Microglia — the brain's resident immune cells — switch between repair states and damaging ones, and the same molecular dial appears to be turned in opposite directions in two diseases of the same organ. In glioblastoma, roughly 45% of tumors lose the gene MTAP, causing its substrate methylthioadenosine (MTA) to pour out into the tissue, where it binds the adenosine A2B receptor on macrophages and microglia and locks them into an immunosuppressive, tumor-protective state MTAP loss correlates with an immunosuppressive profile in GBM and its substrate MTA stimulates alternative macrophage polarization. In Alzheimer's disease, the inflammatory signal interferon-gamma pushes microglia the other way, into a neurotoxic state that accelerates amyloid and tau pathology Neurotoxic Microglial Activation via IFNγ‐Induced Nrf2 Reduction Exacerbating Alzheimer's Disease. Nobody has asked whether the tumor-derived MTA–A2B circuit also runs in the Alzheimer's brain, or in which direction it pushes microglia there. That question is the scientific core of this proposal.
The methodological core is what happens after you find a candidate A2B drug: it must reach the brain at a free, active concentration, and this is where Alzheimer's drug development keeps failing. The field's standard tools — potency assays in flat cell cultures, binary blood-brain-barrier classifiers, static docking scores — each ignore a piece of the problem, and the casualties are documented: the failed Alzheimer's candidate Tarenflurbil was predicted brain-penetrant by ordinary models and only flagged as unreliable once uncertainty estimation was added Blood-brain barrier penetration prediction enhanced by uncertainty estimation. We propose a single pipeline that couples all four directions: a computational screen that front-loads uncertainty-quantified pharmacokinetic (PK) prediction, a bidirectional microglial phenotyping assay, and a perfused three-dimensional human "mini-brain" that measures drug exposure and microglial state at the same time, so that only compounds predicted to be both active and brain-exposed advance to mice.
Core Claim
We claim two things. Biologically, we hypothesize that the MTA–A2B–STAT3 axis identified in glioblastoma MTAP loss correlates with an immunosuppressive profile in GBM and its substrate MTA stimulates alternative macrophage polarization is also active in Alzheimer's microglia, interacting with the interferon-gamma–Nrf2 circuit Neurotoxic Microglial Activation via IFNγ‐Induced Nrf2 Reduction Exacerbating Alzheimer's Disease, and that its effect is context-dependent: the same antagonist that breaks immunosuppression in a tumor could worsen neurotoxic inflammation in an Alzheimer's brain, or could instead relieve a suppressive state that impairs aggregate clearance. The direction is an empirical question, and our assay is built to answer it in both directions (antagonists and agonists).
Methodologically, we claim that ranking candidate compounds by a joint score — predicted A2B activity, high-confidence predicted blood-brain-barrier penetration, and predicted microsomal stability, with an explicit uncertainty penalty — produces a hit list in which a substantially larger fraction of phenotypically active compounds is also CNS-exposed in vivo, compared to screening on potency first and checking PK later. The mitophagy-discovery paper shows what happens without this: its cross-species machine-learning screen achieved a 44% in vitro hit rate (8 active compounds out of 18 tested, versus roughly 0.14% for conventional high-throughput screening), yet one of its two validated leads, Rhapontigenin, then failed on brain bioavailability Amelioration of Alzheimer’s disease pathology by mitophagy inducers identified via machine learning and a cross-species workflow. Half an excellent hit list was lost to a problem that was, in principle, predictable before any mouse was dosed.
Research Program
The computational workstream builds a joint PK model rather than a new architecture from scratch. We adopt the sequential multi-task structure of ADME-DL, which trains a molecular foundation model through the biological order Absorption → Distribution → Metabolism → Excretion across 21 Therapeutics Data Commons endpoints and uses gradient surgery to stop conflicting tasks from degrading each other ADME-drug-likeness: enriching molecular foundation models via pharmacokinetics-guided multi-task learning for drug-likeness prediction; ADME-DL stops at generic drug-likeness, and we extend it with a blood-brain-barrier head trained on the curated M-data benchmark (1,937 compounds) and the independent S-data clinical set (527 drugs) from the BBB-uncertainty study Blood-brain barrier penetration prediction enhanced by uncertainty estimation. Each prediction carries an evidential uncertainty in the style of TrustworthyMS, which couples dual-view graph encoders to a Beta-Binomial uncertainty output and showed that filtering uncertain predictions lifted out-of-distribution selective accuracy from 83.1% to 94.2% on metabolic stability Uncertainty-Aware Metabolic Stability Prediction with Dual-View Contrastive Learning. The screen itself is small and tractable: known and computationally enumerable A2B ligands plus the 3,274-compound Macau Natural Product Library used by the mitophagy study Amelioration of Alzheimer’s disease pathology by mitophagy inducers identified via machine learning and a cross-species workflow, scored on the joint activity-PK-uncertainty objective.
The biological workstream first establishes whether the axis exists in Alzheimer's-relevant cells. We dose human induced-pluripotent-stem-cell-derived microglia and BV-2 microglia with MTA, with interferon-gamma, and with both, and measure the markers that define each state across the two parent papers: Nrf2, CD86, CD11b, nitric oxide and reactive oxygen species (the Alzheimer's neurotoxic signature) and Arg1, VEGFa, STAT3 and CREB phosphorylation (the glioblastoma immunosuppressive signature). We then test whether selective A2B antagonists — the same tool compounds that fully blocked MTA-driven polarization in the GBM study's cell assays MTAP loss correlates with an immunosuppressive profile in GBM and its substrate MTA stimulates alternative macrophage polarization — and A2B agonists shift the phenotype in the Alzheimer's context, and whether microglial Nrf2 overexpression (rescue in the mini-brain paper) changes A2B responsiveness. The decisive output is a directionality map: which interventions help, hurt, or do nothing, on which markers, at which concentrations.
The translational workstream is where the in vitro-to-in vivo gap is attacked directly. We adapt the 3D human Alzheimer's mini-brain microfluidic platform — a co-culture of Alzheimer's-derived neurons, reactive astrocytes, and microglia in a hydrogel that reproduces amyloid deposition, tau phosphorylation, and synaptic loss, and which was cross-validated against post-mortem Alzheimer's tissue and 5XFAD mice Neurotoxic Microglial Activation via IFNγ‐Induced Nrf2 Reduction Exacerbating Alzheimer's Disease — into a dynamic PK/PD instrument. The key modification is perfusion with inline sampling: instead of a static bolus whose concentration is effectively infinite, compound flows through the ECM-like matrix at a controlled rate with measured inlet and outlet free concentrations, giving a per-compound exposure profile inside the model. We then read out the microglial phenotype and pathology markers against the measured free exposure, not the nominal dose. The platform's predictions — free exposure and phenotype effect — are checked against 5XFAD mouse brain unbound partitioning and microglial markers for the top compounds. This is what no platform in the file does: the mini-brain paper performs no pharmacokinetics, and the macrophage-reprogramming polypeptide work De Novo Design and Synthesis of Polypeptide Immunomodulators for Resetting Macrophage Polarization is static 2D culture with an entirely uncharacterized in vivo disposition.
Proposed Experiments
The model evaluation is against published numbers on the cited benchmarks. On the HLM metabolic-stability benchmark, TrustworthyMS reports an F1-score of 0.866 and Matthews Correlation Coefficient of 0.622 Uncertainty-Aware Metabolic Stability Prediction with Dual-View Contrastive Learning; our joint model should match these within noise while simultaneously predicting BBB penetration, since multi-task training should not cost accuracy on either head. On M-data, GROVER-BBBp with uncertainty filtering reached an MCC approaching 1.0 on the top 20% most certain predictions, and over 99% accuracy when the uncertainty score was below 0.1 Blood-brain barrier penetration prediction enhanced by uncertainty estimation; we reproduce this selective-classification operating point and additionally require the model to assign Tarenflurbil an uncertainty above 0.5 and Niraparib a low-uncertainty brain-penetrant call, as that paper did, before we trust it for screening. Screening yield is measured as in vitro hit rate (phenotypic modulation in the 2D assay) and, critically, as the fraction of hits retaining activity at measured free concentrations in the perfused mini-brain — the number the mitophagy study could not achieve for Rhapontigenin. For in vivo, the primary metric is the rank correlation between mini-brain measured free exposure and mouse brain unbound partition coefficients across compounds; a strong correlation is the evidence that the perfused model is a genuine translational bridge rather than another static assay. No reported result found on Lacuna for measured unbound brain partitioning of any A2B antagonist, which is itself part of the gap.
Baselines
The naive reference point is unfiltered phenotypic screening: test the same library in the 2D microglia assay with no computational pre-ranking, at fixed high concentrations, exactly as a conventional high-throughput campaign would. The file's reference for what this yields is the roughly 0.14% active hit rate of traditional screening quoted in the mitophagy paper Amelioration of Alzheimer’s disease pathology by mitophagy inducers identified via machine learning and a cross-species workflow. Our pipeline beats this baseline if its hit rate is materially higher and, more importantly, if its hits survive the perfused-model exposure test that naive hits largely fail.
The strongest existing competitor is the cross-species phenotypic machine-learning workflow itself Amelioration of Alzheimer’s disease pathology by mitophagy inducers identified via machine learning and a cross-species workflow, which achieved a 44% hit rate (8 of 18) by multi-representation similarity to known actives, followed by worm and mouse validation. We will run that workflow unchanged on our library with the same compute and tuning budget, and compare on the joint endpoint: hit rate multiplied by the fraction of hits that are brain-exposed in vivo. That paper's own Rhapontigenin failure defines the expected weakness; our pipeline wins only if it closes the exposure gap without sacrificing comparable hit rate. If its joint yield matches ours, the PK machinery is not earning its cost and we should say so.
The strongest PK-specific competitor is GROVER-BBBp with uncertainty filtering Blood-brain barrier penetration prediction enhanced by uncertainty estimation, run as a standalone filter in front of the same 2D assay, tuned on M-data/S-data with the same budget. It has the reported selective-classification performance quoted above; it lacks any metabolic-stability or joint-activity component, so the comparison isolates whether the sequential multi-task structure adds anything beyond a good BBB filter.
Finally, an ablation removes the proposal's key ingredient: the same joint pipeline run with point estimates only, uncertainty disabled. The expected signature of failure is Tarenflurbil-style false positives — compounds the pipeline passes that the uncertainty-aware version correctly flags — and a lower fraction of in vivo-exposed hits.
Why not just…?
Why not just screen A2B ligands in standard 2D microglia cultures? This is the cheapest path and would answer the directionality question at the cell level. What it cannot answer is the translation question: the GBM paper's own A2B-antagonist rescue experiments were done exactly this way, in 2D cultures, with no in vivo or PK follow-up MTAP loss correlates with an immunosuppressive profile in GBM and its substrate MTA stimulates alternative macrophage polarization], and the polypeptide immunomodulator study explicitly leaves clearance, biodistribution, and barrier penetration uncharacterized for the same reason De Novo Design and Synthesis of Polypeptide Immunomodulators for Resetting Macrophage Polarization. A static well holds drug at a constant, unlimited concentration; a brain does not. We include the 2D assay as the first tier of our own pipeline — it is necessary — but a 2D-only result would recreate, not close, the gap.
Why not just dock compounds against the A2B receptor? Docking scores capture static geometric and electrostatic fit in a binding pocket and ignore metabolic stability, permeability, and systemic survival — a disconnect documented directly in the multi-omics docking study Integrating Clinical Multi-Omics, Machine Learning, and Molecular Docking Strategies to Reveal the Mechanism of Yiqi Huoxue Prescription in the Treatment of Ischemic Heart Failure and illustrated by kinase inhibitors that docked beautifully and cleared too fast in vivo A virtual screen identified C96 as a novel inhibitor of phosphatidylinositol 3-kinase that displays potent preclinical activity against multiple myeloma in vitro and in vivo. Moreover, our central biological question is not binding affinity at A2B but the direction of the phenotypic shift in a specific diseased tissue, which no docking score speaks to at all.
Why not just apply the existing BBB predictor as a filter and be done? This is the most serious alternative, and we treat it as a baseline rather than dismissing it. Two things limit it. First, it predicts a binary total brain-to-plasma ratio, not the unbound partition coefficient that drives target engagement, and not metabolic stability or activity — so it filters on one axis of three. Second, its own paper's "DL Hype Check" showed that on the uncurated real-world S-data set, deep models performed no better than random forests on simple descriptors, and that only uncertainty filtering rescued reliability Blood-brain barrier penetration prediction enhanced by uncertainty estimation; TDC's benchmarking across 21 methods tells the same story, with drug-target interaction correlation falling from about 0.7 to about 0.4 under realistic out-of-distribution splits Therapeutics Data Commons: Machine Learning Datasets and Tasks for Drug Discovery and Development. If the standalone-filter baseline matches our joint pipeline on the combined yield metric, the honest conclusion is that the multi-task machinery is unnecessary for this compound class, and the experiments as designed will reveal that.
Why not just run the cross-species phenotypic workflow and check PK afterwards? This is precisely the ordering we argue against, and the evidence is the mitophagy study itself: a 44% hit rate, then a 50% loss of validated leads to CNS bioavailability failure discovered only after worm and mouse work Amelioration of Alzheimer’s disease pathology by mitophagy inducers identified via machine learning and a cross-species workflow. Post-hoc PK checking spends the most expensive experiments on compounds that were predictable failures. Our design keeps that workflow as a baseline so the cost of the alternative ordering is measured, not asserted.
Decisive Checks
Three results would settle the proposal. First, the directionality map: a clean demonstration that MTA and interferon-gamma produce distinguishable, interacting microglial states in human Alzheimer's-relevant cells, and that A2B ligands shift them — in whichever direction — would establish the shared purine checkpoint as real beyond glioblastoma. Second, the yield comparison: our pipeline must beat the cross-species workflow baseline on hits that are both active and brain-exposed, not on hit rate alone, with the uncertainty-off ablation showing where the advantage comes from. Third, the bridge test: rank correlation between measured free exposure in the perfused mini-brain and unbound brain partitioning in mice, across the advancing compounds. If that correlation is strong, the perfused organotypic model earns its place as the first in vitro platform in this literature that predicts in vivo brain exposure of immunomodulators before they enter an animal; if it is weak, the field learns that too, and earlier than it currently does.
References
- MTAP loss correlates with an immunosuppressive profile in GBM and its substrate MTA stimulates alternative macrophage polarization
- Neurotoxic Microglial Activation via IFNγ‐Induced Nrf2 Reduction Exacerbating Alzheimer's Disease
- Blood-brain barrier penetration prediction enhanced by uncertainty estimation
- Amelioration of Alzheimer’s disease pathology by mitophagy inducers identified via machine learning and a cross-species workflow
- ADME-drug-likeness: enriching molecular foundation models via pharmacokinetics-guided multi-task learning for drug-likeness prediction
- Uncertainty-Aware Metabolic Stability Prediction with Dual-View Contrastive Learning
- De Novo Design and Synthesis of Polypeptide Immunomodulators for Resetting Macrophage Polarization
- Integrating Clinical Multi-Omics, Machine Learning, and Molecular Docking Strategies to Reveal the Mechanism of Yiqi Huoxue Prescription in the Treatment of Ischemic Heart Failure
- A virtual screen identified C96 as a novel inhibitor of phosphatidylinositol 3-kinase that displays potent preclinical activity against multiple myelomain vitroandin vivo
- Therapeutics Data Commons: Machine Learning Datasets and Tasks for Drug Discovery and Development