Reproduce “TwinWeaver: An LLM-Based Foundation Model Framework for Pan-Cancer Digital Twins” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Nikita Makarov, Maria Bordukova, Lena Voith von Voithenberg, Estrella Pivel-Villanueva, Sabrina Mielke, Jonathan Wickes, Hanchen Wang, Mingyu Derek Ma and 6 more
TwinWeaver uses text-serialized patient histories to train Genie Digital Twin (GDT), a language-model framework that jointly forecasts laboratory values and predicts clinical events from longitudinal oncology records. On retrospective data from 93,054 patients across 20 cancer indications, it outperformed leading baselines on laboratory forecasting and several event-prediction tasks, though metastasis prediction was weaker and clinical use remains unvalidated.
Key insights
- A shared language-model representation can support both numerical trajectory forecasting and clinical event prediction without separating them into distinct modeling pipelines.
- Longer patient histories improved survival discrimination; using the full available history raised the one-year C-index by 9.9% compared with baseline-only information.
- The text-generation approach also enabled explanatory outputs, but these are hypothesis-generating rather than verified as faithful accounts of the model’s reasoning.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.