Interpretable Forecasting of Kidney Cancer Progression via Generative AI and Symbolic Reasoning

Predicting the progression of kidney cancer stages from omics data and understanding the underlying mechanisms presents a major hurdle, partly due to the scarcity of longitudinal data and the interpretability issues of existing forecasting models. Cancer datasets like TCGA provide cross-sectional patient profiles rather than longitudinal ones, making timely treatment decisions challenging as tumors progress.

Deep neural networks commonly employed for forecasting, including LSTMs, exacerbate this issue by maintaining opacity, leaving clinicians without a clear means to verify their predictions. Clear cell renal cell carcinoma (ccRCC) exemplifies the critical nature of these challenges. The five-year survival rate drops significantly from over 94% at stage I to 28% at stage IV, yet early-stage tumors are often handled through active surveillance, a strategy hindered by a lack of molecular evidence for progression risk.

Detecting progression in a timely manner necessitates forecasts that clinicians can interpret and rely on, rather than opaque predictions. To tackle these challenges simultaneously, we introduce a novel approach that integrates generative and symbolic AI: a Variational Autoencoder trained on bulk RNA-Seq profiles of 530 TCGA ccRCC patients produces synthetic pseudo-time trajectories to compensate for the lack of longitudinal data.

A symbolic rule-induction framework, ASAL, then learns finite-state automata from these trajectories, translating stage transitions into human-readable Boolean conditions based on gene expression. These transitions are subsequently converted into probabilistic forecasts of stage progression by a complex event forecasting system, Wayeb.

To validate our method, an independent XGBoost classifier trained on real patients achieved an F1 score ranging from 0.71 to 0.81, demonstrating a gradual shift in the probability of progression along the synthetic trajectories, which is absent in non-progressing control trajectories. By analyzing the trajectories, we identified stage-dependent alterations in well-known kidney cancer-related pathways, such as the TCA cycle and DNA repair.

Notably, our symbolic forecaster achieved performance comparable to an LSTM baseline (macro F1 score of 0.928 compared to 0.964), but went a step further by providing an inspectable rule set and a probability distribution over the timing of transitions, rather than relying on a single opaque score. Our findings demonstrate that combining generative and symbolic AI can transform cross-sectional cancer datasets into a transparent, forecast-oriented framework for modeling disease progression, as illustrated in the case of ccRCC.