The project established a multiscale framework from molecular information to many-body simulations. First, antibody structures were converted into amino acid-based coarse-grained models, where each amino acid is represented by one bead. Charges were assigned under experimentally relevant conditions using constant-pH Monte Carlo, which captures realistic protonation states and therefore heterogeneous charge patterns on the protein surface. These models are endowed with a detailed electrostatic representation while still being feasible for simulations of solutions. Second, to access formulation-relevant concentrations and longer times, antibodies were also represented as coarser, colloid-inspired bead models with a Y-shaped geometry. Specifically, few beads per domain were used, preserving the overall shape and excluded volume of the real antibody. These simplified models enable systematic scans of parameters, such as where positive and negative charges are placed, and how salt and concentration affect the overall solution behavior. In this context, machine learning was used not only as a technical shortcut but as a design tool. Neural networks were trained to infer bead-wise charge patterns from solution structure factors, enabling an inverse-design step that identifies coarse-grained charge maps consistent with experiments. Explainable analysis highlighted which electrostatic descriptors, such as dipole moments and localized negative patches, carry the strongest information for solution structure, helping to translate scattering curves into interpretable molecular features that can be tested in simulations or, further, in real antibody formulation development.
Core part of the work was the direct test of two different antibodies with distinct charge distributions in many-body simulations. Two different electrostatic representations were used, namely implicit-ion screened Coulomb interactions and explicit Coulomb charges including counterions and salt ions. By comparing simulations with experimental fingerprints from small-angle X-ray scattering (structure factors) and static/dynamic light scattering (compressibility and collective diffusion), the work showed that implicit screening can be adequate only for relatively uniform charge distributions, but it breaks down for strongly heterogeneous antibodies; in that case, explicit ions are required to reproduce experimental solutions structure and thermodynamics. Further theoretical analysis from liquid-state theory supported these findings.
The project also connected microscopic dynamics to macroscopic viscosity. Viscosity was computed directly from equilibrium molecular dynamics simulations using stress correlations by exploiting a Green–Kubo relation. Models with implicit ions could fail to reproduce experimental viscosity trends even when static structure showed appropriate agreement with experimental outcomes. Including heterogeneous charges together with explicit ions recovered the correct qualitative behavior and revealed a mechanism according to which viscosity is controlled by transient but strongly correlated antibody structures that persist long enough to slow stress relaxation. Additional ongoing work is aimed at systematically addressing how internal structural features such antibody hinge flexibility and hydrophobic interactions contribute to determine the raise in viscosity typically observed at high concentrations.
