Co-Designing Lipid and Recipe: Corona-Aware Search over Structure and Formulation Space

Less than one percent of an injected nanoparticle dose typically reaches a solid tumor — a deficit so consistent it is called the "0.7% problem" (Biopolymer-Based Nanomedicine for Cancer Therapy: Opportunities and Challenges). The dose is lost in a fixed sequence of stages: serum proteins coat the particle within minutes and flag it for clearance by liver and spleen macrophages; high interstitial fluid pressure pushes back against entry into the tumor (Enabling Anticancer Therapeutics by Nanoparticle Carriers: The Delivery of Paclitaxel); and particles that do get in stall in the dense extracellular matrix near blood vessels. Meanwhile, the machine-learning community has built impressive tools for exploring chemical space — the astronomically large universe of possible molecules — but it points those tools at the wrong objective. Generative models design single molecules against simple filters; formulation optimizers shuffle the ratios of a handful of already-known ingredients. Neither searches the space that actually determines delivery: the joint choice of what the molecules are and how they are assembled together, under pressure from the real biological barriers.

This proposal builds a search algorithm that operates on that joint space, with the protein corona — the layer of serum proteins that adsorbs onto a particle and largely decides its fate — as an explicit, learned objective rather than an afterthought. The core move is a two-level design loop: an inner search that grows new ionizable lipid structures through validated chemical reactions, and an outer search that simultaneously optimizes the formulation recipe (helper lipid, sterol, PEG lipid fractions, and how many of each), all scored by surrogate models that predict transfection, endosomal escape, and corona composition. Crucially, the loop can co-design two novel lipids to blend in one particle, exploiting a synergy that formulation work has observed but never been able to search for chemically.

Core Claim

Existing work splits cleanly into two camps that do not talk to each other. On one side, Generative Model for Synthesizing Ionizable Lipids: A Monte Carlo Tree Search Approach uses Monte Carlo Tree Search (MCTS — a game-tree search that treats molecule construction as a sequence of moves pairing building blocks with validated reaction templates) to generate more than 5,000 synthesizable ionizable lipids, but scores them only on being lipid-like and having switchable charge; it never predicts delivery and never considers what else is in the particle. On the other side, Designing lipid nanoparticles using a transformer-based neural network builds COMET, a transformer that treats a whole formulation like a sentence and ranks it (test Spearman correlation 0.873 on its LANCE dataset of 3,028 formulations), and even discovers that blending two different ionizable lipids can beat either alone — but it can only shuffle a palette of 7 known lipids; it cannot invent a new one. Between them sits LANTERN, which shows that on ~1,100 clean data points a simple model on fingerprints and expert descriptors predicts transfection far better than deep graph networks ($R^2 = 0.8161$ versus 0.2655 for the AGILE deep model on the same test set), and the protein corona predictor, which forecasts how strongly 178 serum proteins adsorb onto a nanoparticle (top classifier AUROC 0.969) but is purely descriptive — no one has ever put it inside a design loop.

The claim of this proposal is that a search over the product space of molecular structure and formulation recipe, rewarded by a composite objective that includes predicted corona composition, will find lipid nanoparticle designs that no existing pipeline can reach — and that this can be demonstrated with a bounded, purified wet-lab campaign. The composite reward for a candidate structure $x$ (or pair of structures) and recipe $f$ is

$$R(x, f) = \alpha\,\hat{T}(x, f) + \beta\,\hat{D}(x, f) - \gamma\,\hat{O}(x, f),$$

where $\hat{T}$ is predicted transfection efficiency from a descriptor-based surrogate, $\hat{D}$ is predicted adsorption of dysopsonins (protective proteins like albumin and clusterin), $\hat{O}$ is predicted adsorption of opsonins (clearance-driving immunoglobulins and complement proteins), and $\alpha, \beta, \gamma$ are weights set on validation data. The ratio $\hat{D}/\hat{O}$ is the corona-quality term, and it is what makes the search aim at surviving the bloodstream rather than merely at a Petri-dish assay.

Research Program

The program has three components, ordered by risk.

First, re-featurize the corona model. The existing predictor was trained on 652 data points whose 21 features are mostly macroscopic — core material, incubation serum, temperature — and its dominant drivers were unmodified inorganic cores (silica, gold, silicon), not organic lipid surfaces. No reported result found on Lacuna for corona prediction on lipid nanoparticle surfaces specifically, so transfer to LNPs is an open empirical question, not a given. We will augment the 21 features with molecular descriptors of the surface-forming lipids (the LANTERN lesson: in low-data regimes, fingerprints plus expert physicochemical descriptors beat deep learned representations), retrain the Extremely Randomized Trees / Random Forest stack on the Ban et al. dataset, and then extend it with in-house corona measurements: liquid-chromatography mass-spectrometry profiling of the corona on roughly 80–100 LNP formulations spanning structural families, collected during the loop itself. We deliberately predict only a ranked opsonin-versus-dysopsonin panel rather than the full competitive corona, because the Vroman effect (proteins displacing each other over time) is unmodeled even in the source work and we do not want the reward to rest on dynamics nobody can yet predict.

Second, build the two-level search. The inner level reuses the SyntheMol-style MCTS machinery of Zhao et al. — building blocks from ZINC20 and LipidMaps, reaction templates, a policy network — but its terminal reward is the composite $R$, not a lipid-likeness filter. The outer level optimizes the recipe: molar fractions of helper lipid, cholesterol, PEG lipid, and the split between the one or two generated ionizable lipids, using sparse-axis Bayesian optimization of the kind validated for high-dimensional, low-sample autonomous search in ScatterLab, which converged on an 11-dimensional inorganic synthesis space in 41 autonomous experiments. Allowing the outer loop to request a pair of novel lipids is the feature that directly exploits COMET's dual-ionizable synergy finding with structures no fixed palette contains.

Third, close the loop in the lab with purified chemistry. AGILE synthesized 1,200 lipids robotically but used crude, unpurified one-pot products, injecting reaction noise into every training label; we accept a tenfold smaller throughput (three batches of 32 candidates) in exchange for chromatographic purification, so that our labels measure the compound, not the mixture. Each candidate is assayed on three axes that mirror the biological gauntlet: transfection in HeLa cells and RAW 264.7 macrophages (AGILE's protocol, giving direct comparability); corona composition by LC-MS after serum incubation; and penetration into multicellular tumor spheroids, an in vitro stand-in for the extracellular-matrix barrier, since no reported result found on Lacuna for computational prediction of ECM penetration. Results feed back into the surrogates and the search continues.

Baselines

The do-nothing reference is random draws from the same building-block space, which Zhao et al. report meet all ionizable-lipid criteria at a rate of only about 0.15, alongside the unmodified clinical lipid MC3 as the formulation reference point. The strongest existing competitor is the obvious sequential pipeline: run the MCTS generator of Zhao et al. unmodified, rank its output with the LANTERN surrogate ($R^2 = 0.8161$ on its curated 1,100-lipid test set, top-lipid identification 73.7% of the time), then hand the top single lipids to a COMET-style ratio optimizer (test Spearman 0.873 on LANCE). This pipeline uses the same components, the same surrogates, and the same reaction templates as our proposal — it differs only in searching the spaces separately and in ignoring the corona — which makes it the cleanest possible test of whether joint search and the corona term earn their complexity. It will be given the identical wet-lab budget (96 purified syntheses total), the identical surrogate-call budget, and the same hyperparameter tuning effort, with its own tuning done by an operator blind to our method's results. The third baseline is Gaussian-process Bayesian optimization over a fixed library of known lipids and their ratios, using fingerprint descriptors — the classical alternative with calibrated uncertainty. The key ablation removes the corona term ($\beta = \gamma = 0$) from our own search, isolating whether corona-awareness specifically, rather than joint search generally, drives any gain. We beat the sequential pipeline if, at equal synthesis budget, our candidates achieve a higher hit rate — a hit being a candidate whose measured composite (transfection times spheroid penetration, divided by measured opsonin share) lands in the top decile of all synthesized candidates across all arms — and if the ablation without the corona term underperforms the full method on the corona axis without fully explaining away the overall gap.

Why not just…?

Why not just generate with MCTS and filter with LANTERN? That is exactly the sequential-pipeline baseline, and we treat it as the real threat. It will likely match us on the transfection axis alone, because LANTERN's surrogate is strong on exactly that quantity. The honest expectation is that our advantage must come from the two things it structurally cannot do: it cannot see the recipe (so it cannot find dual-lipid synergies involving novel structures), and it cannot see the corona (so it may pick chemically excellent lipids that get cleared before reaching anything). Our corona LC-MS measurements will directly test the second point: if the sequential pipeline's top picks show high measured opsonin share, that is evidence the filter is optimizing the wrong target; if they do not, the corona term was unnecessary and we will say so.

Why not just extend COMET's palette? COMET already ranks formulations well (Spearman 0.873) and its in vivo results are strong — designed LNPs achieved more than 5-fold higher protein expression than SM-102 and more than 40-fold higher than DLin-MC3-DMA. But it is a ranker over known components, and the space of possible ionizable lipids is combinatorially vast; adding lipids one at a time to its training set explores that space at a rate of one experiment per dimension, which is precisely the brute-force bottleneck that AGILE already hit when its 1,200 syntheses covered a tenth of a 12,000-candidate virtual library. The decisive comparison is structural novelty: if the top hits from joint search are structurally distant (low Tanimoto similarity) from every lipid in LANCE and still win on measured composite score, the fixed-palette approach could not have found them at any budget.

Why not just Gaussian-process Bayesian optimization with descriptors? On a fixed library this is a serious competitor, and we run it with full tuning rather than strawmanning it. The countervailing evidence is behavioral: in Agentic Systems for Sample-Efficient Drug Formulation Design, a gradient-boosted BO baseline needed 64 experiments to reach what an agentic system achieved in 16, and was observed to cluster its experiments around a single safe chemical family. Over a joint discrete-structure-plus-continuous-recipe space, kernel-based BO additionally has no natural distance metric, since two molecules can be near-identical in recipe yet one bond apart in structure with wholly different behavior — the "activity cliff" problem noted across chemical-space work. Our experiment settles it empirically: same budget, same assays, hit rate compared directly.

Why not just wrap the particles in cell membranes instead of doing chemistry? Biomimetic cloaking works — an RBC-membrane-coated, iRGD-functionalized system achieved a fourfold increase in intratumoral accumulation and complete tumor regression in 60% of mice with anti-PD-1 (Intravenous iRGD-Guided, RBC-Membrane Camouflaged Lactococcus Lactis Remodels Cold NSCLC and Enhances PD-1 Blockade) — but that same paper shows its limits: the membrane covered only about 60% of the surface, leaving the rest exposed to opsonization, and cloaking alone did not address matrix penetration or endosomal escape, which required the engineered peptide. Membrane coatings also carry the batch-to-batch variability that plagues all natural-source materials (Biopolymer-Based Nanomedicine). The two approaches are composable rather than competing: a corona-optimized synthetic surface under a membrane cloak is a natural follow-up, not a refutation.

Decisive Checks

Three outcomes would decide the proposal's fate. If the re-featurized corona model cannot rank opsonin-versus-dysopsonin balance on held-out LNP measurements better than chance (target: held-out Spearman above 0.5 against LC-MS), the corona term is not yet usable and the honest fallback is to make corona a measured-feedback signal in the loop rather than a predicted one — a slower but still viable design criterion. If the joint search, at 96 purified syntheses, does not beat the sequential pipeline on measured composite hit rate, the central claim fails, and the ablation will tell us whether joint structure-recipe search, the corona term, or both were the dead weight. And if the top joint-search hits turn out to be structurally close to lipids already in LANTERN's curated set, then the added search machinery bought nothing the existing library did not contain. Conversely, a small set of novel, structurally distant, corona-favorable lipids whose formulations out-perform MC3 on the composite metric in the spheroid-and-serum assays would be the first demonstration that chemical-space exploration can be pointed at the delivery gauntlet itself rather than at its surrogate assays.