AI and Data Science Approaches in Precision Oncology

This summary covers AI and data science components from AACR 2024, especially how it is integrated into basic science, drug development, patient recruitment in clinical trials, and advancing the promise of precision medicine for the entire cancer continuum.

Artificial Intelligence (AI): A world of sophisticated computer programs that learn from the data to make decisions. Dr. Vivek Subbiah, in the AACR plenary session, beautifully described how we have evolved our perception of oncology research from using light microscopes to molecular microscopes (next-generation sequencing) and to AI to look much more profound in disease biology.

Technology Advancements in Precision Era: Computational advancements in data integration algorithms, along with approaches such as single-cell spatial analysis, digital pathology, and next-generation sequencing, are shaping the oncology workflows at the bench and bedside.

"Good" Data is the New Oil: The enormous amount of data generated every time needs to be diverse, representative, curated, de-identified, harmonized, and re-usable (submitted to the open resources, e.g., MIDRC, funded by the ARPA-H). The right data should undergo the right AI prediction workflow (by Dr. Maryellen Giger).

Discovery and Tumor Biology: With innovative approaches, researchers can uncover critical cellular states, tumor microenvironment compositions, and architecture during therapy-driven tumor evolution and resistance mechanisms.

AI Approach to Pan-Tumor Biomarker Development (by Dr. Mia Levy):

o   AI-based tools face several challenges, such as biomarker testing needing vigorous clinical validation, reimbursement issues after clinical use, feasibility assessment, and lack of harmonized clinical utility workflows.

o   Dr. Levy’s team has developed biomarker datasets after comprehensive molecular profiling for clinical care and research using complex biomarkers such as genomic loss of heterozygosity (gLOH) and homologous recombination deficiency (HRD) signatures.

o   HRD involves the inactivation of BRCA genes and other members in the homologous recombination repair (HRR) pathway. Targeting HRD has been used in ovarian, breast, pancreatic, and prostate cancers.

o   The BRCA1/2 mutational landscape, observed across tumor types, has spurred the development of a pan-tumor HRDsig method. This machine learning comprehensive scar-based copy number signature readout for pan-tumor analysis, HRDsig, holds immense promise in identifying patients who could benefit from PARP inhibitors, pending clinical validation studies.

Predicting a Diagnosis Score Using EHR: A hospital-based study developed a diagnosis score algorithm by utilizing electronic health records by doctors to speed up the ovarian cancer diagnosis process, complementing the blood workup and histological findings. The study predicted the proportion of patients later diagnosed with ovarian cancer (Poster: Minh Tung Phung, PhD, MPH at the University of Michigan).

Improving Clinical Trial Design: 

o   Genomics and spatial biology data, with their potential to subgroup biology-based characteristics, are poised to revolutionize patient recruitment. Their early analysis and decision-making capabilities, coupled with AI approaches, hold promise for the future of clinical research.

o   Innovating trial designs can modernize evidence generation. For example, efficient patient-centric trials using master protocol, adaptive design, common controls, and adding decentralized elements could help advance trials quickly.

o   The role of expanding data sources, such as patient-generated data, EHR, and registries, in enabling advanced data integration and analysis cannot be overstated. These sources, when harnessed with the aid of advanced technologies (Project GENIE being a prime example), can significantly enhance our understanding of real-world clinico-genomic data.

Multimodal Data Integration: In an Exhibitor Spotlight Session- Patterns in Patients discussed why precision medicine needs multimodal data and AI.

o   Technological advancements enable us to combine information resources to increase predictive power, such as improving patient stratification for personalized medicine.

o   AI can uncover a deep understanding of a patient's disease biology. Such approaches allow experts to identify patients who will respond to the therapy before taking a drug to trial.

Patient Outcome Predictions:  

o   Using multimodal real-world data based on clinical, radiological, pathological, and genomic signatures can predict patient outcomes better than individual parameters. For example, Dr. Sohrab Shah presented the MSK MIND initiative using an advanced NSCLC study comprising 247 patients treated with anti-PD-1/PD-L1 therapy (PMID:36038778). 

o   Multi-omic data integration was tested in ovarian cancer in the MSK SPECTRUM initiative, where single-cell RNA sequencing, digital pathology, multiplexed immunofluorescence, and whole genome sequencing (MSK IMPACT) data could track the clonal evolution over time in patients to define drug resistance.

AI in Medical Imaging, Data Science, and Health Disparities (by Dr. Maryellen Giger):

o   AI improves efficiency, and workflows empower medical imaging teams to interpret data, predict/diagnose the disease early, and give the right patient the right treatment at the right time (e.g., Computer-aided detection (CADe) and diagnosis (CADx) system). However, diverse data is needed to train and yield trustworthy AI systems.

o   The imaging workflow involves (a) the patient undergoing imaging, (b) the FDA-approved AI system evaluating images, (c) the AI system predicting the patient outcome, and (d) the provider using the AI-guided prediction report for decision-making in diagnosis and therapy outcomes.

o   For example, one of the human-engineered AI (deep learning) of breast cancer on MRI can be analyzed as a virtual biopsy with the help of 3D/4D lesion features of the image (e.g., size, shape, morphology of the features following contrast enhancement).

o   The READER study demonstrated the value (improved prediction) of AI-based methodologies when comparing outcomes (radiologists using with or without the AI aid) that led to the FDA approval of the QuantX system in 2017 for breast cancer diagnosis.

o   In a multiscale model, such methodologies are integrated with genomics data for improved prediction and patient stratification.

AI for Accessible, Affordable, and Equitable Precision Medicine:

o   The study presented by Dr. Anant Madabhushi integrated the Oncotype recurrence score genomics data with the digital pathology, uncovering improved prediction/identification of high-risk disease in early-stage ER+/Her2- breast cancer.

o   His team mapped tumor features (epithelia, stroma, tumor cell clusters, tumor-infiltrating lymphocytes, etc.) on digital images and developed an image-based risk score (ibRiS) that conserves tissue features compared to genomics-based methodologies that destroy tissue features during the sample preparation.

o   The structural features like collagen fiber orientation on digital images can predict the outcome of breast cancer patients better than genomics (Chen et al.; Breast Cancer, 2023) and can identify a group of patients who can forgo chemotherapy.

o   He concluded that low-cost digital pathology-based methods are easy to adopt in low-income countries that lack NGS resources in diagnosing patients.

AI for Precision Drug Development and Discovery (by Dr. Thomas Clozel):

o   Drug development processes should be centered on biology, from finding new targets using spatial technologies to uncovering vast data on microenvironments and tumor heterogeneity.

o   High-quality data must be shared by collaboration and breaking silos and offered centralized and decentralized manner.

o   With the help of AI, multiscale algorithms (body, organ, cells, and molecules) can be generated and tested to analyze multimodal data on patient outcomes following therapy.

o   Approaches like digital pathology anchor the images' key features, emphasizing biology's critical role in drug development workflows.

AI and Data Science Fostering Collaboration: AI data sources are growing and integrating and are being shared collaboratively with private and government organizations. Alliances are growing, helping institutions previously working in silos.

NCI's Investment and Funding Resources (by Drs. Jennifer Couch, Juli Klemm, Emily Greenspan, and Christopher Gibbon): Through various grants and funding programs, NCI supports extramural research on AI in cancer research.

o   NCI supports the development of data-driven models, software tools, and datasets for broad use by cancer research researchers (e.g., GAuDI, a framework for preprocessing electronic health data, and FrESCO, a modular deep learning library for extracting information from clinical text documents).

o   NCI's activities are designed to be accessible and inclusive, enabling virtual panel discussions (e.g., Cancer AI Conversations) to explore emerging topics in AI and cancer. In this spirit, NCI has recently launched NanCI, an AI-driven mobile application that empowers users to engage with and contribute to scientific topic searches.

AI Challenges in Oncology (by Dr. Thomas Clozel): With the boom of AI in oncology and medicine, several challenges still exist, e.g., biopharma research silos, lack of generative chemistry use, lack of wet lab validation, lack of proof of concepts for foundation models, and lack of reimbursement of AI-driven approaches in clinical care.