Abstract
The application of data science in cancer research has been boosted by major advances in three primary areas: (1) Data: diversity, amount, and availability of biomedical data; (2) Advances in Artificial Intelligence (AI) and Machine Learning (ML) algorithms that enable learning from complex, large-scale data; and (3) Advances in computer architectures allowing unprecedented acceleration of simulation and machine learning algorithms. These advances help build in silico ML models that can provide transformative insights from data including: molecular dynamics simulations, next-generation sequencing, omics, imaging, and unstructured clinical text documents. Unique challenges persist, however, in building ML models related to cancer, including: (1) access, sharing, labeling, and integration of multimodal and multi-institutional data across different cancer types; (2) developing AI models for cancer research capable of scaling on next generation high performance computers; and (3) assessing robustness and reliability in the AI models. In this paper, we review the National Cancer Institute (NCI) -Department of Energy (DOE) collaboration, Joint Design of Advanced Computing Solutions for Cancer (JDACS4C), a multi-institution collaborative effort focused on advancing computing and data technologies to accelerate cancer research on three levels: molecular, cellular, and population. This collaboration integrates various types of generated data, pre-exascale compute resources, and advances in ML models to increase understanding of basic cancer biology, identify promising new treatment options, predict outcomes, and eventually prescribe specialized treatments for patients with cancer.
Introduction: The Complexity of Cancer Research
Cancer research is currently bottlenecked by the sheer complexity and multiscale nature of the disease. While the scientific community has access to more data than ever—ranging from genomics and imaging to clinical records—three major challenges persist. First, Data Integration remains technically and logistically difficult when merging multimodal and multi-institutional datasets. Second, Computational Scaling is a hurdle, as standard AI models often fail to scale effectively on the next generation of exascale supercomputers. Finally, Reliability is paramount; for AI to be useful in a clinical setting, robust Uncertainty Quantification (UQ) is required to determine when a model’s prediction can be trusted.
The JDACS4C initiative was formed to bridge these gaps by "co-designing" AI algorithms alongside the hardware they run on, leveraging the Department of Energy's high-performance computing (HPC) capabilities and the National Cancer Institute's vast biomedical data repositories.
The Three Pilots of JDACS4C
The initiative is structured around three "Pilots" that utilize distinct but overlapping technical approaches to address cancer at different biological scales.
Pilot One: Cellular Level Drug Response
This pilot uses multimodal deep learning to predict how tumors respond to specific drug pairs. By integrating diverse data types—including gene expression, proteomics, and drug descriptors—into a single neural network architecture, researchers can identify synergistic effects. A key insight from this work is that combining gene expression, microRNA, and proteomics data is significantly more powerful than using any single data type, explaining over 94% of the variation in tumor response.
Pilot Two: Molecular Dynamics and AI Steering
Pilot Two employs a novel multiscale modeling workflow known as MuMMI. In this framework, AI acts as a "smart filter" to monitor macroscopic simulations of RAS proteins on a lipid membrane. When the AI detects a unique or interesting molecular state, it automatically triggers a high-resolution Coarse-Grained Molecular Dynamics (CGMD) simulation.
This "active learning" approach allows for simulating biological processes over much longer timescales—milliseconds versus the traditional microseconds—representing a nearly 100-fold improvement over previous state-of-the-art simulations.
Pilot Three: Population-Scale Surveillance
Focusing on Natural Language Processing (NLP), Pilot Three uses multi-task learning and attention mechanisms to automate the extraction of clinical features from millions of unstructured pathology reports. Preliminary tests showed these tools could accurately classify five key data elements for 42.5% of cancer cases automatically, significantly reducing the manual burden on national cancer registries like SEER.
Infrastructure and Reliability
To support these pilots, the collaboration developed the CANcer Distributed Learning Environment (CANDLE). This software stack allows researchers to run deep learning models across thousands of nodes on supercomputers like Summit and Theta.
A critical component of CANDLE is the integration of Uncertainty Quantification (UQ). The authors argue that a prediction without a confidence interval is insufficient for clinical decision-making. UQ helps identify "out-of-distribution" data where the model might be guessing rather than predicting based on learned patterns.
Key Results and Future Directions
The JDACS4C initiative has already yielded significant quantitative results:
- Drug Response Accuracy: Pilot One achieved an R-squared value of 0.944 for drug pair response models using the NCI-ALMANAC dataset.
- Simulation Breakthrough: Pilot Two successfully simulated a 1x1µm lipid membrane with 300 RAS proteins, capturing over 200ms of protein behavior.
Despite these successes, challenges remain. Accessing and labeling multi-institutional data is still hindered by privacy barriers, and "opening the black box" of AI to understand the mechanistic biology behind drug responses is an ongoing goal. Future work will focus on moving toward fully atomistic resolution in molecular models and integrating population-level tools directly into clinical APIs for real-world application.
Summary updated