What a formulation robot can learn from yesterday’s experiments

A useful laboratory memory should change the next experiment. Storing a measurement is easy; recognizing when an older result should influence a new formulation is the harder task. Andromeda 2, described in a September preprint by Michael Craig and colleagues at Intrepid Labs, makes that distinction experimentally testable. It couples formulation proposals to an automated laboratory and asks whether accumulated evidence improves the use of a fixed experimental budget.¹

The most informative result is not an exceptional winning formulation. At an equal budget, the earlier probabilistic system reached almost the same maximum assay score. The newer system instead produced a better distribution of candidates and more formulations satisfying the complete specification. That changes the practical reading: this is evidence about where experimental effort goes, rather than proof that a new agent alone can reach a previously inaccessible chemical optimum.¹

Follow the sample before judging the planner

The study concerns self-emulsifying drug-delivery formulations containing paclitaxel. The experimental space combines excipient identities, composition and drug loading. Each campaign ran six batches of sixteen formulations. Robotic preparation fed into dispersion and dissolution measurements, particle-size characterization and chromatographic drug quantification. The physical comparison used the same laboratory, assay and experimental budget for Andromeda 2, its predecessor Andromeda 1, and an experimentally executed design-of-experiments campaign.¹

This common measurement chain is a substantial strength. An optimization comparison becomes difficult to interpret if one algorithm receives cleaner measurements or a more permissive composition space. Matching the laboratory and constraints makes the observed difference more directly relevant to selection. It does not make every information input identical: historical evidence is deliberately part of the newer system's advantage. The distinction matters when deciding which claim the experiment answers. The design-of-experiments arm also explored the shared composition grid without a separate expert excipient-solubility prescreen; it represents that implemented comparator, not every conventional formulation-development strategy.¹

The main endpoint integrates apparently solubilized paclitaxel concentration from ten to 240 minutes in fasted-state simulated intestinal fluid, following an initial gastric-fluid stage. HPLC supplies the concentration measurements. This area under the concentration–time curve describes sustained performance in the assay. It is not the blood concentration–time curve used to quantify systemic drug exposure.¹

Calling both quantities AUC can obscure the difference. Here, a high result says that more drug remained apparently solubilized over the measured interval under the stated conditions. It does not measure passage across the intestine, metabolism or the concentration reaching a tumour. Those are subsequent questions. Keeping the endpoint attached to the assay gives this experiment its proper value: a measured formulation-selection test that can identify candidates for further development.¹

A winning sample and a productive campaign

Table 1 reports median assay AUCs of 70.1, 12.0 and 3.5 mg·min/mL for Andromeda 2, Andromeda 1 and experimental design of experiments, respectively. Yet the two Andromeda maxima were 157.7 and 156.4 mg·min/mL. Each campaign contained 96 unique formulations. The separation between typical and best performance is therefore central to interpreting the result.¹

A laboratory choosing between strategies should care about both. If the sole objective is to recover one exceptional candidate and all follow-up properties are equivalent, similar maxima are meaningful. If development benefits from several plausible alternatives, an upward-shifted distribution can be valuable even without a higher record. Alternative formulations can become useful when later tests introduce a new constraint. That is a practical interpretation of the distribution, not a claim that those later advantages were measured here.

The paper also reports how often candidates reached the pooled upper quartile of measured AUC: 48 of 96 for Andromeda 2, sixteen of 96 for Andromeda 1, and two of 96 for experimental design. This relative threshold helps compare allocation within the experiment. Because the threshold is defined from the pooled results, it should not be read as a universal pharmaceutical quality standard.¹

The separate target product profile is more demanding. It combines an integrated dissolution threshold, a late concentration floor, a late fraction-dissolved floor and resistance to precipitation. Twelve Andromeda 2 formulations met all four requirements, compared with six for Andromeda 1 and none for experimental design. The highest-AUC candidate satisfied only two requirements. Optimizing a single summary score and satisfying the intended product profile were visibly different achievements.¹

That is an unusually useful result to report. Multi-objective development can be misrepresented by selecting whichever endpoint looks most favourable after the experiment. Here the complete specification allows the reader to see the trade-off. The experiment supports a broader pool of acceptable candidates under this profile; it does not justify replacing that profile with maximum AUC when describing success.

What was actually autonomous?

Andromeda 2 organizes proposal, criticism, review and planning around a shared formulation context. It can use computational models, including the probabilistic models available to its predecessor. A separate deterministic layer checks executable compositions, permitted ingredients, composition totals, distinctness and robotic readiness. A formulation scientist inspected proposed batches for scientific reasonableness and feasibility without substituting or retuning their compositions.¹

This division is scientifically sensible. Choosing a promising experiment and verifying that a robot can execute its specification are different tasks. A plausible verbal explanation cannot certify a composition total or instrument constraint. By giving those checks an explicit engineering role, the platform makes the passage from proposal to experiment easier to audit. The human inspection also belongs in the description of autonomy; it is part of the reported workflow.

After a batch, measurements returned to the working context. The authors report no model retraining between batches. Adaptation therefore occurred through new experimental information supplied to the agent, rather than updated model weights. That is a concrete account of learning during the campaign and avoids treating every improvement as evidence that the underlying language model changed.¹

The distinction has implications for reproducibility. To understand a later proposal, one needs the evidence then available, the allowed tools and the execution constraints, not merely the model's name. The relevant experimental object is the coupled decision process. This is an inference about how such systems should be evaluated, rather than a claim that the public manuscript exposes every internal implementation detail.

The experiment that makes memory matter

A reduced-evidence Andromeda 2 campaign withheld historical in-house experimental evidence while retaining the target, executable space, constraints, batch size and wet-lab workflow. Both versions still received results from their own current campaign. Mean assay AUC was 62.0 with historical evidence and 46.2 mg·min/mL without it, a reported increase of 34%. The reduced-evidence arm remained capable of producing executable formulations.¹

This ablation asks a narrower and stronger question than comparing two differently constructed systems. It tests the contribution of prior experimental evidence within the newer workflow. It does not assign the entire difference between Andromeda 2 and Andromeda 1 to memory. The system comparison includes other differences, while the ablation more directly targets one input.¹

The trajectory is also interesting: the reduced-evidence campaign was competitive early but failed to sustain that performance. The authors interpret this as consistent with historical evidence helping the system use newly arriving measurements, rather than merely supplying a better first guess. “Consistent with” is the right level of confidence. The observed sequence supports that interpretation without revealing a uniquely proven internal reasoning mechanism.¹

Historical evidence contained no paclitaxel formulations at campaign initiation. The task therefore tested transfer from broader formulation experience to a new active ingredient, followed by adaptation using that ingredient's measurements. It was not a campaign devoid of prior knowledge. The scientifically useful question is whether prior knowledge travels productively across that boundary.¹

The longer history of learning from formulations

Earlier work by Bannigan and colleagues illustrates a different way to reuse formulation measurements. Their 2023 study trained models to predict release from polymeric long-acting injectables using a literature-derived dataset, with evaluation grouped by drug–polymer combination. It subsequently prepared and measured new formulations informed by the model analysis.²

This is relevant context rather than a head-to-head competitor. The injectable study predicts a release profile; Andromeda selects successive experiments in another formulation system. Their errors, assay endpoints and development goals cannot be ranked against each other. Together they illustrate two distinct uses of historical data: estimating a property for a candidate and deciding which candidate should consume the next experimental slot. Improvement in one task does not automatically establish improvement in the other.

The earlier study's grouping strategy also emphasizes a useful principle: the test boundary should follow the intended transfer. For Andromeda, withholding prior paclitaxel measurements addresses transfer to that active ingredient. Repeating campaigns and extending to chemically different ingredients would address additional questions. These are complementary tests with different units of generalization, not interchangeable evidence of universality.²¹

Count campaigns as well as formulations

The strongest qualification is statistical and is stated by the authors themselves. There was one independently initialized campaign per strategy. Later samples were chosen using earlier outcomes, so the 96 formulations in an arm are not 96 independent repetitions of the optimization strategy. Supplement A.1 treats the formulation-level significance calculation as descriptive, and the main methods apply the same caution to bootstrap intervals.¹

That transparency deserves credit. Many points in a plot can create an impression of replication that the experimental design does not provide. The present evidence shows a substantial difference between the observed campaigns. Independent campaign repetition would test how consistently that difference recurs when starting conditions and sequential choices vary. It is the appropriate extension for a claim about strategy reliability.

The supplementary composition analysis strengthens the mechanistic account of search allocation. Andromeda concentrated on productive formulation classes, but also achieved a better hit rate within those classes. The authors therefore provide evidence against the simplest explanation that the system merely found a favourable broad region. This remains a descriptive analysis of the same campaigns, rather than a separate replicate of the result.¹

The simulated design-of-experiments curves warrant a different evidential weight from the physically executed comparator. Their emulator was checked on held-out formulations, and the supplement reports poorer ranking for composition families absent from its data. A simulation can provide useful directional context while remaining dependent on the surrogate landscape. It should not be added to the number of wet-lab campaigns when judging experimental replication.¹

The study's contribution is consequently precise and worthwhile: within a controlled physical workflow, access to accumulated evidence helped an agent allocate experiments more productively, while a complete product specification exposed differences a maximum score would conceal. Its next test is whether that advantage persists across independent campaigns and new formulation challenges. The robot's useful memory is the one that repeatedly earns its place in the next measured decision.¹