In Silico Formulation Development: What CMC Teams Can Decide Before the Lab

Why are formulation and CMC decisions often made too late?

Critical formulation and CMC decisions are often made under high uncertainty and late in development, when significant capital is already committed and the options are limited. The lab methods to characterize a molecule exist. The problem is that, early in development, there is usually too little drug substance to run them broadly, and timelines are too tight to wait for the results. In silico formulation development offers a way to close that gap.

The result is a familiar trade-off for CMC leaders. Commit to a candidate and a formulation strategy before the data is in, or wait and lose time. Either way, liabilities that were present in the molecule from the start can surface late in development or at tech transfer, when the fix is reformulation and delay.

Coriolis’ in silico platform for formulatability assessment moves part of this work forward. Using only the primary sequence, it predicts liabilities and suggests formulation conditions before the main development investment is made. Lab work then focuses on a shortlist of candidates and conditions instead of a broad screen.

Which CMC decisions can in silico modeling inform?

The platform takes the amino acid sequence, predicts the 3D structure, computes more than 890 descriptors for monoclonal antibodies (mAbs) and more than 150 for non-mAb proteins, and combines machine learning, benchmarking and molecular dynamics. For a CMC leader, the useful question is what each output lets you decide, from in silico candidate selection to the starting point for formulation screening.

CMC question	What the in silico platform provides
Which biologic drug candidates should we take forward into formulation development?	Liability flags for self-association, viscosity, deamidation, isomerization and oxidation, and each candidate’s position against more than 600 clinical-stage or marketed mAbs or more than 100 marketed non-mAb proteins
Can this antibody be formulated at high concentration without aggregation or viscosity problems?	Self-interaction propensity from self-interaction by molecular dynamics (SIMOD), a coarse-grained simulation method, combined with machine learning models for viscosity and self-association
Which buffer, pH and excipients should formulation screening start from?	A formulation corridor covering concentration, buffer, pH and excipients, drawn from the closest marketed products and from physics-based simulations under different conditions
Where should limited drug substance and lab time be spent during early formulation work?	A ranked view of which liabilities need experimental confirmation, and under which conditions

In the published article, three antibody candidates received three different answers: one suited high concentration (a corridor up to 180 mg/mL) but needed control of deamidation; one should be kept at low concentration; one improved at lower pH. That is a triage result, not a screening plan.

How do molecular dynamics simulations guide pH and excipient selection for biologics?

For a CMC leader, there is a big difference between “this candidate is risky” and “this candidate is risky unless formulated at slightly acidic pH”. Only the second is something a program can act on. Machine learning gives the first statement: it flags a liability, but cannot say whether a different pH or excipient would remove it. Physics-based simulation under different formulation conditions, in effect in silico excipient and pH screening, is what gives the second.

In the antibody example from our article, two candidates (mAb 2 and mAb 3) were flagged by machine learning for self-association and viscosity at high concentration. Diffusion interaction parameter (k_D) measurements in histidine at pH 6 confirmed self-association for mAb 2. For mAb 3, k_D sat very close to the published cut-off, so the classification was uncertain.

Self-interaction by molecular dynamics (SIMOD) simulations were then run at pH 5, 6 and 7. The two molecules diverged:

• mAb 2 showed net attractive interactions at every pH tested. Formulation could not tune the problem out, so the recommendation was low concentration.
• mAb 3 was dominated by repulsion below pH 7. It has better formulatability than mAb 2, and could be taken to higher concentration if pH is kept slightly acidic.

The ReFOLD assay, which predicts long-term stability from the relative monomer yield, agreed: low yield for mAb 2 in all conditions, good yield for mAb 3 at pH 5 and 6.

The article’s second example, granulocyte colony-stimulating factor (G-CSF), shows the same logic for a non-antibody protein. Simulations with sucrose and NaCl predicted aggregation at pH 7, and at pH 3.5 when NaCl was present. An 11-day stability study using backgrounded membrane imaging matched those trends, pointing to acidic pH without salt.

How accurate are in silico formulation predictions compared with lab data?

Predictive tools are only useful in CMC decision-making when their accuracy and limitations are clearly understood. Both are described below.

The machine learning models for antibody liabilities are right in the large majority of test cases, and SIMOD sorted well-behaved from self-associating antibodies with >80% accurac.

As with any predictive approach, false positives can occur, and coverage is broader for antibodies than for other protein modalities, where fewer published datasets exist.

That is why we see in silico work as a way to focus the lab, not replace it. You get a clear rationale for every recommendation, and the lab gets a specific hypothesis to test instead of a broad screen.

What does this change for a biologics CMC program?

The effect is less about any single prediction and more about where capital and material go.

• Earlier triage. Assessments that relied on sequential, manual analysis now take days rather than weeks or months, and several candidates can be screened in parallel.
• Reduced drug substance consumption. Lab work starts from a defined formulation corridor, so less broad empirical screening is needed.
• Cleaner hand-offs. A manufacturing partner receives CMC-informed formulation strategies and clearer specifications up front. That means fewer development iterations, less technical ambiguity and better tech transfer readiness.
• Less late-stage risk. Identifying high-risk candidates early helps avoid reformulation and program delays downstream.

How can Coriolis help with in silico formulation development?

The assessment is available stand-alone, in a hybrid model with the developer’s preferred CMO, or as part of a fully integrated Coriolis formulation development package. Because Coriolis runs in silico prediction and wet-lab screening in the same team, predictions can move straight into experimental confirmation. Clear rules on IP ownership and no lock-in apply across all three.
The lab remains the final word on formulation. The change is that it is asked better questions, earlier.

References

1. Arsiccio A, Le Vay K, Sao Pedro A, et al. A novel in silico platform combining data-driven and physics-based models for protein formulation developability assessment. Journal of Pharmaceutical Sciences 115 (2026) 104196. Open access.