Ion channels are tunnel-shaped proteins that let charged atoms flow across membranes, and the textbook picture says a gate at the tunnel's mouth opens and closes to control that flow. A newer and less exploited picture is that the flow can be shut off deep inside the tunnel itself — by a membrane lipid whose tail pokes through a side window and physically plugs the pore, or by a small molecule that gets trapped in the cavity. The clearest Alzheimer's-relevant example is GolpHCat (TMEM87A), a channel sitting in the Golgi apparatus (the cell's protein-processing and shipping station) of neurons and astrocytes. Work by GolpHCat (TMEM87A), a unique voltage-dependent cation channel in Golgi apparatus, contributes to Golgi-pH maintenance and hippocampus-dependent memory resolved its atomic structures (up to 3.1 Å; PDB 8HSI, 8HTT, 8KB4), showed that a phospholipid, phosphatidylethanolamine, sits inside the conduction pocket as an intrinsic plug, and found that deleting the channel in mice fragments the Golgi, strips sugar decorations off hippocampal proteins, and causes severe memory deficits — organellar pathology of exactly the kind now viewed as central to Alzheimer's disease.
The problem is that the only known blocker of this channel is gluconate, a sugar-acid metabolite with sub-micromolar potency ($\text{IC}_{50} \approx 0.10\ \mu\text{M}$ in that same paper) but no prospect of crossing the blood-brain barrier or surviving as a drug. Meanwhile, the machine-learning community has built generative lead-optimization systems — models that iteratively rewrite molecules to improve potency and drug-like properties — but has only ever pointed them at soluble enzymes and kinases. This proposal argues that the lipid-plugged pocket of GolpHCat is a specific, structurally defined, AD-relevant pore-blockade site, and that a lead-optimization pipeline built around it needs two ingredients no existing pipeline has: a physics-based oracle that understands lipid-plug displacement rather than flat docking, and functional enrichment analysis — the statistical method that asks which biological pathways shift together in disease tissue — used prospectively to set the therapeutic direction and a system-level reward, rather than after the fact to describe results.
Core Claim
The claim is threefold. First, the phosphatidylethanolamine binding pocket of GolpHCat is a druggable pore-blockade site for an organellar channel, and molecules that control plug occupancy — stabilizing or displacing it — can be designed, not just screened for. Second, because GO and KEGG pathway databases under-annotate recently characterized organellar channels (a known annotation bias of functional enrichment), a custom, rank-based gene-set analysis of Alzheimer's expression data is needed to decide the therapeutic sign (whether the diseased brain wants the channel more open or more blocked) and to supply a pathway-level reward term. Third, a preference-guided multi-turn reinforcement-learning optimizer — an agent that edits molecules over several turns and learns from preferences over whole trajectories — equipped with a molecular-dynamics plug-displacement oracle and an explicit cardiac-safety penalty, can produce central-nervous-system drug-like modulators where static screening cannot.
Research Program
The program has three coupled workstreams.
Workstream 1: Enrichment sets the target and the sign. Standard overrepresentation analysis, which tests whether a predefined gene list is over-represented among differentially expressed genes using Fisher's exact test, depends on arbitrary cutoffs and on curated databases that had not even heard of TMEM87A before 2024. We will instead build a custom "organellar ion homeostasis" gene set from the GolpHCat paper's own glycomics and proteomics readouts (Golgi pH maintenance, protein glycosylation, organellar cation transport) plus first neighbors in the Alzheimer's Knowledge Base, a graph of 118,902 nodes and 1,309,527 relationships built from 22 databases. We then run rank-based enrichment (no fold-change cutoff, so subtle coordinated shifts are visible) on AD single-cell and bulk expression datasets, following the methodological precedent of Polygenic enrichment distinguishes disease associations of individual cells in single-cell RNA-seq data, which linked cell-type-specific profiles to disease susceptibility, and Mechanistic Investigation of Exercise Interventions in Rodent Models of Alzheimer's Disease and Prospects for Clinical Translation, which used enrichment to map multi-target neuroprotective pathways in AD models. The output is twofold: a cell-type-resolved decision on whether the GolpHCat module looks like a driver or a compensation — the exact disentangling problem flagged for AD microglial cascades in Neurotoxic Microglial Activation via IFNγ‐Induced Nrf2 Reduction Exacerbating Alzheimer's Disease — and a pathway score $s_{\text{path}}(m)$ that ranks a candidate molecule by the predicted proximity of its targets to the disease module, inverting the post-hoc network-pharmacology style of Integrated Network Pharmacology, Single‐Cell Transcriptomics Unveil the Mechanistic Role of Morusin in Aortic Dissection and Multi-target modulation of the homologous recombination pathway by resveratrol promotes DNA damage repair in POI, which used enrichment to explain molecules rather than to steer them.
Workstream 2: A physics oracle for plug displacement. Static docking into the GolpHCat pocket is misleading because the pocket is occupied by a lipid whose displacement costs real energy: in the directly analogous inward-rectifier potassium channels, Ion currents through Kir potassium channels are gated by anionic lipids measured displacement barriers up to 41 kJ/mol, and A constricted opening in Kir channels does not impede potassium conduction showed that conduction physics cannot be read off the static pore geometry at all. We will run molecular dynamics on the cryo-EM structures to quantify phosphatidylethanolamine plug occupancy and its free-energy cost, calibrating the oracle so that gluconate's measured $0.10\ \mu\text{M}$ potency lands where it should. The oracle $s_{\text{plug}}(m)$ returns the predicted change in plug occupancy (or pore block, for stabilizers) when molecule $m$ occupies the pocket. This follows the lead of Computational Identification of Novel Kir6 Channel Inhibitors, which showed small molecules can productively target lipid-activation pockets in channels — but that work used one-shot virtual screening on a plasma-membrane channel with no disease context and no iterative optimization, which is precisely the gap we fill.
Workstream 3: Preference-guided optimization with a safety counter-objective. We adapt the multi-turn reinforcement-learning recipe of POLO: Preference-Guided Multi-Turn Reinforcement Learning for Lead Optimization, which framed lead optimization as a multi-turn decision process for a language-model agent and handled conflicting objectives with turn-level preferences rather than a single blended score — important because POLO's own authors report failure when objectives conflict, as they do here (potency wants lipophilicity; the brain and the heart do not). The agent is warm-started on the 500,000 molecular-modification pairs of the MolOptIns dataset, as in POLO, and optimizes
$$R(m) = s_{\text{plug}}(m) + \lambda_p\, s_{\text{path}}(m) - \lambda_h\, \text{hERG}(m), \quad \text{subject to } \text{QED}(m) \ge 0.5 \text{ and } \text{BBB}(m) \ge b_0,$$
where $s_{\text{plug}}$ is the Workstream-2 oracle, $s_{\text{path}}$ the Workstream-1 pathway score, hERG a predicted cardiac potassium-channel block liability, QED a quantitative estimate of drug-likeness, BBB predicted blood-brain-barrier permeability, and the $\lambda$ weights tuned on held-out objectives. The hERG term matters because deep hydrophobic pore cavities are exactly what cardiac hERG channels present, and pore-block cross-reactivity there causes fatal arrhythmias; Therapeutics Data Commons provides the hERG endpoint and 65 other AI-ready tasks for this filter, and Time Is a Critical Factor When Evaluating Oligonucleotide Therapeutics in hERG Assays shows how easily hERG measurements mislead if done carelessly. For scaffold diversity we also compare against ScaffoldGVAE: scaffold generation and hopping of drug molecules via a variational autoencoder based on multi-view graph neural networks, which decouples scaffold from side-chain representations to escape the "forgetting problem" of fingerprint models; its generator is our diversity engine, its RL agent our efficiency engine.
Evaluation Plan
We first reproduce the published reference points of the methods we build on, on their own benchmarks: POLO reports an 84% average success rate on single-property tasks using only 500 oracle calls, a 2.3× improvement over baselines, and 50% success on simultaneous three-property optimization on ZINC-250k oracles; ScaffoldGVAE reports a hop rate near 100% and a 92.4% success rate on LRRK2 with MM/GBSA binding energies between $-40$ and $-60$ kcal/mol. Reproducing these within a few points establishes that our reimplementations are faithful. The main experiment is the GolpHCat plug task: all methods get the same oracle budget (500 MD-based $s_{\text{plug}}$ calls, equal compute) and the same property oracles, and are scored on multi-objective success rate (plug score, QED, BBB, hERG below threshold), scaffold hop rate, and chemical novelty. Because generative models are easily gamed when the oracle is a surrogate — a limitation POLO itself notes — every headline number is followed by re-scoring the top molecules with longer, independently seeded MD, and the best few go to wet-lab validation: electrophysiology on GolpHCat with gluconate's $0.10\ \mu\text{M}$ as the in-house anchor, a Golgi-pH assay, glycosylation-rescue in TMEM87A-deficient cellular models, and a hERG patch-clamp counter-screen. We also evaluate generalization honestly, since the Therapeutics Data Commons benchmarking showed drug-property models dropping from roughly 0.7 to roughly 0.4 Pearson correlation between random and realistic scaffold/temporal splits; our plug predictors are evaluated under scaffold-split as well, with that gap as the expected failure mode. For the hERG predictor itself, no reported result found on Lacuna for a baseline hERG ROC-AUC, so we will establish our own split-validated reference rather than compare to a literature number.
Baselines
The naive reference point is gluconate itself, the only known GolpHCat blocker, with a reported $\text{IC}_{50} \approx 0.10\ \mu\text{M}$ from GolpHCat (TMEM87A). It anchors potency but fails every other axis: it is an open-chain polyhydroxy anion with no blood-brain-barrier prospect. Beating this baseline does not require higher potency; it requires sub-micromolar plug activity together with CNS-drug-like properties and a clean hERG profile, which gluconate trivially cannot meet on the latter counts.
The strongest existing method is POLO, which reports 84% average success on single-property tasks with only 500 oracle calls and 50% on three-property optimization on ZINC-250k oracles. It is the strongest competitor because it is the current best sample-efficient multi-objective optimizer and because our proposal is a direct modification of it; the comparison is fair by construction — POLO runs unmodified on our plug oracle with the same 500-call budget, the same property oracles, and the same tuning effort — and we beat it only if adding the enrichment reward and preference decomposition for the hERG conflict raises multi-objective success (potency + BBB + drug-likeness + hERG) at equal or fewer oracle calls. ScaffoldGVAE is the second competitor on the scaffold-hopping axis specifically, with its near-100% hop rate and 92.4% LRRK2 success rate; it gets the same oracle budget and is beaten on hop-into-potency, i.e., novel scaffolds that retain plug activity.
The matched-compute docking baseline is a large-library structure-based virtual screen of the plug pocket in the Glide style demonstrated by Computational Discovery of Potent Nucleoprotein Inhibitors for Influenza A Virus, run with the same total compute as the full RL pipeline and allowed as many docked compounds as that budget permits; this is the honest test of whether iteration beats brute force on this pocket. Finally, two ablations isolate our key ingredients: physics-only (plug oracle and properties, no $s_{\text{path}}$ term) isolates the enrichment contribution, and static-docking-only (a flat docking score replacing the MD plug oracle) isolates the physics contribution. If either ablation matches the full method, that ingredient is not earning its cost and we will say so.
Why not just…?
Why not just repurpose approved pore blockers like memantine or ketamine? These are real drugs with real mechanisms — ketamine is a trapped open-channel blocker of the NMDA receptor whose target occupancy outlasts its 13-minute plasma half-life, as Sustained antidepressant effect of ketamine through NMDAR trapping in the LHb showed — but they act on plasma-membrane synaptic channels and do nothing for Golgi ion homeostasis or Golgi fragmentation, which is the pathology the GolpHCat paper ties to memory. Knowledge-graph repurposing does not escape this: AlzKB used RotatE link prediction to propose ten repurposed candidates with 30% concordance with existing AD trials, but it can only recommend drugs that already exist, and no approved drug targets an intracellular lipid plug. The threat is genuine for symptomatic relief, so we will run the decisive check: dock the AlzKB-predicted candidates into the GolpHCat plug pocket and, for any plausible hit, predict BBB permeability. If one binds well and is brain-penetrant, repurposing wins on cost and speed, and we would report that.
Why not just run a huge docking screen instead of reinforcement learning? Two cited lines of evidence say screening alone falls short here. First, the pocket is lipid-occupied: the Kir work measured up to 41 kJ/mol barriers for lipid-tail displacement, which static docking scores cannot represent, so a screen would rank molecules against the wrong geometry. Second, POLO showed that single-turn approaches need roughly 2.3× more oracle calls to reach comparable success, and MD-based calls are far too expensive to waste at screening throughput. But the honest answer is that this is an empirical question, which is exactly why the matched-compute docking baseline is in the plan; if screening finds plug-active, brain-penetrant, hERG-clean molecules within the same compute budget, the RL machinery is not earning its cost and the proposal's practical contribution collapses to the physics oracle and the gene set.
Why not just run standard GO/KEGG enrichment after picking the compounds? Post-hoc enrichment is what every network-pharmacology paper in the file already does, and it cannot help here for three reasons grounded in the literature: curated databases did not annotate TMEM87A as a channel until 2024, so a standard GO/KEGG analysis would not see the target at all; cutoff-based overrepresentation discards the subtle coordinated shifts across channel and transporter complexes that rank-based analysis preserves; and description after the fact cannot set the therapeutic sign — whether to stabilize or displace the plug — before the optimization budget is spent, nor can it distinguish a driver from a compensatory response in the diseased tissue. We also take our own medicine here: the physics-only ablation tests whether the enrichment reward actually changes the generated molecules. If it does not, the enrichment component belongs in the analysis section, not the objective, and we will write it that way.
Decisive Checks
Three results would make or break this proposal. The enrichment analysis must show a coordinate, cell-type-specific shift of the organellar ion-homeostasis module in AD tissue that points clearly toward plug stabilization or displacement; ambiguity here means the sign is undecided and the pipeline optimizes toward an arbitrary goal. The matched-compute comparison must show the physics-plus-preference optimizer beating POLO-unmodified and beating the docking screen on multi-objective success at 500 oracle calls; if the screen ties, the contribution shrinks to the oracle. And at least one generated molecule must show sub-micromolar, MD-reproduced plug activity with predicted brain penetration and a clean hERG counter-screen in patch clamp — the first drug-like modulator of an organellar channel's lipid plug, and a concrete opening on an Alzheimer's pathway that no amyloid-directed strategy, including the $k_{\text{on}}$-optimized sequestration approach of Small-molecule sequestration of amyloid-β as a drug discovery strategy for Alzheimer's disease, has ever touched.
References
- GolpHCat (TMEM87A), a unique voltage-dependent cation channel in Golgi apparatus, contributes to Golgi-pH maintenance and hippocampus-dependent memory
- The Alzheimer’s Knowledge Base: A Knowledge Graph for Alzheimer Disease Research
- Polygenic enrichment distinguishes disease associations of individual cells in single-cell RNA-seq data
- Mechanistic Investigation of Exercise Interventions in Rodent Models of Alzheimer’s Disease and Prospects for Clinical Translation
- Neurotoxic Microglial Activation via IFNγ‐Induced Nrf2 Reduction Exacerbating Alzheimer's Disease
- Integrated Network Pharmacology, Single‐Cell Transcriptomics Unveil the Mechanistic Role of Morusin in Aortic Dissection
- Multi-target modulation of the homologous recombination pathway by resveratrol promotes DNA damage repair in POI: integrated network pharmacology, molecular dynamics simulation, and experimental validation
- Ion currents through Kir potassium channels are gated by anionic lipids
- A constricted opening in Kir channels does not impede potassium conduction
- Computational Identification of Novel Kir6 Channel Inhibitors
- POLO: Preference-Guided Multi-Turn Reinforcement Learning for Lead Optimization
- Therapeutics Data Commons: Machine Learning Datasets and Tasks for Therapeutics
- Time Is a Critical Factor When Evaluating Oligonucleotide Therapeutics in hERG Assays
- ScaffoldGVAE: scaffold generation and hopping of drug molecules via a variational autoencoder based on multi-view graph neural networks
- Computational Discovery of Potent Nucleoprotein Inhibitors for Influenza A Virus: Validation Through QM/MM Analysis and Experimental Binding Assays
- Sustained antidepressant effect of ketamine through NMDAR trapping in the LHb
- Small-molecule sequestration of amyloid-β as a drug discovery strategy for Alzheimer’s disease