Abstract A044: Developing a digital twin framework for colorectal cancer outcome prediction and robust clinical trial simulation

Abstract Introduction: Digital twin (DT) technology enables in silico patient representations for counterfactual outcome simulations and precise target population identification, thereby facilitating in silico hypothesis testing while reducing clinical trial costs and resource utilization. We present a novel DT framework for colorectal cancer (CRC) patient outcome prediction and clinical trial simulation. Methods: We used multiomics data from 2,100 CRC patients from our Precision 360™ database. The data were partitioned into training (80%), validation (10%), and testing (10%) cohorts. The DT architecture incorporates an encoder-decoder neural network framework that compresses patient data into latent vectors (serving as digital twins) and subsequently decodes clinically relevant observables including overall survival (OS). Training employed a self-supervised approach with dual optimization criteria: minimization of reconstruction loss and maximization of latent space entropy. Training proceeded for 10,000 epochs until convergence of validation and training curves. We conducted two validation experiments related to AST/ALT ratio, a parameter shown to be associated with overall survival: (1) to assess how well the DT could generate prognosis-related hypotheses, we compared the predictions made by the DT and those from univariate analysis between AST/ALT ratio at baseline (closet lab value after initial diagnosis) and OS using concordance indices (CI); and (2) evaluation of counterfactual stability through simulated AST/ALT manipulations across a spectrum from low to high values. Results: Among the CRC cohort, data demonstrated AST/ALT > 0.96 was associated with median OS=31±4 months versus AST/ALT < 0.96 with median OS=38±4 months (c-index=0.57, log-rank p<0.005). In the test set, DT-decoded survival predictions exhibited improved prognostic accuracy with c-index=0.60 (log-rank p<0.015). In counterfactual simulations, systematic AST/ALT modulation from low to high values (AST/ALT= 0.37, 1.0, 2.7) produced corresponding decreases in median OS from 37±3, 35±3, to 34±3 months, with intermediate AST/ALT values following the monotonic relationship. Conclusion: Our encoder-decoder based DT framework robustly simulates CRC patient outcomes and predicts survival outcomes. The model maintains consistent trends during counterfactual analysis. Further, the DT framework shows considerable potential in revealing subtle relationships between potential features such as AST/ALT and survival outcomes. This ability to generate new prognostic insights paves the way for hypothesis formation for future studies. As such, our framework could prove to be a valuable tool for clinical trial simulation and insight generation, potentially reducing unnecessary costs and adverse effects in CRC drug development and management. Citation Format: Ian Ruchlin, Mirella Altoe, Maria Monberg, Todd Oakland, Michael Rossi, Vivek Agarwal, Jitesh Chawla, Pyeush Gurha, Stephen Yip. Developing a digital twin framework for colorectal cancer outcome prediction and robust clinical trial simulation [abstract]. In: Proceedings of the AACR Special Conference in Cancer Research: Artificial Intelligence and Machine Learning; 2025 Jul 10-12; Montreal, QC, Canada. Philadelphia (PA): AACR; Clin Cancer Res 2025;31(13_Suppl):Abstract nr A044.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.