Aggressive brain tumors feed themselves by switching their metabolism from oxygen-based energy production to a fermentative, glucose-guzzling mode — the same shift that lets a cell build biomass fast. The biological levers of that switch are increasingly well mapped: in glioma, a long non-coding RNA called PVT1 protects the glycolytic program by soaking up an inhibitory microRNA, and enzymes like pyruvate dehydrogenase kinase 1 (PDHK1) act as dams that divert sugar away from mitochondria. Yet the interventions that work in a dish fail in the animal. The small-molecule drug dichloroacetate shut down glycolysis in cultured glioblastoma cells but, given systemically, paradoxically shortened mouse survival, while genetic knockout of the same enzyme extended it by roughly ten days — the problem is not the target but getting a specific, targeted intervention through the blood–brain barrier (BBB) to the tumor Refining the Role of Pyruvate Dehydrogenase Kinases in Glioblastoma Development.
Meanwhile, machine learning has learned to design RNA rather than merely analyze it. Reinforcement learning (RL) fine-tuning — taking a generative model that already knows the grammar of biological sequences and then applying reward-driven selection pressure to push it toward desired properties — has produced messenger RNAs with longer half-lives and higher translation efficiency Full-length mRNA Design with Reward-Guided Masked Diffusion Model Fine-Tuning, and lipid nanoparticles decorated with brain-targeting small molecules now deliver functional mRNA cargo to mouse neurons AI-Validated Brain Targeted mRNA Lipid Nanoparticles with Neuronal Tropism. These two threads have never met. This proposal builds the missing link: an RL design loop whose reward function includes not just metabolic potency and RNA stability, but the physical demands of nanoparticle packaging and BBB transit — so that the RNA that comes out is optimized for the delivery vehicle it will actually ride in.
Core Claim
Existing generative-RNA work optimizes transcripts against intracellular predictors only, with no notion of a delivery vehicle or a physiological barrier; existing brain-delivery work ferries computationally untouched reporter mRNAs. I claim that closing this loop matters: an RNA sequence designed jointly for (a) knockdown of a glycolytic driver, (b) serum and intracellular stability, and (c) lipid-nanoparticle packaging behavior will outperform, in trans-BBB assays and orthotopic glioma models, the same knockdown payload taken from the literature and passively loaded into an identical, proven brain-targeted nanoparticle. The corollary is methodological: without an uncertainty-gated reward, the RL agent will exploit blind spots in the predictive oracles and collapse into narrow, high-scoring-but-bogus sequences, a failure mode already visible in prior sequence-RL work.
The concrete target is the PVT1/miR-187-3p axis in glioma, where siRNA knockdown of PVT1 cut glucose uptake and lactate output by 40–50% in culture but was never tested in vivo or across the BBB LncRNA-PVT1 enhances glucose metabolism of pediatric low-grade glioma cells through sponging miR-187-3p, with PDHK1 as a second, harder target in glioblastoma proper.
Research Program
The program has three connected components.
First, the payload generator. I will fine-tune a pre-trained masked diffusion model over full RNA transcripts and short interfering/antisense designs, using the off-policy RL recipe of T3PO-mRNA — Monte Carlo tree search tracing a Pareto frontier, with the pre-trained model acting as a prior that guarantees grammatical validity Full-length mRNA Design with Reward-Guided Masked Diffusion Model Fine-Tuning. The contribution is not the optimizer but the reward. Prior work scored transcripts only for predicted half-life and translation efficiency using XGBoost regressors trained on transcriptome-wide stability and translation datasets; I reuse those oracles and add two new reward terms that no sequence-design loop has yet included: a packaging term scoring GC distribution, predicted secondary-structure folding energy, and length against the enclosure constraints of ionizable lipid nanoparticles, and a metabolic term scoring predicted knockdown of PVT1 (via microRNA-binding and target-prediction models anchored to the Res186/BT66 phenotype data) or PDHK1. The composite reward is
$$R(x) = w_m \hat{M}(x) + w_s \hat{S}(x) + w_p \hat{P}(x) - \lambda\, D_{\mathrm{KL}}\!\left(\pi_\theta \,\|\, \pi_0\right),$$
where $\hat{M}$, $\hat{S}$, and $\hat{P}$ are surrogate scores for metabolic knockdown, stability (the T3PO XGBoost oracles), and LNP packaging, $\pi_\theta$ is the fine-tuned policy, $\pi_0$ the pre-trained prior, and the KL term penalizes drift into degenerate, repetitive sequences. The KL guard is essential: DDPO fine-tuning of a diffusion transformer for DNA regulatory elements produced a 38x improvement in predicted activity but left 92.8% of generations nearly identical to each other — entropy collapse that would be fatal here, where we need a diverse candidate library to survive wet-lab filtering Continuous Diffusion Transformers for Designing Synthetic Regulatory Elements.
Second, uncertainty-gated oracles. Every reward term is a machine-learning predictor, and RL agents are ruthless at gaming them. One BBB-permeability study showed that on an independent clinical test set, sophisticated graph neural networks performed no better than basic physicochemical heuristics — the "DL Hype Check" — unless predictions were filtered by an uncertainty score, under which accuracy exceeded 99% for the most certain 20% of compounds Blood-brain barrier penetration prediction enhanced by uncertainty estimation. I will therefore attach Shannon-entropy/MC-dropout uncertainty estimates to each reward surrogate and zero out reward for candidates whose oracle uncertainty exceeds a calibrated threshold, so the agent is never rewarded inside the region where its judges are guessing. This directly answers a documented open problem: reward exploitation and out-of-distribution hallucination against in silico oracles.
Third, the delivery vehicle is fixed, deliberately. Rather than also generating novel brain-targeting ligands, I use the acetylcholine-functionalized LNP already validated to achieve a ~3.6-fold increase in mouse brain uptake and 59.1% neuronal transgene expression versus 7.7% for untargeted particles, including in a human BBB-on-a-chip AI-Validated Brain Targeted mRNA Lipid Nanoparticles with Neuronal Tropism. Holding the vehicle constant is what makes the payload-design claim testable: any in vivo difference between arms is attributable to the RNA design layer, not the nanoparticle.
Baselines
The naive reference point is the untargeted LNP carrying the same payload — the Sela group's untargeted particles reached only 7.7% of neurons, versus 59.1% for acetylcholine-targeted ones, so any payload-level improvement must be interpreted against this floor AI-Validated Brain Targeted mRNA Lipid Nanoparticles with Neuronal Tropism. A second pharmacological reference is systemic dichloroacetate monotherapy, which paradoxically shortened survival in orthotopic glioblastoma mice even though it suppressed glycolysis in vitro Refining the Role of Pyruvate Dehydrogenase Kinases in Glioblastoma Development; it defines what a non-specific systemic metabolic intervention buys you in this model, which is negative value.
The strongest competitor is the strongest existing combination that requires no new computation: the published PVT1 siRNA sequence from the pediatric low-grade glioma study — the construct that achieved the 40–50% reduction in glucose uptake and lactate secretion in Res186 and BT66 cells — loaded into the acetylcholine-targeted LNP at matched dose LncRNA-PVT1 enhances glucose metabolism of pediatric low-grade glioma cells through sponging miR-187-3p. This baseline is formidable precisely because both of its halves are individually validated in the literature. It will be run with the same nanoparticle formulation, same dosing schedule, same mouse strain and orthotopic implantation protocol, and the in silico arms will receive the same compute and oracle-query budget as my pipeline. Beating it means: higher encapsulation efficiency and serum stability of the payload, greater trans-BBB flux in the spheroid and chip assays, measurably lower intratumoral 2-deoxyglucose-6-phosphate uptake and lactate output in vivo, and — the decisive endpoint — extension of median survival in orthotopic xenografts beyond the literature-siRNA-in-targeted-LNP arm.
Two ablations isolate the proposal's key ingredients. Removing the packaging reward term while keeping everything else yields a payload optimized exactly as T3PO-mRNA optimized its luciferase transcripts — purely for intracellular properties — and tests whether delivery-aware design is what closes the in vivo gap. Removing the uncertainty gate on the reward surrogates tests whether oracle exploitation and entropy collapse, not delivery constraints, are the binding limitation; I expect this arm to look excellent on predicted scores and fail in the spheroid assay, reproducing the in-silico-to-wet-lab gap on purpose. In silico, the internal baselines are inference-time guidance and supervised fine-tuning: T3PO-mRNA's own comparison showed reward-guided fine-tuning beating Monte Carlo tree guidance at inference time by +0.375 in predicted translation efficiency and +0.489 (a 167% increase) in predicted half-life, and the same paper's framing explains why — guidance is trapped inside the prior's support while fine-tuning moves the manifold Full-length mRNA Design with Reward-Guided Masked Diffusion Model Fine-Tuning.
Why not just…?
Why not just wrap the published siRNA in the proven nanoparticle and skip the machine learning entirely? This is the real threat, and I treat it as the primary baseline rather than arguing it away. The honest case for the design layer is that the published siRNA was never asked to survive nanoparticle encapsulation, endosomal escape, and bloodstream transit — it was delivered by ordinary lipid transfection to cells in a plastic dish, a regime with none of those constraints, and its inventors flagged delivery across the BBB as an open problem LncRNA-PVT1 enhances glucose metabolism of pediatric low-grade glioma cells through sponging miR-187-3p. The head-to-head experiment settles it: if the literature siRNA matches my designed payload in encapsulation efficiency, transcytosis, and in vivo metabolic suppression, then the design layer adds nothing for this axis and the claim is falsified — though the pipeline would still be the only route to targets like PDHK1, for which no delivery-validated RNA construct exists at all, only ex vivo CRISPR knockouts that cannot be administered systemically Refining the Role of Pyruvate Dehydrogenase Kinases in Glioblastoma Development.
Why not just use inference-time guidance or supervised fine-tuning instead of RL policy optimization? Both cheaper options were tested head-to-head in the closest prior work and lost. Reward-guided fine-tuning beat inference-time tree guidance by +0.375 predicted translation efficiency and +0.489 predicted half-life on the same model and data Full-length mRNA Design with Reward-Guided Masked Diffusion Model Fine-Tuning, and answer-only supervised fine-tuning of an LLM policy caused documented "reasoning collapse" — the model jumps straight to a final answer and loses the ability to balance competing objectives, which is exactly the multi-objective regime this proposal lives in Reference-guided Policy Optimization for Molecular Optimization via LLM Reasoning. The fairness safeguard is that these baselines run with the same oracles and compute budget, so if they match under my reward, the simpler method wins.
Why not just use a brain-penetrant small molecule like dichloroacetate against the same enzyme? Because it was tried and failed in the exact model system this proposal uses: DCA suppressed glycolysis and invasion in vitro yet, as in vivo monotherapy, failed to shrink orthotopic tumors and shortened mouse survival, while genetic knockout of the same target extended survival Refining the Role of Pyruvate Dehydrogenase Kinases in Glioblastoma Development. That result is the strongest empirical motivation for a targeted, locally delivered genetic intervention, and it is why DCA stays in the experiment as a negative-control reference arm.
Why not just pick ligands by lipophilicity rules instead of learned BBB predictors? Passive diffusion favors small, fat-loving molecules, but the lipophilicity paradox is that the same property degrades solubility, drives plasma-protein binding and peripheral clearance, and hands the molecule to P-glycoprotein efflux pumps that pump it straight back out of the brain; over 98% of small molecules fail these combined filters Overcoming the Blood Brain Barrier for CNS Drug Delivery. Since this proposal fixes the vehicle rather than designing ligands, the learned-predictor question matters mainly for the uncertainty-gating methodology, where the "DL Hype Check" result showed un-gated deep predictors are no better than physicochemical heuristics on independent clinical data Blood-brain barrier penetration prediction enhanced by uncertainty estimation — which is why the gating, not the architecture, is the ingredient I ablate.
Proposed Experiments and Decisive Checks
The in silico phase trains the reward oracles on data already assembled in the literature: the transcriptome-wide half-life and translation datasets behind the T3PO XGBoost regressors Full-length mRNA Design with Reward-Guided Masked Diffusion Model Fine-Tuning, and TCGA/IVYGAP glioma expression for the metabolic targets Refining the Role of Pyruvate Dehydrogenase Kinases in Glioblastoma Development. Oracle discrimination is checked on the MoleculeNet BBBP benchmark for any delivery-related classifier — the reference numbers to beat or match are the MegaMolBART-XGBoost AUC of 0.88 with full correct classification of 21 permeable and 5 impermeable compounds in 3D human BBB spheroids Predicting blood–brain barrier permeability of molecules with a large language model and machine learning, and GROVER-BBBp's greater than 99% accuracy on the low-uncertainty subset of the 1,937-molecule M-data benchmark Blood-brain barrier penetration prediction enhanced by uncertainty estimation. For the no reported result found on Lacuna for any BBB predictor applied to oligonucleotide payloads themselves — the models are small-molecule-only — the payload-side delivery reward must rely on the physical packaging term plus wet-lab measurement, and I state this limitation openly.
Generation quality is judged on predicted half-life and translation gains over the inference-time-guidance baseline (T3PO reported +0.375 and +0.489 on its luciferase libraries; my target is to match that margin while additionally satisfying packaging constraints), on library diversity (self-alignment far below the 92.8% entropy-collapse figure reported for DDPO Continuous Diffusion Transformers for Designing Synthetic Regulatory Elements), and on BLAT-style novelty against training sequences, using the DiT paper's 1.7% memorization rate as the yardstick.
The decisive wet-lab cascade then runs the top library candidates against the literature-siRNA baseline at every stage: encapsulation efficiency and serum nuclease stability; transcytosis across 3D human BBB spheroids quantified by LC-MS/MS, the same assay that validated the 0.88-AUC predictor Predicting blood–brain barrier permeability of molecules with a large language model and machine learning; the human BBB-on-a-chip and iPSC-neuron system used by the brain-targeted LNP work AI-Validated Brain Targeted mRNA Lipid Nanoparticles with Neuronal Tropism; then glioma stem-like cells with 2-deoxyglucose-6-phosphate uptake, lactate secretion, and Seahorse oxygen-consumption and acidification rates as the metabolic readouts, mirroring the assays in both the PVT1 and PDHK studies. The final check is the orthotopic patient-derived xenograft with a survival endpoint, interpreted against three anchors from the file: DCA monotherapy shortening survival, PDHK knockout extending it by roughly ten days, and the biomimetic nanoparticle M@HLPC extending median survival from 26–29 days to 48 days in the same model class Engineered biomimetic nanoparticles achieve targeted delivery and efficient metabolism-based synergistic therapy against glioblastoma — the last of these being an enzyme-payload strategy that consumes lactate downstream rather than reprogramming the upstream genetics, which is precisely the distinction this proposal bets on.
References
- Refining the Role of Pyruvate Dehydrogenase Kinases in Glioblastoma Development
- Full-length mRNA Design with Reward-Guided Masked Diffusion Model Fine-Tuning
- AI-Validated Brain Targeted mRNA Lipid Nanoparticles with Neuronal Tropism
- LncRNA-PVT1 enhances glucose metabolism of pediatric low-grade glioma cells through sponging miR-187-3p
- Continuous Diffusion Transformers for Designing Synthetic Regulatory Elements
- Blood-brain barrier penetration prediction enhanced by uncertainty estimation
- Reference-guided Policy Optimization for Molecular Optimization via LLM Reasoning
- Predicting blood–brain barrier permeability of molecules with a large language model and machine learning
- Engineered biomimetic nanoparticles achieve targeted delivery and efficient metabolism-based synergistic therapy against glioblastoma