TMatt is a prototype transformer-based model under active development, designed to integrate tumor microenvironment (TME) signals with cell-intrinsic features for enhanced prediction of patient outcomes. By fusing diverse data modalities—including transcriptomic and spatial features alongside clinical variables—TMatt leverages self-attention mechanisms and Bayesian uncertainty quantification to deliver robust, interpretable predictions.
TMatt combines cell-intrinsic data with TME signals to generate a unified latent representation of patient samples. Its transformer architecture uses self-attention to capture complex intra- and inter-modality interactions, ensuring strong performance even when some data modalities are missing. This integrated approach not only improves predictive accuracy but also provides deeper insights into the biological underpinnings of treatment response.
- Data Integration:
TMatt incorporates:- Cell-intrinsic features.
- TME signals (e.g., transcriptome and spatial features).
- Clinical and outcome data from harmonized patient cohorts.
- Training Sources:
Training leverages large-scale, harmonized datasets, ensuring diverse and comprehensive input for model development.
- Unified Co-Embedding:
All inputs are projected into a shared latent space via tokenization and RBF encoding, integrating categorical, continuous, and survival variables. - Self-Attention Framework:
The model employs transformer-based self-attention to model spatial and relational interactions, effectively handling missing modalities through feature masking.
- Masked Modeling:
By masking missing features during training, TMatt learns to maintain predictive performance even in the absence of certain data types. - Bayesian Dropout:
Multiple stochastic forward passes approximate a Bayesian posterior, quantifying uncertainty in predictions for survival, progression-free survival, or treatment response.
- High-Level Validation:
TMatt undergoes rigorous cross-validation on held-out subsets and unseen cancer types, ensuring robust performance. - Comparative Analysis:
Integrated TME predictions are benchmarked against cell-intrinsic-only models to quantify the added value of TME data.
- Systematic Evaluation:
Various configurations—from single modalities to integrated multi-modal setups—are systematically compared to determine the optimal combination of TME features for reliable outcome prediction.
Explore the modules for data preprocessing, model training, self-attention implementation, and evaluation workflows. For further details or to contribute, please consult the issue tracker or contact the maintainers.
TMatt is a dynamic prototype aimed at advancing precision oncology by integrating multi-modal data to generate robust, interpretable predictions that drive meaningful clinical insights.