Interesting to know: Frequently asked questions

How does the modeling approach work?

The in-silico model requires key training data to characterize the net strength of the drug molecules’ intermolecular interactions with its environment. The combination of key experimental data and the physical model allows predicting molecular properties with minimal experimental effort.

What is the benefit compared to Machine Learning tools?

Succesful machine learning models require thousands of data points as these algorithms need to elaporate physical principles and laws from the training data. Our approach already contains physicochemical priciples and thus requires only few training data points, a true benefit at early stage of drug discovery.

How many data points are required for setting up the model?

The training set should comprise at least 5 experimental values, e.g. solubilities in different organic solvents. The solubility data in different organic solvents allows deriving the most propable strength of the individual intermolecular interactions (e.g. hydrogen bonding strength, van-der-Waals interactions).

The drug molecule is a poor crystallizer - can it be characterized as well by the model?

Solubilities of weakly crystallizing molecules are usually hard to determine – an excellent alternative is considering other thermopysical properties of the molecule instead, e.g. dynamic vapor sorption data or partitioning coefficients.

Can the model also handle a modern complex drug candidate?

We observed that the intermolecular interactions especially of highly complex molecules are very well described.

What kind of intermolecular interactions are considered?

The model considers diverse contributions to the overall intermolecular interaction profile: van-der-Waals interactions, hydrogen bonding, polarity, charge…

Do you need to know the chemical structure to set up the model?

The chemical structure is not required for the modeling!

What type of experimental data is required as training data?

The input requirements are flexible: Possible input data are crystalline solubilities, miscibilities, partitioning coefficients, sorption data, osmotic data, density data and many more. We aim at reducing the experimental effort and thus try to set up the model with the data that is already available from early-stage screenings.

What can be predicted by the model?

The model can predict any thermodynamic property of a formulation, e.g. thermodynamic equilibria (sorption, crystalline and amorphous solubilities) and also metastable states (e.g. crystallization risk in metastable ASDs, driving forces to undergo a phase change).

Also kinetic properties (e.g. crystallization onset times or release rates) are predictable by a combination of the thermodynamic properties with information about the molecular mobility.

This allows performing excipient screenings, amorphous solid dispersion screenings, humidity risk assessments, shelf-life predictions, solvent screenings, dissolution behavior predictions and many more.

How fast is a calculation?

A typical phase diagram within few moments. This is highly advantageous compared to computationally more intensive predicitons as full screenings may be realized within an economic time frame.

Can you also predict ternary ASDs?

The number of components in a mixture is not limited: A typical binary formulation (drug+ excipient) works just as good as a highly complex formulation containing many different excipients.

What substance classes can be modeled?

Any kind of substance such as solvents, gases, drugs, polymers, sugars, amino acids, lipids, ….