Research

ImmuneWorld is the first foundation model-based world simulator for the tumor immune microenvironment (TIME). For the first time, we formulate TIME dynamics as a world-modeling problem — a trajectory-aware transformer autoregressively pre-trained on 12.4M single cells, validated across 7 cancer types and 14 independent clinical cohorts, outperforming 10 state-of-the-art baselines (scGPT, scFoundation, GEARS, VCWorld, et al.) across four downstream tasks.

Key Contributions

• First to frame TIME dynamics as a world-modeling problem — a 12-layer trajectory-aware Transformer (68M params, FlashAttention-2)
• Trajectory-aware attention with learnable temporal decay bias: cell-state transition Pearson r lifted from scGPT 0.554 to 0.914 (+65% relative)
• Cross-Cancer Transfer Module with gradient-reversal adversarial training: ICB response AUC jumps from 0.762 to 0.891 across 14 independent cohorts and 7 cancer types
• +24.7% DTW improvement over UniTVelo on trajectory reconstruction, +19.3% F1 over CIBERSORTx on cell-type deconvolution
• Perturbation Engine simulates counterfactual immune trajectories via cross-attention: +37.2% Pearson r over GEARS on Perturb-Seq tasks
• Inference at 14,200 cells/s — 2.8× faster than scGPT, 11.6× faster than UCE; deployable on a single RTX 3090 (6.2 GB)

RST2G is the first 4D framework that combines a residual-guided spatiotemporal transformer with a graph fusion mechanism for DCE-MRI breast tumor segmentation. It explicitly leverages pre-contrast MRI, post-contrast MRI, and their residual differences to capture fine-grained kinetic dynamics of tumors across the contrast-enhancement sequence — systematically outperforming 10 state-of-the-art 2D / 3D / 4D baselines on both Breast-MRI-NACT-Pilot and TCGA-BRCA public datasets.

Key Contributions

• CFormerEncoder: weight-sharing CNN + Vision Transformer dual-branch hybrid, capturing local texture and global semantics simultaneously
• Residual-guided Multi-Scale Refinement (MSR) module: uses the Pre/Post-contrast residual to recalibrate spatial attention
• Spatiotemporal graph fusion: Inter-Slice Attention (ISA) for slice-space dependencies, Inter-Temporal Attention (ITA) for contrast-phase changes
• Breast-MRI-NACT-Pilot: DSC 61.8% vs runner-up MultiResUNet 55.6% (+6.2 percentage points)
• TCGA-BRCA: DSC 80.1% vs runner-up MTLN 73.5% (+6.6 percentage points)
• Systematically outperforms 10 SOTA baselines across 2D / 3D / 4D, end-to-end trained on 2× V100 GPUs

PDF DOI