Illuminating the Depot: Quality-Diversity Search for Long-Acting Polymeric Nanoparticles

Encapsulating a drug inside a biodegradable polymer particle — a slowly eroding matrix of materials like PLGA — turns a toxic concentration spike into a depot that releases the drug steadily for weeks or months. This solves the core bioavailability problem for poorly soluble, rapidly cleared drugs, but only if you can find a working recipe, and that search is still empirical trial-and-error across polymer chemistry, molecular weight, surfactant, and processing conditions. Machine learning has entered this space in two limited ways: as a passive predictor of release curves, and as a greedy optimizer that hunts for the single best formulation. The second approach has a documented failure mode — Bayesian optimization collapses onto one "safe" chemical family and never explores the rest of the space.

Meanwhile, the generative modeling community has spent years studying exactly this tension between per-sample quality and population diversity, and has produced algorithms that deliberately maintain a collection of high-performing, mutually distinct solutions. This proposal brings those quality-diversity ideas to solid polymeric nanoparticle design. The deliverable is not one optimized formulation but an illuminated map: several chemically distinct formulation families — different polymer backbones, different diffusion-barrier shells — that all pass hard bioavailability gates, so that if the lead candidate fails in toxicology, manufacturing, or patent review, high-performing backups already exist.

Core Claim

The claim has three parts. First, formulation search for polymeric nanoparticles should be an illumination problem, not a point-optimization problem: we want many distinct, gate-passing families, not one winner. Craig et al. demonstrated multi-family exploration, but only for liquid surfactant-oil mixtures (SEDDS) targeting immediate oral dissolution, with no sustained release and a dependence on a large proprietary measurement database (Agentic Systems for Sample-Efficient Drug Formulation Design); we target solid biodegradable matrices with multi-week release kinetics, a setting they explicitly leave open. Second, the small-data regime of formulation science (roughly 100–200 measured curves) breaks existing quality-diversity machinery: soft quality-diversity search needs differentiable models (Soft Quality-Diversity Optimization), yet tree models beat neural networks on formulation data (Machine learning models to accelerate the design of polymeric long-acting injectables; Data-driven development of an oral lipid-based nanoparticle formulation of a hydrophobic drug). We resolve this with a hybrid: tree ensembles as teachers, distilled into a small differentiable student that carries an uncertainty estimate. Third, the search itself should be two-phase, mirroring both the two-stage generative architectures that resolve the quality-diversity conflict in vision (DUET: A Diversity-Quality Duet of Distillation Experts for Two-Step Video Generation; Draft-and-Revise: Effective Image Generation with Contextual RQ-Transformer) and the two-stage physical architecture of the particles themselves — a drug-loaded porous core plus an outer coating that gates burst release (Bioinspired surface functionalization of biodegradable mesoporous silica nanoparticles for enhanced lubrication and drug release; Biodegradable Polymer Encapsulation for Sustained Drug Release). A diversity-focused "draft" phase proposes distinct backbone-and-shell families; a quality-focused "revise" phase tunes continuous parameters within each family.

Research Program

The search operates over a formulation vector $x$ covering polymer backbone (PLGA, PLA, PCL, chitosan, and blends), lactide:glycolide ratio, polymer molecular weight, surfactant type and concentration, and a categorical choice of shell or gatekeeper chemistry (none, chitosan electrostatic layer, bioinspired polymer coating). Each candidate is scored by a forward surrogate predicting drug loading capacity, encapsulation efficiency, particle size and polydispersity, the 24-hour burst fraction, and the cumulative release curve from which we read the release duration. Hard bioavailability gates $G(x)$ zero out candidates that fail thresholds — loading at or above 15% weight-for-weight (beating the 1–5% typical of polymer-drug conjugates and approaching the 19.4% achieved manually for SN-38 in albumin-PLA particles (Albumin-based nanoparticles encapsulating SN-38 demonstrate superior antitumor efficacy compared to irinotecan)), encapsulation efficiency above 90%, diameter under roughly 200 nm with polydispersity index under 0.2, and a 24-hour burst fraction below a preset ceiling. Quality $q(x) \in [0,1]$ is a composite of how far each predicted property exceeds its gate. The behavior descriptor $b(x)$ is a low-dimensional vector — predicted loading, burst fraction, time-to-50%-release, and a chemistry fingerprint of backbone and shell — and the search maximizes a soft quality-diversity objective of the SQUAD form:

$$\mathcal{J} \;=\; \sum_i q(x_i)\, G(x_i) \;-\; \lambda \sum_{i \neq j} q(x_i)\, q(x_j)\, \exp\!\big(-d(b(x_i), b(x_j))^2 / 2\sigma^2\big),$$

where $d$ is a distance in descriptor space and $\lambda, \sigma$ control how strongly equally good but chemically similar formulations repel one another. This penalizes redundant near-duplicate winners, which is precisely the redundancy that greedy optimizers produce.

The surrogate is the methodological crux. We train LightGBM ensembles on the curated long-acting injectable dataset of 181 release profiles and over 3,700 fractional release measurements (Machine learning models to accelerate the design of polymeric long-acting injectables) plus our own new nanoprecipitation data, then distill the ensemble into a small differentiable network whose gradient disagreement across teachers provides an epistemic uncertainty estimate; candidates with high uncertainty are routed to the wet lab rather than trusted. The two-phase loop runs batched: the draft phase proposes whole-family switches (categorical moves scored by family-level expected quality), the revise phase runs gradient ascent on the distilled surrogate within each promising family, and a fixed batch (e.g., 16 candidates) is synthesized per round in 96-well automated nanoprecipitation, following the CBD lipid-nanoparticle campaign that characterized a 1,215-formulation space by synthesizing only 128 of them (Data-driven development of an oral lipid-based nanoparticle formulation of a hydrophobic drug). Automation matters for reproducibility as well as throughput: robotic control of addition rates near 0.3 mL/s proved necessary to reliably reproduce nanoparticle synthesis in prior autonomous work (Autonomous nanoparticle synthesis by design).

Baselines

The naive reference point is passive uniform sampling: a random or grid-selected set of 128 formulations from the design space, synthesized and measured with no search algorithm, exactly as the CBD campaign sampled 10% of its lipid formulation space (Data-driven development of an oral lipid-based nanoparticle formulation of a hydrophobic drug). This tells us how much of the space passes the gates by luck alone. The strongest existing competitor is batch Bayesian optimization with a gradient-boosted-tree surrogate and an expected-improvement acquisition — the method that Craig et al. used as their baseline and showed reaches a given performance level in four 16-experiment batches where a diversity-preserving system needed one (Agentic Systems for Sample-Efficient Drug Formulation Design). We run it on the same surrogate, the same wet-lab batch size, the same total experiment budget, and the same tuning effort, and we track not only its best loading but the number of distinct chemical families among its gate-passing candidates. The third baseline is the strongest predictive method used as-is: the LightGBM forward model in the style of Bannigan et al., whose optimized model reported a mean absolute error of 0.114 on fractional release (Machine learning models to accelerate the design of polymeric long-acting injectables), used to rank the design space once and synthesize its top predictions with no iterative search — this isolates the value of the search loop versus the surrogate itself. The key ablation removes the repulsion term ($\lambda = 0$), reducing our method to pure quality maximization on the same surrogate; this isolates the quality-diversity ingredient, since any gain in family coverage must then come from the repulsion and the two-phase structure rather than from the surrogate or the batch design. Beating the baselines means: within a fixed 128-experiment wet-lab budget, at least three chemically distinct families (different polymer backbones or shell chemistries) pass all gates, with the best loading within 10% of the Bayesian optimization winner — whereas we expect Bayesian optimization to certify one family, as it clustered around a single chemical family in the SEDDS setting.

Why not just…?

Why not just run Bayesian optimization? It is a real threat on sample efficiency, and if the only deliverable were one high-loading formulation, Bayesian optimization would likely win per-experiment. But the evidence is direct that it mode-collapses: Craig et al. documented their gradient-boosted Bayesian optimization baseline clustering around a single "safe" chemical family and needing 64 experiments to match what a diversity-preserving system achieved in 16 (Agentic Systems for Sample-Efficient Drug Formulation Design). The head-to-head in our experiments settles the trade-off empirically: we report both best-quality and family-coverage metrics, and if Bayesian optimization happens to find multiple distinct families in this space, our claim weakens and we will say so.

Why not just use design-of-experiments or a grid screen? The CBD campaign itself shows the ceiling of this approach: even a modest multi-component space held 1,215 candidate formulations, and full screening was resource-prohibitive, forcing a 10% passive slice (Data-driven development of an oral lipid-based nanoparticle formulation of a hydrophobic drug). Grid sampling also treats all regions equally, spending expensive active pharmaceutical ingredient on regions the surrogate already knows fail. Our random-sampling baseline quantifies exactly this gap in our space.

Why not just train a deep generative model end-to-end on formulation data? Both studies that tested this empirically found deep networks lose to tree models at this data scale: Bannigan et al. found tree-based models systematically outperforming neural architectures on 181 release profiles, and the CBD study found a decision tree at $R^2 = 0.98$ for drug loading beat its neural networks (Machine learning models to accelerate the design of polymeric long-acting injectables; Data-driven development of an oral lipid-based nanoparticle formulation of a hydrophobic drug). This is why our differentiable model is a distilled student of tree teachers rather than a from-scratch network — but the honest version of this concern is that distillation might erase the tree models' accuracy advantage, which is exactly what the teacher-versus-student ablation measures.

Why not just solve bioavailability by chemistry — prodrugs, conjugates, or the pH trick? Xiong et al. achieved 19.4% loading and 97.1% encapsulation efficiency for SN-38 by exploiting its lactone-carboxylate equilibrium (Albumin-based nanoparticles encapsulating SN-38 demonstrate superior antitumor efficacy compared to irinotecan), and we adopt their system as our reference point rather than a rival. But their route is compound-specific (it requires a hydrolyzable lactone ring), polymer-drug conjugates are stuck at 1–5% loading, and neither route optimizes sustained-release kinetics or provides alternative families. Our proposal generalizes across poorly soluble drugs and outputs backup chemistries by construction.

Proposed Experiments

The first phase is offline, on the reconstructed long-acting injectable release dataset (181 profiles, 43 drug-polymer combinations, held out by drug). We validate the distilled surrogate against its tree teachers — the target is to stay within a small margin of the reported LightGBM mean absolute error of 0.114 on fractional release (Machine learning models to accelerate the design of polymeric long-acting injectables) — and then replay the search: quality-diversity versus Bayesian optimization versus random sampling, all querying the same surrogate, scored on quality-diversity score (total gated quality of the archive), coverage (number of distinct polymer families passing gates), and sample efficiency. SQUAD's own scaling evidence — near-double the quality-diversity score of the next best method on a 7-dimensional behavior-space benchmark (Soft Quality-Diversity Optimization) — motivates the descriptor design but does not predict performance here, and we treat it that way.

The second phase is the wet-lab campaign: SN-38 as the payload, chosen because it is nearly insoluble in both water and organic solvents and has a fully characterized manual benchmark (19.4% loading, 97.1% encapsulation efficiency, 163 ± 21 nm diameter, polydispersity index 0.18) (Albumin-based nanoparticles encapsulating SN-38 demonstrate superior antitumor efficacy compared to irinotecan). Across roughly 128 synthesized formulations in 96-well nanoprecipitation, we measure loading, encapsulation efficiency, size, polydispersity, and 80-hour release profiles; the release benchmark comes from the coated-versus-uncoated mesoporous carrier study, where a bioinspired polymer coating held release to 39% of payload over 80 hours against 70% for the uncoated control (Bioinspired surface functionalization of biodegradable mesoporous silica nanoparticles for enhanced lubrication and drug release), and the chitosan-gate study that stretched release from hours to over a week (Biodegradable Polymer Encapsulation for Sustained Drug Release). Success is at least three chemically distinct gate-passing families within budget, with the best family's loading within 10% of the Bayesian optimization winner.

The decisive checks close the loop. We scale up the top two or three archived families and run extended release profiling toward a multi-week depot profile, taking as the existence proof the month-long sustained release of nalmefene from PLGA microspheres in animals (Biodegradable Polymer Encapsulation for Sustained Drug Release). In vivo pharmacokinetics in rats, benchmarked against free drug exactly as the CBD campaign reported a 15-fold exposure improvement (Data-driven development of an oral lipid-based nanoparticle formulation of a hydrophobic drug), would establish that the archived families translate; the file contains no in vivo surrogate model for polymeric nanoparticle pharmacokinetics, so we treat in vivo behavior strictly as a final validation, not something the search optimizes. Finally, the repulsion-off ablation, run at matched budget, is the clean test of the central claim: if removing the diversity pressure does not shrink the number of surviving families, then the quality-diversity machinery is not earning its place in this domain, and the honest conclusion would be that a good surrogate plus Bayesian optimization suffices.