Evidence-Grounded Agentic Formulation Development in an Autonomous Laboratory
Michael M. Craig, Riley J. Hickman, Yingshan Ma, R\'emi Pich\'e-Taillefer, Christine Allen, Pauric Bannigan
cs.LG
Quick verdict
Evidence-grounded agentic design beat optimization and DoE on paclitaxel SEDDS at matched budget; strong internal benchmark but single-run, in vitro only.
Summary
Intrepid Labs presents Andromeda 2, an agentic system that designs SEDDS (self-emulsifying drug delivery) formulations by reasoning over structured in-house experimental evidence and invoking computational and robotic tools. In a matched 96-formulation campaign on paclitaxel in a miniaturized automated laboratory, it achieved a 50% high-AUC hit rate versus 17% for Andromeda 1 (the company's deployed probabilistic optimizer) and 2% for a physically executed DoE campaign, identified 12 formulations meeting all four TPP criteria versus 6 and 0, and raised median FaSSIF AUC10-240 to 70.1 versus 12.0 and 3.5 mg·min/mL; maximum AUC was similar between the two Andromeda systems (157.7 vs 156.4). A controlled ablation withholding historical evidence cut mean AUC by 34% (46.2 vs 62.0 mg·min/mL) and hit rate to 31%, with the reduced-evidence variant matching early batches but regressing later. LFCS composition analysis shows the advantage persists within the productive Type IIIB/IV region (49% vs 29% hit rate), so it is not merely region selection. Key limitations: each arm ran once (no campaign replicates), the endpoint is in vitro FaSSIF solubilization rather than bioavailability, the DoE comparator lacked expert pre-screening, all statistics are explicitly descriptive, and no code or data release is stated. A benchmark with unusually honest statistical framing, but generality and translational value remain unproven.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- At a matched 96-formulation budget on paclitaxel SEDDS, Andromeda 2 achieved a 50% high-AUC hit rate and 12 full-TPP formulations versus 17%/6 for Andromeda 1 and 2%/0 for physical DoE (Table 1).
- Withholding structured in-house historical evidence reduced mean AUC by 34% (46.2 vs 62.0 mg·min/mL) and hit rate from 50% to 31%, with the ablated arm regressing in later batches despite early parity.
- The advantage is not explained merely by finding the productive Type IIIB/IV composition region, since within that region hit rates were 49% vs 29% vs 3% (Figure 6C).
- A selected full-TPP formulation showed an apparent effective paclitaxel loading of about 19±5% w/w, roughly 3.3x the 5.7% w/w of a published S-SEDDS, though the paper concedes measurement methods differ.
Strengths
Matched-budget comparison against a production-proven baseline, not a strawman
Section 2.1/Table 1: all arms ran on the same miniaturized robot, assay, TPP, composition grid, and 96-formulation budget; Andromeda 1 is described as deployed across dozens of live development programs, making the 50% vs 17% hit-rate gap a comparison against a genuinely competitive optimizer rather than a weak surrogate.
A real controlled ablation isolates the evidence contribution
Section 2.2/Figure 3: the reduced-evidence arm held objective, TPP, design space, constraints, batch size, and wet-lab workflow constant and still received in-campaign measurements; the batch-resolved trajectories (early parity, later regression for the ablated arm: mean 46.2 vs 62.0 mg·min/mL, hit rate 31% vs 50%) specifically rule out a 'stronger first batch' explanation.
Advantage demonstrably not just region selection
Unusually candid statistical framing
Careful endpoint and literature-comparison hygiene
Weaknesses
Single campaign per arm; no replicates
In vitro-only endpoint with a derived loading headline
DoE comparator is weaker than real-world practice
Human-in-the-loop step unquantified
No code, data, or model availability; core asset is proprietary
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is Andromeda 2 and what does it do?
Andromeda 2 is an AI agentic system from Intrepid Labs that designs self-emulsifying drug delivery system (SEDDS) formulations. It reasons over a company's structured historical lab evidence, proposes batches of candidate formulations with written rationales, checks them against compositional and robotic constraints, and has a miniaturized automated laboratory execute and measure them. On a paclitaxel benchmark it produced high-performing formulations roughly 3x as often as the company's previous probabilistic-optimization system.
Did Andromeda 2 improve oral bioavailability of paclitaxel in animals or patients?
No. The study's endpoint is an in vitro dissolution test: how well a formulation keeps paclitaxel apparently dissolved in simulated intestinal fluid (FaSSIF) over 240 minutes. The authors explicitly state this does not measure intestinal absorption, metabolism, or in vivo exposure, and that digestion-coupled tests and pharmacokinetic studies would be needed to establish translational value.
How much did the historical experimental evidence actually contribute?
In a controlled ablation, an otherwise identical version of Andromeda 2 without access to historical in-house evidence achieved 34% lower mean AUC (46.2 vs 62.0 mg·min/mL) and a 31% versus 50% hit rate. The ablated version was competitive in early batches but regressed later, suggesting the evidence mainly helped the system interpret and exploit new measurements during the campaign. Both results come from single runs, so uncertainty is not quantified.
Is the code, data, or the formulation list from this paper publicly available?
No public release is stated in the paper. Both Andromeda platforms and all underlying language and probabilistic models are described as fully in-house to Intrepid Labs and trained exclusively on internal experimental data, so independent reproduction is currently not possible from the published materials.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures are available only as captions and embedded axis labels in the extracted text; distribution shapes (Figure 1) and batch trajectories (Figure 3D/6D) could not be independently inspected.
- Supplementary sections A.1-A.4 appear complete, but any content beyond Table 3 may have been truncated in extraction.
- Assessment is based solely on the provided text; no access to underlying data, code, or the actual evidence corpus.
- References [16] (a 2026 ICML workshop version of related work by the same group) could not be consulted to check novelty claims against prior art.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.