Researchers from Genentech Propose A Deep Learning Methodology to Discover a Predictive Tumor Dynamic Model from Longitudinal Clinical Data

Researchers from Genentech introduced tumor dynamic neural-ODE (TDNODE) as a pharmacology-informed neural network for enhancing tumor dynamic modeling in oncology drug development. Overcoming the limitations of existing models, TDNODE allows unbiased predictions from truncated data. Its encoder-decoder architecture expresses an underlying dynamical law with generalized homogeneity, representing kinetic rate metrics with inverse time as the unit. The generated metrics accurately predict patients’ overall survival, showcasing TDNODE’s utility in principled oncology disease modeling and improving personalized therapy decision-making.

TDNODE’s encoder-decoder architecture expresses a time-homogeneous dynamical law, generating metrics for accurate patients’ overall survival predictions. The proposed formalism enables principled integration of multimodal dynamical datasets in oncology disease modeling. The study specifies dimensions for the initial condition encoder output and GRU hidden layers. The implementation uses torchdiffeq, PyTorch, Pandas, Numpy, Scipy, Lifelines, Shap, and Matplotlib for solving, development, and analysis.

The study explores tumor growth dynamics using mathematical models, emphasizing the historical success of such models in describing experimental data. While non-linear mixed-effects modeling is common in pharmacometrics, machine learning has been underutilized for deriving metrics. The TDNODE framework integrates neural ODEs and ML, aiming to mine large oncology datasets for accurate predictions and enhanced understanding. The study aims to predict future patient outcomes early, enabling personalized therapy and advancing drug development through interpretable ML models.

TDNODE is a system that uses two encoders and a decoder based on an ODE solver. It employs a recurrent neural network to determine initial conditions and an attention-based LSTM to assess tumor kinetic parameters. Using numerical integration, the decoder represents the ODE system as a neural network and predicts tumor size over time. The Reducer component condenses the state vector for comparison with the tumor size.

The TDNODE model surpasses existing limitations by making unbiased predictions from truncated data and generating kinetic rate metrics for highly accurate overall survival predictions. TDNODE integrated multimodal dynamical datasets in oncology disease modeling, demonstrating its versatility and providing a principled approach for combining diverse data types. Continuous longitudinal tumor size predictions were generated for training and test sets, employing an ADAM optimization approach during 150 epochs with specified hyperparameters, achieving accurate predictions through careful configuration of L2 weight decay, learning rate, ODE tolerance, batch size, and observation window.

By utilizing kinetic rate metrics, TDNODE can provide highly precise predictions of survival rates even when working with incomplete or truncated data sets. This advanced approach overcomes the limitations of traditional survival analysis methods, which often need to be able to account for incomplete or missing data accurately. With TDNODE’s cutting-edge technology, researchers and healthcare professionals can obtain a more detailed understanding of patient outcomes, leading to better-informed treatment decisions and improved clinical outcomes. 

Further research avenues for TDNODE include exploring the incorporation of dosing or pharmacokinetics factors and enhancing the model’s comprehensiveness. Validation across diverse datasets will assess TDNODE’s generalizability in predicting future tumor sizes. Investigating TDNODE’s potential in personalized therapy is a promising direction, leveraging its ability for model discovery from longitudinal tumor data to support individualized treatment decisions. Exploring TDNODE in disease modeling beyond oncology could offer insights into its applicability and effectiveness in diverse medical contexts.

Check out the Paper. All credit for this research goes to the researchers of this project. Also, don’t forget to join our 33k+ ML SubReddit, 41k+ Facebook Community, Discord Channel, and Email Newsletter, where we share the latest AI research news, cool AI projects, and more.

If you like our work, you will love our newsletter..