cancer with 100% accuracy (EBioMedicine. 2018 Jan;27:317-28).
- Recognize distinct breast cancer subtypes (ductal, lobular, mucinous, papillary) and biomarkers (bioRxiv 242818. doi: 10.1101/242818; EBioMedicine. 2018 Jan;27:317-28)
- Predict mesothelioma prognosis (Nat Med. 2019 Oct;25:1519-25).
- Predict prostate biopsy results (unpublished data) and calculate Gleason scores that can predict survival in prostate cancer patients (AACR 2020, Abstract 867).
Drug development through applied AI
In another study, Dr. Elemento and colleagues used a Bayesian machine learning approach to predict targets of molecules without a known mechanism of action ().
The method involved using data on gene expression profiles, cell line viability, side effects in animals, and structures of the molecules. The researchers applied this method to a large library of orphan small molecules and found it could predict targets in about 40% of cases.
Of 24 AI-predicted microtubule-targeting molecules, 14 depolymerized microtubules in the lab. Five of these molecules were effective in cell lines that were resistant to other microtubule-targeted drugs.
Dr. Elemento went on to describe how Oncoceutics was developing an antineoplastic agent called ONC201, but the company lacked information about the agent’s target. Using AI, the target was identified as dopamine receptor 2 (DRD2;).
With that information, Oncoceutics initiated trials of ONC201 in tumors expressing high levels of DRD2, including a highly resistant glioma (