Thirty Days of Release, One Needle to Pass

TL;DR for operators

Developing a monthly injectable medicine requires more than slowing drug release. The formulation must also pass through a needle, form a stable depot, and release its active ingredient at an acceptable rate. In an industrial study, Bannigan and colleagues¹ combined automated experimentation with AI-guided optimization to evaluate 181 formulations over approximately 15 weeks, identifying four leads with acceptable reported administration properties and differentiated 30-day in vitro release profiles.

The stronger efficiency numbers come from a separate retrospective simulation. The AI-guided system recovered at least one previously identified lead in 30 of 30 simulated campaigns, versus 19 for random search and 17 for conventional design of experiments. Estimated peptide consumption was also lower among successful simulated campaigns.

For pharmaceutical R&D teams, the study demonstrates a workable approach to searching across competing formulation requirements. It does not yet establish that the AI system outperforms conventional methods in independent prospective laboratory trials.

A formulation that releases slowly may be too thick to inject

Consider a medicine intended to be administered once a month. Its formulation needs to release the therapeutic ingredient gradually rather than delivering too much immediately after injection.

A stronger polymer matrix might retain the drug longer, but the same change can increase viscosity. If the formulation becomes too thick to pass through a fine needle, better release control has not produced a usable injectable product.

This is the difficulty behind developing PLGA in situ forming depots. A liquid containing a biodegradable polymer and a drug is injected into the body, where solvent exchange causes a polymer-based reservoir to form. The drug then escapes progressively as the depot evolves and degrades.

The formulation team must manage several coupled requirements: early drug release, month-long release behavior, drug loading, viscosity, physical stability, and injection compatibility. Improving one attribute can impair another.

The experimental results illustrate why this matters. Formulations with similar initial drug release could have substantially different viscosities. A candidate selected solely for its release performance might therefore fail an administration requirement.

The resulting problem is multi-objective optimization: finding combinations that satisfy several requirements simultaneously, rather than maximizing one desirable characteristic.

Four leads emerged from 181 laboratory formulations

Corbion and Intrepid Labs assembled a library of 15 PURASORB polymers, including commercially available and developmental materials.

Intrepid Labs’ VALIANT automated platform prepared and characterized peptide-containing formulations. Its proprietary ANDROMEDA 1 optimization engine incorporated experimental observations, modeled formulation-performance relationships and predictive uncertainty, and guided subsequent experiment selection.

Over approximately 15 weeks, the team tested 181 unique formulations spanning drug loadings of 6–12% by weight.

The target product profile established several requirements before optimization:

Source: Table 1, page 3. These are development objectives, not achieved measurements for every formulation.

Four leads—F136, F139, F148, and F162—met the reported viscosity and injectability requirements while exhibiting distinct 30-day in vitro release profiles. Their drug loadings included 6%, 9%, and 12%.

Importantly, three of the four leads incorporated developmental polymers rather than materials from the existing commercial portfolio. Most formulations satisfying both the viscosity and early-release criteria also used developmental materials.

This result points to a complementary role for materials innovation and AI-guided search. A more sophisticated selection algorithm cannot discover an attractive formulation if the necessary polymer properties are absent from its available materials library.

Within the tested library, intermediate-molecular-weight polymers provided a favorable compromise between sustained release and viscosity. Choosing the highest-molecular-weight material was not generally the best way to satisfy the full target profile.

These are directly observed development outcomes. They establish that the integrated workflow could identify feasible candidates in this experimental setting, without isolating how much of the result came specifically from the AI selection strategy.

The model identified different drivers of release and viscosity

The researchers used SHAP feature importance and partial-dependence analysis to examine the formulation-performance relationships learned from the experimental dataset.

Two results help explain why a single-variable formulation strategy would be inadequate.

Polymer loading was the dominant modeled predictor of 24-hour burst release, accounting for approximately 44% of total mean absolute SHAP importance. Higher polymer loading was associated with lower predicted burst release.

Solvent fraction dominated modeled viscosity, accounting for approximately 64% of corresponding SHAP importance. Increasing solvent fraction was associated with sharply lower predicted viscosity.

Molecular weight also mattered. Higher polymer molecular weight was associated with increased viscosity, while higher molecular weight and residual monomer content were associated with greater predicted early burst release in the fitted models.

These relationships identify different controls over different aspects of formulation performance. Raising polymer loading may help suppress early release, but it can increase viscosity. Increasing solvent content may reduce viscosity while also changing depot formation and release behavior.

The supplementary partial-dependence analysis extended this interpretation to zero-order release deviation and the overall optimization score, finding additional nonlinear relationships.

These are model-based associations, not independently identified causal effects. SHAP allocates predictive importance within a fitted model, while partial dependence describes modeled responses to changes in input attributes. Neither establishes that changing one ingredient in isolation will produce the predicted effect under every manufacturing condition.

For experimental teams, their immediate value is in identifying promising regions and variables to test next.

The 30-of-30 advantage came from a virtual laboratory

After completing the physical experiments, the researchers conducted a retrospective comparison of three search strategies.

They trained an oracle model on the completed formulation dataset. This predictive model served as a virtual laboratory, allowing search strategies to simulate experiments without preparing additional physical formulations.

ANDROMEDA 1, random search, and conventional design of experiments (DoE) were each evaluated across 30 simulated campaigns. Each campaign had a maximum budget of 160 experiments, organized into batches of 16.

Success meant recovering at least one of the four leads already identified in the original campaign.

Source: Figure 4, pages 6–7, and Supplementary Table S1, page 10. Peptide-consumption averages cover only simulated campaigns that successfully reached a lead. Experiment counts are the paper’s reported first-lead comparison and should be read in light of the campaigns that exhausted their budgets without success.

The difference in simulated discovery rates is substantial. ANDROMEDA 1 found at least one known lead in every campaign, while both alternatives failed to recover any lead in more than one-third of their campaigns.

The individual-lead results provide additional detail. F148, the candidate with the most favorable reported multi-objective score, appeared in 28 of 30 ANDROMEDA 1 campaigns, compared with six random-search campaigns and three DoE campaigns.

Supplementary Table S1 therefore supports a more specific conclusion than overall lead discovery alone: within the simulated environment, ANDROMEDA 1 repeatedly found a particularly attractive known candidate. The supplementary results are a breakdown of the same retrospective comparison, not independent experimental validation.

The critical distinction is that these simulated campaigns were not 90 new laboratory development campaigns. The oracle was learned from the completed experimental dataset, and the success criterion concerned recovering previously identified formulations.

The benchmark supports a favorable within-model comparison of search behavior. It cannot establish how frequently the advantage would persist when all three strategies face previously unobserved experimental outcomes.

Lower material consumption is promising, but conditional

The retrospective benchmark also estimated total material consumption to first-lead discovery.

ANDROMEDA 1 required approximately 92 g of materials among successful campaigns, compared with 141 g for random search and 134 g for DoE.

For expensive therapeutic peptides, even relatively small reductions in experimental consumption could affect development budgets and the number of formulation candidates a team can afford to explore.

However, the averages exclude comparator campaigns that never found a lead within their simulated budgets. They are therefore not unconditional estimates of the materials consumed by an entire development program.

Cognaptus inference: pharmaceutical R&D managers developing costly peptide formulations have a reasonable basis to test whether model-guided experimentation can improve lead discovery per gram of active ingredient consumed.

The appropriate procurement comparison would use prospective laboratory campaigns under matched constraints, tracking successful lead discovery, actual material consumption, laboratory effort, and total elapsed development time.

The observed 15-week project duration is a description of the campaign, not a measured reduction in development time relative to conventional practice.

Formulation success is not clinical validation

The evidence has three principal boundaries.

First, the experimental findings concern one therapeutic peptide and one particular polymer library. Different drugs, solvents, polymer systems, and manufacturing processes may change the feasible formulation region.

Second, ANDROMEDA 1 and the retrospective oracle are insufficiently specified for independent algorithmic reproduction. The paper does not disclose enough about the optimization architecture, acquisition functions, hyperparameters, or oracle validation to determine whether the simulated advantages would survive alternative modeling assumptions.

Third, the four leads were evaluated through in vitro release and formulation properties. The paper does not establish in vivo pharmacokinetics, clinical safety, therapeutic efficacy, long-term storage stability, or prospective manufacturing scale-up.

These boundaries imply separate development decisions. Formulation teams can use the study to inform how they combine material libraries, feasibility constraints, and adaptive experimentation. Technology purchasers still need prospective comparisons before attributing a specific efficiency gain to the AI engine. Product-development teams must then validate stability, manufacturability, and biological performance before advancing a candidate toward clinical use.

The next advantage must be demonstrated experimentally

The most persuasive contribution of this study is the completed industrial workflow: polymer expertise, automated formulation experiments, and iterative optimization produced four differentiated candidates satisfying reported administration requirements.

The retrospective benchmark strengthens the case for testing model-guided search, particularly when experimental materials are expensive. Its unusually favorable discovery rates are measurable evidence within a virtual laboratory, but not a substitute for independent prospective comparison.

For pharmaceutical R&D, the next step is to establish whether the same advantage persists when the optimizer must make decisions using genuinely new experimental feedback, with material use, lead quality, and practical feasibility measured together.

That is the evidence needed to turn a promising optimization workflow into a defensible development investment.

Cognaptus: Automate the Present, Incubate the Future.