scCTM: Continuous Thought Machine Reconstruction of Glioma Cell Fate Dynamics Attractor Landscapes and Algorithmic Perturbation Signatures

Abstract

Cell fate plasticity dynamics in glioma ecosystems is usually inferred from static single-cell snapshots, leaving a gap between manifold geometry, predictive modelling, and mechanistic intervention ranking. We developed scCTM, a complexity theory and machine learning framework that combines the Sakana AI Continuous Thought Machine (CTM) with autoencoding, a continuous actor-critic policy over synchronization-derived latent states for predicting single-cell fate dynamics. We also incorporated integrated-gradient and decoder-weight explainability, network science, empirical Waddington-like landscape (attractor) reconstruction, and post-hoc Block Decomposition Method (BDM) perturbation analysis. The complexity framework was evaluated in two malignant cell secondary analyses: GSE210567 stem-like versus differentiated H3K27M cells (18,490 cells) and longitudinal H3K27M initial versus recurrent disease trajectory (20,000 cells) from Sussman et al. (2026) as a ML binary classification problem. CTM-AE achieved AUROC values of $0.99$ in both datasets. CTM-RL preserved classification almost exactly but did not improve reconstruction. Out-of-fold CTM-RL latent states produced potential basins across PCA, UMAP, and PHATE embeddings. BDM edge perturbation analysis further identified recurrent high-impact network interactions, including epigenetic/chromatin-associated signatures involving ARID1B and developmental or neuronal-state genes such as AUTS2, CTNND2, DPP6, NRXN1, and CADM2. These in silico findings support CTM synchronization as an interpretable latent-state representation and predictive forecasting of cell fate dynamics, as well as BDM perturbation signatures as a complementary in silico intervention ranking layer for systems medicine.