SPARK Agentic AI in Cancer Pathology & AI Drug/Med Momentum
Key Questions
What is the SPARK project at U Cologne and its main hypothesis?
SPARK uses agentic AI to hypothesize from pathology slides and predict immunotherapy responses, validated on 5,400 patients in Nature Medicine.
How does AutoMedBench evaluate medical auto-research agents?
AutoMedBench benchmarks agentic AI models for medical research, with the Validate stage showing the weakest performance at 37.7% verification errors.
What breakthrough did Biohub achieve with AI-designed proteins?
Biohub's world model designed real proteins with 4.3 nM binding affinity and 9.4-second structure prediction, later validated in the lab.
What does the large-scale connectome study reveal about brain aging?
Analyzing 54k scans, it produced normative white matter charts showing a 'last in first out' pattern of aging in brain connectivity.
How is privacy preserved in federated learning for medical imaging?
Techniques like CasLB+ReG-CDN+Hy-COXdCN achieve 98.5% accuracy on KITS while preserving privacy in distributed medical imaging models.
U Cologne SPARK hypothesizes from slides, predicts immunotherapy (Nature Med, 5,400 patients). New: ML doubles depression remission (iMAP/wearables); GNN counterfactuals; biomedical replication crisis (97% invalid stats); Insilico Ph2a rentosertib; Mila/TandemAI world models for drug discovery. Added: Flatiron Health VALID framework for AI-curated oncology data; generative AI for high-stakes decisions (One Health, flow matching, LLM agents). New this run: DANCE for EEG event detection; AutoScientists self-organizing agent teams for scientific experimentation (BioML-Bench, GPT training, ProteinGym); privacy-preserved FL for medical imaging (CasLB+ReG-CDN+Hy-COXdCN, 98.5% on KITS). Also noted: UK-France AI collaboration for women's health using advanced imaging and £900m compute funding. New: Biohub world model designs real proteins (4.3 nM binding, 9.4s structure prediction, lab-validated); AutoMedBench benchmark for medical auto-research agents (Validate stage weakest, 37.7% verification errors); Large-scale connectome study (54k scans, normative white matter charts, 'last in first out' aging).