The era of running hundreds of formulation experiments to find a good-enough answer is ending. What's replacing it is a smarter approach: adaptive, instrument-enabled, and designed to find the right answer, not just a workable one.

Prometheus Panta is built for exactly this kind of workflow. Its combination of nanoDSF-based thermal stability measurement and DLS-based colloidal stability characterization, from the same sample in the same run, with the throughput and precision that machine learning demands, makes it the biophysical foundation for next-generation formulation development. 

Ready to bring quality in, and get quality out? 

What this means for your formulation program

The broader lesson: Optimal formulation is not generic. It's molecule-specific, and it requires the kind of high-quality, systematic experimental data that biophysical tools like Prometheus Panta are built to generate.

Beyond identifying optimal formulations, the Bayesian approach generated a rich dataset that shed light on why the algorithm made the choices it did, and some of the findings challenged conventional formulation wisdom. 

pH converged at ~6. This reflects a fundamental trade-off: higher pH tends to improve thermal stability (Tm), while lower ph drives better colloidal stability (kD) by moving the molecule away from its isoelectric point and increasing intermolecular repulsion. The algorithm balanced both by consistently converging near pH 6, closely aligned with the pH range (~5.8) seen in most commercially approved mAb formulations.

Sorbitol settled at ~300 mm. Intermediate-to-high sorbitol concentrations consistently improved Tm, a well-characterized effect since the 1980s, explained by the preferential exclusion of polyols from the protein surface, which promotes a more compact and thermodynamically stable protein structure.

Arginine was consistently excluded. This is the headline finding. Arginine is one of the most widely used excipients in biopharmaceutical formulation, and yet, for this molecule, the algorithm avoided it in nearly every optimization round. Across all three target properties, arginine had a negative effect.

The likely mechanisms: at formulation-relevant pH, arginine is positively charged, significantly increasing ionic strength and reducing the repulsive intermolecular interactions that maintain colloidal stability. Its guanidinium group also has known chaotropic properties that can destabilize protein structure, contributing to reduced Tm.

This is precisely the kind of system-specific insight that prior-knowledge-based formulation design misses. Arginine works well for many molecules. For Bocutizumab, it doesn't, and only a data-driven approach that interrogates the actual behavior of the actual molecule would have revealed that. 

The surprising story the data told about Arginine

This is the practical meaning of "quality in, quality out." The machine learning model doesn't know your instrument. It only knows the numbers you give it. Prometheus Panta's measurement precision is what makes those numbers trustworthy, and that trustworthiness is what makes the entire optimization framework function.

Isabel Waibel

Scientist, ETH Zürich

"Garbage in, garbage out": Why measurement precision makes or breaks the model 

The Results

After 33 total experiments (13 for initialization, then four rounds of five) the algorithm identified near-Pareto optimal formulations. That's less than half the experiment count of comparable doe methods, and crucially, it comes with a guarantee of global search coverage that static doe methods cannot offer.

In terms of speed: Tm and kD measurements on Prometheus Panta take under one hour per round for up to 48 samples in parallel. The rate-limiting step in this study was the agitation assay, which requires one week of shaking at 1 mg/mL. With alternative interfacial stress assays, even that bottleneck could be removed, making the full optimization achievable in a single day.

How does Bayesian optimization work?

Bayesian optimization is a sequential, adaptive approach to searching complex design spaces. Here's how it works in practice:

• Initialize: Run a small set of experiments (13 in this study) covering the design space broadly.

• Train: Use the results to train surrogate models: in this case, Gaussian processes, that learn the relationship between formulation variables and stability outcomes.

• Suggest: The model recommends the next experiment by balancing two strategies:

  • Exploitation (75%): Focus on regions of the design space that have already shown promise, guided by the Pareto front of the surrogate models.

  • Exploration (25%): Send experiments into novel, uncharted regions to ensure global coverage and avoid local optima.

• Repeat: Measure the new formulations, retrain the models, and iterate.

To make this experimentally practical, the team used the Kriging Believer approach, retraining the model four times with simulated results to generate five new formulation suggestions per round, all of which can be prepared and measured in parallel. 

The optimization simultaneously improved three properties: thermal stability (Tm), colloidal stability (kD), and resistance to agitation-induced aggregation (remaining monomer after shaking). Osmolality was incorporated as a hard constraint, keeping all generated formulations within the physiologically tolerable range of 100–600 mosm/kg.

What are the current challenges with traditional Design of Experiments (DoE)?

Conventional DoE approaches such as central composite design or Box-Behnken design, fix the experimental plan at the outset. You define your variables and levels upfront, run the full matrix, and analyze the results. This works reasonably well for simple, well-understood systems. But for complex biologics in large design spaces, it has two significant weaknesses:

• Local optima: With a pre-fixed plan, DoE has no mechanism to redirect experiments toward areas that look promising. You can easily converge on a locally good formulation while missing the global optimum entirely.

• Rigidity: The method cannot learn from early results and adapt. Every experiment is equally uninformed. 

From screening hundreds of formulations to finding the optimum in 33 experiments

The flag-based risk ranking

To make the complexity of 15 assays actionable, the team adopted a "flag" system: each time a molecule exceeded a risk threshold for any measured property, it received a flag. The more flags, the higher the overall developability risk. 

The results were instructive. Canonical mabs accumulated the fewest flags, confirming their relative robustness. But the data also showed that format family is not destiny: some bispecifics and scFv variants performed nearly as well as traditional mabs, while scFv-scFv tandems showed the highest risk across the board.

Long-term stability (12 months)

At 30°C, most variants remained stable, though scFv-scFv fragments showed some aggregation. At 40°C, full-length mabs outperformed fragments and bispecifics across the board, and significant fragmentation was observed and attributed largely to GS linkers, which can be optimized further.

Colloidal stability

The diffusion interaction parameter (kD) was also measured by Prometheus Panta and told a more differentiated story. Small fragments, particularly scFv and scFv-scFv constructs, showed the most pronounced colloidal stability challenges. Critically, engineered formats showed far greater within-group variation than canonical mAbs, underscoring how sensitively these formats respond to individual mutations.

15 assays, 73 constructs: Here's what the data told us 

The key finding: Engineered antibody formats carry higher developability risk than traditional mabs, but the specific liabilities vary widely by format and by molecule. That variability is precisely why tailored, data-driven formulation strategies are essential.

But these formats come with a critical liability: they didn't evolve. Their structures are engineered, not optimized by natural selection, and the mutations that give them their therapeutic properties often come at a cost to intrinsic stability. The field also lacks the systematic developability datasets that exist for canonical mabs making it far harder to predict where problems will emerge, or how to fix them.

That gap is exactly what Isabel and her collaborators at Novo Nordisk set out to address.

• Improved therapeutic efficiency — Smaller fragments diffuse faster, achieve better tissue penetration, and are more likely to cross the blood-brain barrier.

• Multi-specific targeting — Bispecific and multi-specific formats can engage several molecular targets simultaneously, opening entirely new therapeutic approaches.

• Lower production costs — Many engineered formats can be produced in prokaryote or yeast systems rather than the expensive mammalian cell lines required for full-length mAbs.

For the better part of three decades, monoclonal antibodies (mAbs) like Adalimumab (Humira) have been the backbone of biologics development. We know how to make them, stabilize them, and formulate them. The field has accumulated a rich body of historical data that makes formulation design, while never trivial, at least navigable.

That foundation is now being tested. Engineered antibody formats including bispecifics, scFv fragments, and multi-domain constructs offer genuine therapeutic advantages that have the field's attention:

Beyond Humira: The promise – and the problem – of next-gen antibody formats 

by Léa Valla      —      10 minute read      —      August 13, 2026

Companion post to our webinar featuring Isabelle Waibel, PhD candidate, ETH Zurich.

Quality in, quality out: How biophysical data powers AI-driven formulation optimization for next-gen biologics