PaperlayerView exact revision on arXiv

Source-linked AI summary

Evidence-Grounded Agentic Formulation Development in an Autonomous Laboratory

Michael M. Craig, Riley J. Hickman, Yingshan Ma, Rémi Piché-Taillefer, Christine Allen, Pauric Bannigan

arXiv:2609.19099v1cs.LG

TL;DR

Identifying high-performing SEDDS formulations is experimentally intensive, and conventional sequential optimization often begins largely de novo rather than using prior formulation evidence. Andromeda 2 reasons over structured in-house experimental evidence to design and execute formulation batches. For paclitaxel, it achieved higher high-AUC hit rates than Andromeda 1 and DoE at a matched budget, while assay scope and single-campaign evaluation constrain interpretation.

  • Problem

    SEDDS formulation development requires optimizing multiple components under constraints, while sequential campaigns typically make limited use of accumulated evidence from prior formulation programs.

    Source: §1 · p. 1

  • Method

    Andromeda 2 reasons over structured in-house formulation evidence and orchestrates computational and experimental tools to design and execute successive formulation batches.

    Source: §1 · p. 1, §4.3 · p. 6, §4.4 · p. 7

  • Results

    At a matched budget, Andromeda 2 achieved a 50% high-AUC hit rate, versus 17% for Andromeda 1 and 2% for DoE, and identified 12 full-TPP formulations versus 6 and 0.

    Source: §3 · p. 4

  • Takeaways & Limitations

    The findings suggest that accumulated experimental data may improve subsequent autonomous formulation decisions if the advantage generalizes across APIs and formulation modalities.

    Source: §3 · p. 6

    FaSSIF testing does not capture digestion, intestinal permeability, metabolism, or in vivo performance, and each strategy was evaluated in a single campaign.

    Source: §3 · pp. 5–6

Abstract

from arXiv · show

Self-emulsifying drug delivery systems (SEDDS) can improve the oral bioavailability of poorly soluble drugs, but identifying high-performing formulations remains experimentally intensive. We present Andromeda 2, an agentic system that reasons over structured in-house experimental evidence and invokes computational and experimental tools to design and execute successive formulation batches. Using a miniaturized automated laboratory at a matched budget, we benchmark it against Andromeda 1, a probabilistic optimization model deployed across dozens of live development projects, and a wet-lab design-of-experiments (DoE) campaign. For paclitaxel, Andromeda 2 achieved a 50% high-performance hit rate versus 17% for Andromeda 1 and 2% for DoE, and identified 12 formulations meeting all four target product profile (TPP) objectives versus 6 and 0, respectively. Median $AUC_{10-240}$ was 70.1, 12.0, and 3.5 mg$\cdot$min/mL, while maximum AUC was comparable between Andromeda 2 and Andromeda 1. A selected full-TPP formulation achieved an apparent effective paclitaxel loading of $19 \pm 5\%$ w/w at the first FaSSIF measurement, approximately 3.3-fold higher than the 5.7% w/w loading reported for a published paclitaxel S-SEDDS. A controlled ablation showed that access to structured in-house experimental evidence increased mean AUC by 34%.

1 Introduction

SEDDS development requires balancing multiple formulation components and constraints, making exhaustive experimental exploration impractical. The paper asks whether an autonomous system can use accumulated laboratory evidence to allocate experiments more productively.

1 Introduction

  • SEDDS development balances oil, surfactant, cosolvent, drug loading, and sometimes precipitation inhibitors under strict constraints, while exhaustive exploration is impractical.Conventional workflows focus experiments through excipient screening, phase diagrams, and DoE-guided campaigns.Source: §1 · p. 1
  • Sequential optimization learns from campaign results, but typically initializes largely de novo and makes limited use of evidence from prior formulation programs.The paper asks whether an autonomous system can reason over existing laboratory evidence to allocate scarce wet-lab experiments more productively.Source: §1 · p. 1
  • Oral paclitaxel is a clinically relevant formulation objective, but its poor aqueous solubility and biological barriers complicate oral lipid-based delivery.The study does not address intestinal P-glycoprotein efflux or first-pass metabolism.Source: §1 · p. 1
  • Andromeda 2 reasons over structured in-house experimental evidence and uses computational and experimental tools to design and execute formulation batches.It is evaluated against Andromeda 1 and a physically executed DoE campaign under the same formulation space, TPP, assay, laboratory, and matched budget.Source: §1 · pp. 1–2

2 Results

At a matched 96-formulation budget, Andromeda 2 outperformed Andromeda 1 and experimental DoE on high-AUC hit rate and full-TPP formulations. Structured in-house evidence was associated with better campaign performance, while the selected formulation’s loading comparison provides context rather than a direct ranking against published formulations.

2.1 Andromeda 2 improves paclitaxel formulation performance at a matched budget

  • At 96 formulations each, high-AUC hit rates were 50%, 17%, and 2% for Andromeda 2, Andromeda 1, and DoE.Median AUC10–240 values were 70.1, 12.0, and 3.5 mg·min/mL, respectively; maximum AUC was similar for Andromeda 2 and Andromeda 1 (157.7 vs 156.4 mg·min/mL).Source: §2.1 · p. 2, Figure 1 · p. 3
  • Andromeda 2 identified 12 formulations meeting all four TPP criteria, compared with 6 for Andromeda 1 and 0 for DoE.Its highest-AUC formulation met only two TPP objectives, while its best formulation maintained apparent solubilized paclitaxel through 240 min and reached an AUC approximately 18× the unformulated control.Source: §2.1 · p. 2

2.2 Structured in-house evidence improves Andromeda 2 performance

  • Full-evidence Andromeda 2 had 34% higher mean AUC10–240 than the reduced-evidence configuration: 62.0 versus 46.2 mg·min/mL.The reduced-evidence configuration matched or outperformed early batches but later regressed, while full-evidence performance improved across successive batches and was sustained.Source: §2.2 · p. 2

2.3 Benchmarking Andromeda 2 against published oral paclitaxel formulations

  • The FaSSIF AUC comparison with published oral paclitaxel formulations is not a direct comparison because formulation type, endpoint, and study setting differ.FaSSIF AUC10–240 should not be interpreted as oral bioavailability or systemic exposure.Source: §2.3 · p. 3
  • The selected full-TPP formulation had an apparent effective paclitaxel loading of 19±5% w/w at the first FaSSIF measurement.This was approximately 3.3-fold higher than the 5.7% w/w reported for Gao et al.’s S-SEDDS, but differing loading and measurement methods preclude direct quantitative ranking.Source: §2.3 · p. 3

3 Discussion

At a matched budget, Andromeda 2 produced more high-performing paclitaxel formulations than Andromeda 1 or DoE, though their maximum AUC values were similar. Structured in-house evidence contributed to Andromeda 2’s sustained performance, while FaSSIF results do not establish oral exposure.

3 Discussion

  • 50% versus 17% and 2% high-AUC hit rates, and 12 versus 6 and 0 full-TPP formulations, favored Andromeda 2 over Andromeda 1 and DoE at 96 formulations.Maximum AUC was similar for Andromeda 2 and Andromeda 1; the highest-AUC Andromeda 2 formulation maintained apparent solubilization through 240 min, while Andromeda 1 declined after 120 min.Source: §3 · p. 4, Figure 2 · p. 4
  • Andromeda 2’s advantage was not explained solely by identifying a favourable composition-space region: within the high-performing Type IIIB/IV region, its hit rate was 49%, compared with 29% for Andromeda 1.Source: §3 · p. 4
  • 34% higher mean AUC with full structured in-house evidence accompanied more consistent improvement across batches than the reduced-evidence configuration.The reduced-evidence configuration could match or exceed full evidence in early batches but did not sustain that performance; the ablation does not attribute the entire Andromeda 2–Andromeda 1 difference to historical evidence.Source: §3 · pp. 4–5, Figure 3 · p. 5
  • The FaSSIF assay does not capture digestion, intestinal permeability, metabolism, or in vivo performance, so observed formulation advantages do not establish oral exposure.Each strategy was evaluated in one independently initialized campaign; prospective evaluation across distinct APIs is also needed to establish generality.Source: §3 · pp. 5–6
  • The authors propose reusable accumulated experimental data as a model for improving subsequent autonomous formulation decisions, conditional on the advantage generalizing across APIs and modalities.Source: §3 · p. 6

4 Methods

The methods combine a miniaturized SEDDS laboratory with sequential formulation design, evidence-grounded batch generation, and matched-budget comparisons. Performance is assessed using FaSSIF AUC and a four-objective target product profile, with formulation-level statistical results treated descriptively.

4.1 Autonomous SEDDS laboratory

  • Each 16-formulation batch used robotic preparation and automated dissolution, particle characterization, HPLC quantification, and structured data capture; FaSSIF measurements were not treated as oral bioavailability.The apparent effective loading estimate was calculated retrospectively from nominal loading and the maximum fraction of the nominal dose observed in the dispersed phase.Source: §4.1 · p. 6

4.2 Andromeda 1: probabilistic optimization

  • Andromeda 1 used probabilistic models to select each successive 16-formulation batch from a constrained composition space, balancing promising regions against uncertain ones.New dissolution, droplet-size, PDI, and HPLC measurements were incorporated before the next cycle.Source: §4.2 · p. 6

4.3 Andromeda 2: agentic formulation design

  • Andromeda 2 orchestrated models and computational tools to propose and critique candidates, while deterministic checks enforced executable compositions before robotic execution.It generated 16-formulation batches with scientific rationales, adapted its working context using prior batch measurements, and a scientist reviewed batches for safety and feasibility without redesigning them.Source: §4.3 · pp. 6–7

4.4 Evidence grounding and ablation

  • Andromeda 2 began with structured in-house formulation evidence but no historical paclitaxel formulations; both evidence configurations received new campaign measurements after each batch.The reduced-evidence comparison withheld historical evidence while holding the objective, TPP, design space, constraints, batch size, and workflow constant.Source: §4.4 · pp. 7–8

4.5 Comparators and paclitaxel design task

  • The paclitaxel comparison gave each arm six 16-formulation batches, while the physically executed DoE used the same composition parameters, constraints, platform, and FaSSIF assay.Simulated adaptive and non-adaptive DoE strategies were evaluated separately against an in-silico emulator.Source: §4.5 · p. 8

4.6 Metrics and statistical analysis

  • The primary endpoint was trapezoidal AUC10–240 in FaSSIF; full TPP passage required all four objectives, and formulation-level tests were descriptive because adaptive observations were not independent.The TPP included minimum dissolution AUC, late-time concentration and fraction-dissolved floors, and a precipitation-resistance ratio of C240/C30 ≥0.85.Source: §4.6 · pp. 8–9, Figure 5 · p. 10

A.1 Descriptive Mann–Whitney comparison of Andromeda 2 and Andromeda 1

The formulation-level Mann–Whitney result is descriptive because adaptive campaign observations are dependent.

A.1 Descriptive Mann–Whitney comparison of Andromeda 2 and Andromeda 1

  • p = 9.6 × 10^-8 was treated as descriptive, since later candidate selection depended on earlier measurements and 96 formulations overstated the effective sample size.The matched-budget comparison therefore emphasized distributional shifts and observed hit rates rather than campaign-level inference.Source: §A.1 · p. 14

A.2 Formulation composition and the LFCS taxonomy

Within the evaluated paclitaxel design space, high-performing formulations concentrated in LFCS Types IIIB and IV. Andromeda 2 focused its search there and achieved higher hit rates within that productive region while retaining chemical diversity.

A.2 Formulation composition and the LFCS taxonomy

  • The LFCS classification used oil, water-insoluble and water-soluble surfactants, and hydrophilic cosolvent proportions to define Types II, IIIA, IIIB, and IV; Type I was absent.Type II has oil and water-insoluble surfactant; IIIA has substantial oil plus water-soluble surfactant; IIIB has less oil and more water-soluble components; IV is oil-free.Source: §A.2 · p. 14
  • 35% of Type IIIB and 32% of Type IV formulations were high-AUC hits, versus 6% of Type IIIA and 0% of Type II.These productive types contain water-soluble surfactants and have low or no oil content.Source: §A.2 · p. 14, Figure 6 · p. 15
  • Within Type IIIB/IV, Andromeda 2 achieved a 49% hit rate, versus 29% for Andromeda 1 and 3% for experimental DoE.Andromeda 2 both concentrated its search in the productive region and selected higher-performing candidates within it.Source: §A.2 · p. 14, Figure 6 · p. 15
  • Andromeda 2 shifted toward Type IIIB/IV as measurements accumulated and AUC increased, while DoE showed little corresponding batch-to-batch improvement.Andromeda 2 sampled 16 oil-surfactant families versus DoE’s 90, yet retained several productive families as chemically distinct alternatives.Source: §A.2 · p. 14, Figure 6 · p. 15

A.3 Literature comparison of oral paclitaxel formulations

The literature comparison provides formulation context, not a direct ranking, because studies differ in endpoints and reporting conventions.

A.3 Literature comparison of oral paclitaxel formulations

  • Table 2 presents selected published oral paclitaxel formulations alongside Andromeda 2 for context, not as a head-to-head comparison.The FaSSIF AUC used in the study should not be equated with oral bioavailability or systemic exposure.Source: Table 2 · p. 16, §A.3 · p. 16
  • The loading footnotes distinguish nominal design input from measured or effective loading and explain conversions used for published formulations.Gao et al.’s 57 mg/g converts to 5.7% w/w; DHP107’s approximate % w/w assumes a formulation density near 1 g/mL.Source: §A.3 · p. 16

A.4 In silico emulator validation

The paclitaxel emulator recovered measured performance ordering on held-out formulations and enriched top-quartile selections over random choice. Its simulated DoE comparisons are directional, with weaker ranking accuracy for oil–surfactant families absent from the validation data.

A.4 In silico emulator validation

  • Spearman 0.89 ranking and 3.02-fold top-quartile enrichment over random selection validated the emulator on held-out formulations (ceiling: 4.0).Table 3 reports held-out ranking skill, error, and top-quartile enrichment; enrichment uses 1.0 as chance baseline and 4.0 as the perfect-selection ceiling.Source: §A.4 · p. 16, Table 3 · p. 16
  • Neither simulated non-adaptive nor sequential adaptive DoE reached the upper-tail AUC values measured for the Andromeda strategies.The experimental DoE campaign likewise identified no formulations in that range.Source: §A.4 · p. 16
  • The emulator's ranking accuracy was lower for oil–surfactant composition families absent from its random held-out validation data, so simulated results are directional rather than matched comparisons.The simulated arms should be weighted accordingly.Source: §A.4 · p. 16
Loading 2609.19099v1…