Many of the most promising drug candidates for the central nervous system come from plants and microbes, but they fail twice over before reaching a clinic. First, the active molecules occur in such tiny amounts — often milligrams from kilograms of wild biomass — that there is no material for animal studies, and isolating minor constituents kills preclinical pipelines outright, as documented for prenylated flavonoids from Macaranga adenantha Minor Prenylated Flavonoids from the Twigs of Macaranga adenantha and Their Cytotoxic Activity. Second, the bulky, heavily glycosylated or polyphenolic forms that plants actually produce in quantity are almost completely excluded by the blood-brain barrier (BBB) — the tight layer of brain capillary cells that blocks some 98% of small molecules — even when their aglycone cores show potent neuroactivity, as shown for tannin-rich plant metabolites Determination of the profile of metabolites of medicinal plants rich in tannins in the context of their antineurodegenerative activity. Meanwhile, enzymes can already relieve the scarcity problem: specialized glycosidases snip terminal sugars from abundant precursors to make rare actives Purification and Characterization of Ginsenoside-.BETA.-Glucosidase from Ginseng., Purification and Characterization of Ginsenoside-.ALPHA.-L-Rhamnosidase., and engineered yeast now produces gram-per-liter titers of previously unobtainable natural products Synthesizing ginsenoside Rh2 in Saccharomyces cerevisiae cell factory at high-efficiency.
This proposal builds a search system that, starting from a plant's abundant metabolites, plans short enzymatic transformation cascades whose products are predicted to cross the BBB. The intuition: nature left the scarce molecules in the same chemical neighborhood as the abundant ones, and the edits that connect them — removing a sugar, adding a methyl, closing a ring — are exactly the edits enzymes perform with high specificity. No existing system searches this enzyme-reachable neighborhood under a brain-permeability objective; current BBB predictors only classify molecules that already exist.
Core Claim
Given a library of abundant natural product precursors (glycosylated ginsenosides, polyphenols, sesquiterpene and sesterterpene scaffolds) and a catalogue of enzymatic transformations, a search that jointly models transformation feasibility and BBB permeability will find short, scalable routes to rare metabolites predicted to be brain-penetrant — molecules that are simultaneously scarce and BBB-accessible today, and that neither supply-focused biocatalysis nor BBB-focused prediction can produce alone.
The system scores each candidate product $m$, reachable from an abundant precursor $p$ through reaction $r$, as
$$ S(m) = P_{\text{BBB}}(m) \cdot P_{\text{enz}}(r) \cdot A(p), $$
where $P_{\text{BBB}}(m)$ is an uncertainty-gated permeability probability, $P_{\text{enz}}(r)$ is the likelihood that a real enzyme catalyzes $p \to m$, and $A(p)$ is the precursor's extractable abundance. The search maximizes $S$ over one- to three-step enzymatic cascades rather than over arbitrary chemical space, which is what makes both the prediction problem (stay near natural product geometry) and the supply problem (start from something you can harvest by the kilogram) tractable.
Why the Gap Is Real
Three tool families exist separately, and none is coupled. BBB prediction has progressed from a MegaMolBART transformer plus XGBoost classifier achieving AUC 0.88 on held-out data and validated on 3D human BBB spheroids, including a natural product library Predicting blood–brain barrier permeability of molecules with a large language model and machine learning, to relational graph networks that model influx and efflux transporters and reach AUROC 0.926 Relational graph convolutional networks for predicting blood–brain barrier penetration of drug molecules. But both consume static molecules; neither generates or plans how to make a permeable molecule. Enzyme prediction goes the other way: BEC-Pred, a BERT model fine-tuned on 56,512 curated enzymatic reactions, classifies transformations across 308 EC sub-subclasses at 91.6% accuracy at EC level 3 A general model for predicting enzyme functions based on enzymatic reactions — but it has never been aimed at a pharmacokinetic objective. And generative retrosynthesis by GFlowNets reaches 63.6% round-trip accuracy on USPTO-50k Diverse and feasible retrosynthesis using GFlowNets — but over organic-chemistry disconnections, with no notion of permeability or of biological production. The closest existing work to our intersection is the MegaMolBART study, which explicitly shows (via t-SNE) that natural products occupy a distinct region of chemical space and correctly classifies 26 natural products against spheroid assays — but it treats natural products as static evaluation inputs from an existing physical repository, with no coupling to biosynthesis, remodeling, or scaling. Our proposal inverts that: the natural product library becomes the starting material of a generative search, and the BBB model becomes the objective function.
Two further facts from the literature shape the design. MoleculeNet showed that scaffold splitting sharply degrades performance, meaning models memorize motifs rather than transport rules MoleculeNet: A Benchmark for Molecular Machine Learning, so any predictor must be co-trained on natural product data and evaluated on scaffold-disjoint splits. And the "DL hype check" of Blood-brain barrier penetration prediction enhanced by uncertainty estimation found that uncalibrated deep models performed no better than simple physicochemical baselines on an independent clinical set of 527 molecules, with reliability restored only by uncertainty filtering — so our $P_{\text{BBB}}$ must be uncertainty-gated, not just deep.
Research Program
We will build three components and close the loop with a small experimental campaign. First, the transformation engine: we compile a reaction catalogue from characterized glycosidases Purification and Characterization of Ginsenoside-.BETA.-Glucosidase from Ginseng., Purification and Characterization of Ginsenoside-.ALPHA.-L-Rhamnosidase., promiscuous plant cyclases such as CbTPS1, which converts one universal substrate into 20 sesterterpenes across ten ring architectures From Monocyclization to Pentacyclization: A Versatile Plant Cyclase Produces Diverse Sesterterpenes with Anti‐Liver Fibrosis Potential, and the ECREACT dataset of 56,512 enzyme reactions; BEC-Pred-style classification then assigns plausible enzyme classes to each proposed substrate–product pair, giving $P_{\text{enz}}(r)$. We will be explicit about the known blind spot: BEC-Pred struggles with isomerases and intramolecular rearrangements, so ring-forming steps will be weighted conservatively and validated first. Second, the permeability scorer: a transporter-aware graph model in the spirit of the RGCN, retrained with natural product data (co-training on NP-subset data is exactly what the MegaMolBART study found necessary to avoid overfitting on the NP manifold), and wrapped in the Shannon-entropy-plus-dropout uncertainty gate of the GROVER-uncertainty work, so that only confident predictions enter $S(m)$. Third, the search: a beam search over the transformation graph from each abundant precursor, with 3D-conformation-aware representations (which Deep Molecular Representation Learning via Fusing Physical and Chemical Information showed improve molecular property prediction without experimental structures) to respect the stereochemical density of natural products.
The output is a ranked list of (abundant precursor, enzymatic steps, predicted-permeable product) triples, each carrying the uncertainty of every prediction in the chain. The top candidates are then made and tested: the enzymatic trimming steps are bench-tractable (isolated glycosidases under mild aqueous conditions), and for any hit that needs supply beyond enzymatic conversion of extracted precursor, the heterologous-expression playbook of the Rh2 yeast work — modular pathway optimization plus directed evolution of the terminal enzyme, which reached 2.25 g/L — provides the scaling route Synthesizing ginsenoside Rh2 in Saccharomyces cerevisiae cell factory at high-efficiency.
Baselines
The naive reference point is a physicochemical filter: rank all enzymatic products of the abundant precursors by a CNS property rule (molecular weight under ~450, few hydrogen-bond donors, moderate calculated logP). This is the "do nothing" of learned prediction, and it is not a strawman — on the independent clinical benchmark of 527 molecules, simple property models matched uncalibrated deep models Blood-brain barrier penetration prediction enhanced by uncertainty estimation. We run it on the same product library, with the same downstream spheroid validation, and treat our pipeline as beating it only if it retrieves a strictly better confirmed-permeable set at equal or smaller synthesis cost.
The strongest existing competitor on the prediction side is the transporter-aware relational graph network of Relational graph convolutional networks for predicting blood–brain barrier penetration of drug molecules, which reports AUROC 0.926 and AUPRC 0.865 with Tanimoto drug-similarity edges and 0.919 / 0.838 with Mordred descriptors alone, on a 70/15/15 split of literature-curated drugs. We will rerun this architecture as a plug-in $P_{\text{BBB}}$ inside our own search, trained on the same training pool we use, with identical tuning budget and scaffold-disjoint natural-product test folds, so that any difference is attributable to natural-product-domain co-training and uncertainty gating rather than to architecture or compute. The MegaMolBART–XGBoost system, at AUC 0.88 on held-out data, is the second learned baseline, run the same way. Beating these means higher AUROC on the NP-specific, scaffold-split test folds and — the decisive criterion — a higher fraction of spheroid-confirmed permeable products among the top-ranked candidates than the baseline scorers retrieve over the identical enzymatic product library.
The key ablation isolates the enzymatic-reachability ingredient: replace the enzyme catalogue with a generic chemical retrosynthetic planner (RetroGFN, 63.6% round-trip accuracy at k=50 on USPTO-50k Diverse and feasible retrosynthesis using GFlowNets) whose routes are filtered by the same BBB scorer, keeping data, compute, and ranking budget fixed. If enzyme-restricted search retrieves more confirmed-permeable, synthetically realizable products than chemistry-restricted search, the central claim — that the biosynthetic neighborhood is the right search space — is supported. A second ablation removes the uncertainty gate and the NP co-training from $P_{\text{BBB}}$, to confirm on our own folds the degradation that the DL-hype-check paper predicts. For completeness we note that on the original MoleculeNet BBBP benchmark itself, no reported result found on Lacuna for a MoleculeNet baseline BBBP number in the file's text snippets, which is why we anchor to the refined M-data (1,937 molecules) and S-data (527 molecules) benchmarks of the uncertainty study, where the top-20%-most-certain predictions reached an MCC approaching 1.0 and GROVER-BBBp exceeded 99% accuracy when uncertainty fell below 0.1.
Why not just…?
Why not just use simple physicochemical filters? This is a genuine threat, and the honest answer is that on uncalibrated prediction the filter is competitive Blood-brain barrier penetration prediction enhanced by uncertainty estimation. Two things decide it. Transporter biology: the RGCN work shows passive descriptors fail specifically for molecules that are SLC or ABC transporter substrates, outperforming a LightGBM property baseline at AUROC 0.926 Relational graph convolutional networks for predicting blood–brain barrier penetration of drug molecules, and glycosylated natural products are classic efflux substrates. And empirically: the filter is one of our baselines, run on the identical library with identical spheroid validation; if it matches us, the pipeline's value collapses to its supply component and we will say so.
Why not just formulate the scarce natural product into a nanoparticle or cell carrier instead of redesigning the molecule? Delivery vehicles work — squalenoylated temozolomide particles reverse resistance A Novel Squalenoylated Temozolomide Nanoparticle with Long Circulating Properties Reverses Drug Resistance in Glioblastoma, brain-targeted lipid nanoparticles show neuronal tropism AI-Validated Brain Targeted mRNA Lipid Nanoparticles with Neuronal Tropism, and macrophage carriers traffic across the barrier Exploiting macrophages as targeted carrier to guide nanoparticles into glioma — but every one of them requires gram quantities of purified active ingredient for formulation development and pharmacokinetics, exactly what wild extraction cannot supply at milligram isolation yields. The two approaches are complements, not rivals: our pipeline is the upstream supply that formulation science needs, and our search can additionally avoid candidates whose permeability depends on a vehicle that never clears. If a vehicle route reaches the clinic faster for a scaffold we produce, that is a success of the supply half of this proposal.
Why not just run a standard retrosynthetic planner with a BBB score appended? This is the closest alternative and our main ablation. Chemical routes to complex natural products suffer multi-step yield losses, protecting-group burden, and solubility problems, as reviewed for indirubin analogues Synthetic progress and anti-leukemic research on indirubin analogues: a review since 2010, whereas glycosidases operate in water with absolute regioselectivity Purification and Characterization of Ginsenoside-.BETA.-Glucosidase from Ginseng. and fermentation reaches gram-per-liter titers Synthesizing ginsenoside Rh2 in Saccharomyces cerevisiae cell factory at high-efficiency. But the honest version of the objection is not about chemistry — it is that RetroGFN plus a BBB filter is functionally our pipeline with a different reaction catalogue. That is why the ablation is decisive rather than rhetorical: both searches run on the same precursors, scorer, and validation, and the comparison measures whether enzyme-reachable space is genuinely the richer neighborhood for this objective.
Proposed Experiments and Decisive Checks
The computational evaluation runs on three folds: M-data (1,937 molecules) and the independent S-data (527 molecules) from the uncertainty study, plus a scaffold-split natural-product fold built from CMUH-NPRL-style data, reporting AUROC, AUPRC, and uncertainty-gated accuracy against the baselines above. The generation-side evaluation measures what fraction of top-ranked enzyme-planned products are (a) predicted permeable with low uncertainty, (b) confirmed permeable in the tri-culture 3D BBB spheroid assay with LC-MS/MS readout — the same assay that confirmed 21 permeable and 5 impermeable compounds in the MegaMolBART study Predicting blood–brain barrier permeability of molecules with a large language model and machine learning — and (c) actually producible: the top five candidates are synthesized enzymatically from extracted abundant precursors, with yield and purity reported, and the single best candidate is pushed through a yeast production module as a feasibility demonstration. Success is a short list of rare, brain-penetrant natural product derivatives that neither a BBB model alone nor a biocatalysis pipeline alone would have produced — and a public, uncertainty-annotated ranking that lets others check the claim.
References
- Minor Prenylated Flavonoids from the Twigs of Macaranga adenantha and Their Cytotoxic Activity
- Determination of the profile of metabolites of medicinal plants rich in tannins in the context of their antineurodegenerative activity
- Purification and Characterization of Ginsenoside-.BETA.-Glucosidase from Ginseng.
- Purification and Characterization of Ginsenoside-.ALPHA.-L-Rhamnosidase.
- Synthesizing ginsenoside Rh2 in Saccharomyces cerevisiae cell factory at high-efficiency
- Predicting blood–brain barrier permeability of molecules with a large language model and machine learning
- Relational graph convolutional networks for predicting blood-brain barrier penetration of drug molecules
- A general model for predicting enzyme functions based on enzymatic reactions
- Diverse and feasible retrosynthesis using GFlowNets
- MoleculeNet: a benchmark for molecular machine learning
- Blood-brain barrier penetration prediction enhanced by uncertainty estimation
- From Monocyclization to Pentacyclization: A Versatile Plant Cyclase Produces Diverse Sesterterpenes with Anti‐Liver Fibrosis Potential
- Deep Molecular Representation Learning via Fusing Physical and Chemical Information
- A Novel Squalenoylated Temozolomide Nanoparticle with Long Circulating Properties Reverses Drug Resistance in Glioblastoma
- AI-Validated Brain Targeted mRNA Lipid Nanoparticles with Neuronal Tropism
- Exploiting macrophages as targeted carrier to guide nanoparticles into glioma
- Synthetic progress and anti-leukemic research on indirubin analogues: a review since 2010