A collective synthesis of AI systems that improve themselves — gated at every step by biological, statistical, and clinical checks. AAIC 2026, London — the verification layer that sits beside the biology, not above it.
Because AAIC is a room full of MDs and biologists, we say the hardest thing first. If our claims cannot survive this sentence, none of the rest matters.
A collective synthesis of the 16 AAIC abstracts: AI systems that improve their own workflows, but only when new evidence survives biological, statistical, and clinical verification.
A new workflow is kept only if it improves external evidence without violating safety constraints.
What gets improved: prompts · tools · memory · simulators · code · verifiers · analysis policies
What improves: physics-informed biomarker models, ODE-gated self-training, conservative RL/MPC world models.
Verifier: AT(N) cascade constraints, Z3/SMT safety checks, NAD⁺ homeostasis bands, coherence audits.
Translational output: earlier treatment windows, safer dosing schedules, calibrated disease-trajectory forecasts.
What improves: latent diffusion virtual cells, co-evolving perturbation planners, AI swarms, causal scratchpads, MCTS.
Verifier: GRN hierarchy, conformal uncertainty, execution traces, evidence retrieval, stability tests.
Translational output: DAM reversal hypotheses, rare resilience cell states, epistatic gene interactions, causal neuroinflammatory nodes.
What improves: discovery agents retrieve literature, write analysis code, run simulations, critique failures, and refine hypotheses.
Verifier: AD knowledge graphs, pathway databases, statistical execution feedback, human expert review.
Translational output: neuroprotective pathways, vascular-resilience targets, executable experimental designs.
What improves: biomarker fusion plus test-time self-verification for difficult, conflicting clinical profiles.
Verifier: pericyte/endothelial markers, p-tau217, imaging, APOE4, expert-like contradiction checking.
Translational output: vascular-dominant AD subtype detection and more reliable atypical-case diagnosis.
What improves: generative chemistry and RL from verifiable rewards for NAMPT modulators and tau inhibitors.
Verifier: BBB permeability, CNS MPO, docking, lipophilicity limits, cytotoxicity filters.
Translational output: brain-penetrant lead libraries with improved developability before wet-lab validation.
Self-improving AI belongs where feedback is external, falsifiable, and reversible. Internal self-critique is a speedup, not scientific closure; promotion requires biological plausibility, execution success, safety bounds, provenance, and rollback.
Four of our AAIC posters build on one of the most inspiring pieces of AD biology we have read this year — Chaubey et al., "Pharmacologic reversal of advanced Alzheimer's disease in mice," Cell Reports Medicine 2026, from the Pieper, Paul, and Kang labs. Restoration of brain NAD⁺ homeostasis via P7C3-A20; reversal of tau pathology and BBB damage; cognitive recovery in mouse models. It is the biology we hope our verification tools can serve.
Peer-reviewed by the AAIC 2026 program committee. AI-methods contributions. Every clinical or biological reference is fully attributed to its source.
1D temporal FNO under AT(N) cascade priors. Answer Set Programming plans, Z3 SMT-verified for cumulative-dose and toxicity limits.
Conditional latent diffusion with GRN priors for counterfactual cell-state simulation. Conformal prediction over multi-omic perturbation shifts.
Action-conditioned world model for NAD⁺ / tau / cognition. Conservative MPC. 47% reduction in cumulative drug burden vs continuous-dose baselines.
Monte Carlo Wavelet Coherence regularization. Detects "good MSE / bad simulator" failure mode in counterfactual rollouts.
Adversarial architecture — VCM simulates resistance while the planner learns to force the latent state back to homeostasis under toxicity penalties.
LLM discovery agent with iterative self-reflection over AD knowledge bases. Combinatorial target prioritization; reduced hallucination vs baselines.
sPDGFRβ + ANGPT-2 + p-tau217 stratification. APOE4-enriched vascular-dominant MCI. Neurovascular dimension added to the AT(N) framework.
CNS-MPO-guided generation targeting the NAMPT rear-channel binding pocket. Brain-penetrant candidates without off-target cytotoxicity.
Autonomous LLM agent with an execution environment on ADNI data. 84% analytical idea implementation rate; novel vascular-resilience node identified.
Recursive self-critique for AD subtyping. 34% accuracy improvement on atypical presentations by trading generation speed for structured self-correction.
Persistent-memory multi-agent framework on SEA-AD single-nucleus RNA-seq. 42% higher code-execution success rate; rare cognitive-resilience microglial state.
Medical AI scientist with clinician-engineer co-reasoning. Zero-hallucination methodological proposals via knowledge-graph gating.
Symbolic verifier gates the recursive self-improvement loop against amyloid / tau kinetics ODEs. 63% long-horizon forecast improvement.
Momentum-driven evolutionary causal reasoning with a reflective scratchpad. Non-canonical lipid-metabolism / cytokine node discovered.
Closed-loop docking + BBB simulator provides scalar rewards. Test-time recursive thinking; 42% success-rate improvement on tau aggregation inhibitors.
Socratic self-refine over multi-hop AD genetics literature. MCTS-guided sub-question decomposition and evidence-retrieval-based logical-branch pruning.
The AAIC portfolio builds on prior peer-reviewed work published earlier this year. All are AI-methods contributions applied to Alzheimer's tasks; none involve wet-lab, in-vivo, or clinical work. All cite Chaubey et al. (Cell Reports Medicine 2026) as the biology anchor.
Full inventory: OpenReview · publications page
A dedicated non-profit research foundation is in incorporation as the long-term home for the platform. Until it is live, the work is supported by AIXC Bio MB. Details of jurisdiction and banking will be published once formally registered.
The campaign covers AAIC 2026 travel, poster printing, on-site meetings, and six months of infrastructure build. If you would like to be notified when it opens, please email us.
Quarterly written progress updates to every supporter. Statutory financial reporting begins once the foundation is registered.
Reversal Initiative is the Alzheimer's research programme of AIXC Bio — an EU-based computational-biology company (Zaragoza HQ, NVIDIA Inception member) focused on AI verification and evaluation for biomedical research. AIXC Bio covers the broader platform; this AAIC campaign is the Alzheimer-focused chapter of that work.