Meet COSMOPharm: A Molecular Head Start for Drug–Polymer Formulation

A new experimental Neurosnap service brings molecular thermodynamics into the search for useful polymer carriers.

A promising drug molecule can generate a surprisingly long formulation to-do list. Choose a polymer. Explore the drug loading. Work out how temperature changes the picture. Each decision adds experiments, consumes material, and raises another question about how the molecules will behave together.

Our new COSMOPharm service gives scientists another way into that problem: estimate equilibrium drug solubility in a polymer from molecular structure and crystal properties, then use that information to sharpen the next experiment.

The initial service supports PVA, PVP K12, and PVP K30. It is experimental, with encouraging results for several tested combinations and a substantial unresolved discrepancy for another. Both belong in the story.

Why formulation scientists should care

Amorphous solid dispersions are one strategy for addressing poor drug solubility, and polymer selection is central to their design. The thermodynamic compatibility of the drug and polymer helps determine how much drug can be accommodated before a crystalline phase becomes favorable. The original COSMOPharm study by Antolović, Vrabec, and Klajmon applies molecular thermodynamics to this challenge.

For a scientist planning a formulation screen, a calculated solubility curve can help organize questions: which temperature range deserves attention, where a proposed loading sits relative to estimated saturation, and which drug–polymer combinations merit closer investigation.

That is the opportunity we want to make accessible on Neurosnap. Start with a defined scientific question, explore the molecular picture, and take a more informed hypothesis to the bench.

From a molecular structure to a solubility curve

Figure 1. The Neurosnap approach combines a drug structure, an approximate polymer model, and supplied crystal properties. The surface profile and example loading curve shown here come from paracetamol calculations; the polymer chain is schematic.

You begin with one supported neutral drug molecule, a polymer choice, and a temperature range below the drug’s melting point. You also supply the melting point and enthalpy of fusion for the relevant crystal form. Those properties matter: the calculation must account for both the drug’s interactions with the polymer and the energetic cost of leaving its crystal.

The molecular workflow prepares and refines a three-dimensional geometry, then uses PySCF quantum chemistry to calculate the molecule’s response to a conducting environment. From that calculation, we obtain a sigma profile: a distribution of surface charge density that summarizes features relevant to molecular interactions, including hydrogen bonding.

For the polymers, we use our own generated oligomer surfaces and approximate longer chains by replicating a central repeat unit. COSMO-SAC then estimates nonideal mixing behavior, and COSMOPharm combines that information with the crystal properties to find candidate saturation loadings across temperature.

We use interaction parameters matched to this PySCF profile-generation approach, credited to Tae Hyun Kim, without refitting them to our drug–polymer benchmark. The service offers both Staverman–Guggenheim and free-volume variants; the latter requires suitable drug volume inputs. Optional heat-capacity corrections can refine the crystal contribution.

The result is a temperature-dependent estimate of equilibrium drug loading, accompanied by an indication of whether the calculation found a numerically acceptable candidate.

Putting the predictions under pressure

A convincing curve starts with checks on the calculation itself. We checked that molecular surfaces remain finite and conserve area, that polymer repeat units map correctly, and that chain replication preserves the intended molecular description. We also verified fusion-property units, mass-to-mole conversions, and thermodynamic limiting cases with known answers.

Actual molecular calculations exercise the complete scientific workflow. Saturation candidates are checked against the equilibrium equation and screened for liquid instability on a composition grid. Multiple roots, missing roots, or failed checks produce an explicit inconclusive result rather than a selected loading.

Then comes the harder question: how closely does the model match measurements?

Figure 2. Our PySCF-based adaptation was evaluated against 22 archived measurements across four systems. Bars show mean absolute errors in drug loading, in percentage points. Both variants are shown; neither performs best for every system. Experimental values are from the COSMOPharm data archive.

For the free-volume variant, mean absolute errors were approximately 6.0 percentage points for ibuprofen–PVA, 4.4 for ibuprofen–PVP K12, and 5.5 for paracetamol–PVP K30. Paracetamol–PVP K12 was considerably harder, with an error of 23.9 percentage points.

These are absolute differences in loading: a prediction of 40 wt% against a measurement of 50 wt% differs by 10 percentage points. The benchmark errors are not confidence intervals for a new formulation.

The difficult case is useful information, too

Figure 3. All nine paracetamol measurements in the benchmark, compared with the free-volume predictions. Lines connect predictions at measured temperatures. The two datasets have different sources, and their original sample conditions have not been fully independently verified. Archived observations are credited to Antolović and colleagues’ published COSMOPharm collection.

The K12 discrepancy remained after auditing units, crystal-property calculations, and composition conversions. Tested variations in chain length, density, and free volume did not close the gap. Its cause remains unresolved.

That finding sets a useful boundary. A reasonable result for one drug–polymer pair does not establish accuracy for another, even when the polymer chemistry looks closely related. Broader validation needs more independent measurements and better characterization of the experimental materials.

Where COSMOPharm fits in your workflow

Use the service to explore equilibrium saturation hypotheses for supported drug–polymer systems, with measured crystal properties and the model’s assumptions in view. The initial scope is one neutral, closed-shell drug with 2–24 heavy atoms within the supported element set. The current polymer representation uses a single oligomer conformer and nominal grade properties; it does not capture every lot, molecular-weight distribution, or moisture condition.

The predicted quantity is drug loading in the dry polymer. Aqueous solubility and dissolution performance require separate evidence. Equilibrium solubility also leaves important formulation questions open. Shelf life, crystallization rates, glass transition, processing history, and complete liquid–liquid phase behavior require additional evidence. A finite-grid stability screen cannot establish the full phase diagram.

COSMOPharm on Neurosnap is an invitation to bring molecular insight into formulation screening. Explore a candidate loading, examine the temperature dependence, and use the result to make your next experiment more deliberate—with the validation limits visible from the start.

Scientific credit: COSMOPharm was developed by Ivan Antolović, Jadran Vrabec, and Martin Klajmon. The profile-generation approach and fitted interaction parameters are credited to Tae Hyun Kim. Published parameter and experimental-data sources used here are acknowledged under CC BY 4.0. Figures are original Neurosnap illustrations and visualizations of our validation results.