MP2D: Multi-Property Protein Diffusion

• MP2D is a framework that models protein motion in heterogeneous environments by simultaneously coupling multiple properties like diffusivity, crowding, and domain affinity.
• It integrates methods such as Langevin dynamics, phase-field simulations, and hydrodynamic corrections to capture anomalous diffusion and phase partitioning in membranes.
• The approach extends to non-equilibrium systems and sequence-space optimizations, offering versatile insights for both biophysical studies and protein design applications.

Searching arXiv for papers and the provided arXiv IDs to ground the article. arxiv_search(query="Multi-Property Protein Diffusion MP2D protein diffusion membranes", max_results=10) arxiv_search(query="Multi-Property Protein Diffusion MP2D protein diffusion membranes", max_results=10) Multi-Property Protein Diffusion (MP2D) denotes a class of formulations in which protein motion is controlled simultaneously by multiple coupled properties rather than by a single homogeneous diffusion coefficient. In membrane biophysics, the term is used for frameworks that combine composition-dependent mobility, crowding, domain affinity, concentration or temperature fields, heterogeneous energy landscapes, and hydrodynamic or surface effects to explain fluctuating diffusivity, anomalous transport, and partitioning in heterogeneous environments (Sakamoto et al., 2023, Jasuja et al., 28 Jun 2025, Reinhardt et al., 2020, Reister-Gottfried et al., 2010, Girelli et al., 2024). In a distinct sequence-design usage, the same acronym labels a discrete diffusion framework for multi-objective protein sequence optimization, where “diffusion” refers to denoising in sequence space rather than physical transport (Kong et al., 7 May 2026).

1. Terminology and conceptual scope

In the physical transport literature, MP2D is most explicitly formulated for heterogeneous biological membranes that phase separate into liquid-ordered ${\rm L_o}$ and liquid-disordered ${\rm L_d}$ domains. In that setting, proteins move through a two-dimensional medium whose local composition, and therefore local mobility, changes in space and time. The central idea is that transport, localization, and reaction propensity are jointly regulated by several properties: local diffusivity, interparticle interactions, domain preference, concentration, temperature, and microstructural energy landscapes (Sakamoto et al., 2023, Jasuja et al., 28 Jun 2025).

A common source of ambiguity is nomenclature. The acronym is not reserved for a single formalism. One strand concerns physical diffusion in heterogeneous media, including phase-separated membranes, fluctuating membranes, crowded solutions, and protein motion near surfaces (Sakamoto et al., 2023, Jasuja et al., 28 Jun 2025, Reinhardt et al., 2020, Reister-Gottfried et al., 2010, Girelli et al., 2024). A later strand uses MP2D as the title of a method for “Multi-Objective Protein Sequence Design,” where discrete denoising trajectories are optimized by constrained search in sequence space (Kong et al., 7 May 2026). A closely related multimodal diffusion-language-model framework, CFP-Gen, is described as conceptually equivalent to multi-property diffusion but does not use the acronym explicitly (Yin et al., 28 May 2025).

Within the membrane setting, the canonical heterogeneity is the coexistence of ${\rm L_o}$ and ${\rm L_d}$ domains. The phase-field variable $c(\mathbf{r},t)$ denotes deviation from the critical composition, with $c<0$ corresponding to ${\rm L_o}$ and $c>0$ to ${\rm L_d}$. A normalized order parameter $0<\overline{c}<1$ is used for assigning local diffusivities and identifying domains, with the practical threshold ${\rm L_d}$0 for ${\rm L_d}$1 and ${\rm L_d}$2 for ${\rm L_d}$3 (Sakamoto et al., 2023). These domains are biophysically significant because they organize signaling and trafficking by concentrating specific proteins, modulating mobility, and altering encounter rates; ${\rm L_d}$4 domains often act as nanoscale reaction platforms, with typical sizes tens of nanometers and lifetimes ${\rm L_d}$5–${\rm L_d}$6 (Sakamoto et al., 2023).

2. Phase-separated membrane formulation

The best-defined physical MP2D construction is the coupled Langevin dynamics plus phase-field framework, denoted LDPF, introduced for proteins diffusing in phase-separated membranes (Sakamoto et al., 2023). For isolated proteins, the overdamped dynamics are

${\rm L_d}$7

with Gaussian white noise satisfying ${\rm L_d}$8. The diffusivity is composition-dependent,

${\rm L_d}$9

In the single-particle case, ${\rm L_o}$0 and ${\rm L_o}$1, giving ${\rm L_o}$2. In the multiparticle case, ${\rm L_o}$3 and ${\rm L_o}$4, giving ${\rm L_o}$5 (Sakamoto et al., 2023).

Crowding is introduced through interparticle Lennard–Jones interactions,

${\rm L_o}$6

with

${\rm L_o}$7

where ${\rm L_o}$8 and ${\rm L_o}$9 (Sakamoto et al., 2023). Domain preference is represented not by a bulk free-energy fit but by an interface reflection rule with probability ${\rm L_d}$0: for Lo preference (Lo${\rm L_d}$1), particles attempting to leave ${\rm L_d}$2 are reflected with probability ${\rm L_d}$3; for Ld preference (Ld${\rm L_d}$4), reflection is imposed when leaving ${\rm L_d}$5 (Sakamoto et al., 2023).

The composition field itself evolves by a Cahn–Hilliard-type phase-field equation. In the baseline heterogeneous membrane models, the evolution is

${\rm L_d}$6

with Ginzburg–Landau free energy

${\rm L_d}$7

Separate tests allow protein-induced field modification through a short-ranged coupling ${\rm L_d}$8 with parameters ${\rm L_d}$9 and $c(\mathbf{r},t)$0 (Sakamoto et al., 2023).

The simulations were performed on a $c(\mathbf{r},t)$1 grid with $c(\mathbf{r},t)$2, periodic boundaries, and system size $c(\mathbf{r},t)$3. The Langevin time step was $c(\mathbf{r},t)$4, mapped to $c(\mathbf{r},t)$5, and typical trajectories comprised $c(\mathbf{r},t)$6 steps, corresponding to $c(\mathbf{r},t)$7, with analysis after at least $c(\mathbf{r},t)$8 equilibration (Sakamoto et al., 2023). This construction makes the local mobility a direct function of membrane composition and thereby renders diffusivity itself a fluctuating dynamical variable.

3. Fluctuating diffusivity, subdiffusion, and partitioning

A defining MP2D result is that composition-dependent mobility in phase-separated membranes produces bimodal and time-dependent diffusivity. In the single-particle setting, $c(\mathbf{r},t)$9 corresponds to the slow state near $c<0$0 and $c<0$1 to the fast state near $c<0$2; bimodal $c<0$3 arises from bimodal PDFs of $c<0$4 in phase-separated fields (Sakamoto et al., 2023). The basic diagnostics are the mean squared displacement,

$c<0$5

with $c<0$6 and $c<0$7 for subdiffusion, and the time-dependent diffusivity,

$c<0$8

Fluctuations in diffusivity are quantified through the relative standard deviation of TAMSDs, which exhibits a plateau over $c<0$9–${\rm L_o}$0 and crosses over to ${\rm L_o}$1 decay at short and long times (Sakamoto et al., 2023).

The membrane work separates two mechanisms of anomalous transport. Crowd­ing-induced subdiffusion originates from collisional caging and dynamic clustering. As particle number increases, the occupancy of ${\rm L_o}$2 domain area rises from approximately ${\rm L_o}$3 to ${\rm L_o}$4 for ${\rm L_o}$5 to ${\rm L_o}$6, the TAMSD amplitude decreases, and the anomalous exponent drops from about ${\rm L_o}$7 to about ${\rm L_o}$8 up to about ${\rm L_o}$9. Stronger interactions, $c>0$0 versus $c>0$1, further suppress TAMSD and reduce $c>0$2 (Sakamoto et al., 2023). Confinement-induced subdiffusion arises from domain preference. Lo$c>0$3 reduces TAMSD and yields $c>0$4, while Ld$c>0$5 increases TAMSD and gives $c>0$6 (Sakamoto et al., 2023).

Partitioning is expressed thermodynamically through the equilibrium occupancy ratio

$c>0$7

Its reported determinants are the diffusivity contrast $c>0$8, the interface reflection probability $c>0$9, and molecular concentration ${\rm L_d}$0 together with interaction strength ${\rm L_d}$1 (Sakamoto et al., 2023). Without explicit preference, ${\rm L_d}$2, proteins still spontaneously enrich in ${\rm L_d}$3 because slower mobility increases residence times and aggregation; the Lo fraction rises sigmoidal in time and reaches a plateau by about ${\rm L_d}$4. Higher ${\rm L_d}$5 accelerates enrichment and increases the plateau fraction, while larger ${\rm L_d}$6 increases the plateau fraction but has little effect on the rate. With explicit preference, Lo${\rm L_d}$7 increases both rate and plateau fraction in ${\rm L_d}$8, whereas Ld${\rm L_d}$9 reduces the Lo fraction, with a crossover around $0<\overline{c}<1$0 in Model 5 where Ld affinity counterbalances retention in Lo from slow mobility and clustering (Sakamoto et al., 2023).

A related misconception is that subdiffusion in such systems can be attributed to a single universal cause. The reported results instead separate crowding and confinement as distinct mechanisms, each extending the anomalous regime, and show that RSD is only weakly affected by crowding but decreases as $0<\overline{c}<1$1 increases because stronger confinement reduces mobility fluctuations (Sakamoto et al., 2023).

4. Non-equilibrium hybrid continuum–discrete formulations

A more general MP2D framework treats protein transport in membranes with spatially varying concentration and temperature under non-equilibrium statistical mechanics (Jasuja et al., 28 Jun 2025). The system state is

$0<\overline{c}<1$2

with total energy $0<\overline{c}<1$3 and entropy $0<\overline{c}<1$4 defined in the GENERIC formalism. Dissipative operators $0<\overline{c}<1$5 encode irreversible processes, and fluctuations obey $0<\overline{c}<1$6; the stochastic forcing is multiplicative and the Stratonovich interpretation is used in time integration (Jasuja et al., 28 Jun 2025).

The protein position follows an overdamped Langevin equation with an explicit divergence-drift term,

$0<\overline{c}<1$7

where the force derives from

$0<\overline{c}<1$8

The corresponding Smoluchowski equation has spatially varying coefficients and an explicit spurious-drift contribution inherited from multiplicative noise (Jasuja et al., 28 Jun 2025). Local diffusion is determined by the Einstein relation

$0<\overline{c}<1$9

In an isotropic approximation, this reduces to ${\rm L_d}$00 (Jasuja et al., 28 Jun 2025).

The continuum fields are evolved self-consistently. The concentration field obeys

${\rm L_d}$01

and the membrane temperature obeys a heat equation with conduction, interfacial exchange, and dissipation terms. Additional ODEs govern the protein and interfacial temperatures ${\rm L_d}$02 and ${\rm L_d}$03 (Jasuja et al., 28 Jun 2025). The hybrid solver uses a two-stage Euler–Heun scheme consistent with Stratonovich calculus, together with a finite volume discretization whose discrete gradient and divergence operators satisfy ${\rm L_d}$04 (Jasuja et al., 28 Jun 2025).

Within this non-equilibrium setting, diffusiophoresis is represented by the chemical-force term ${\rm L_d}$05, while thermophoretic drift is not parameterized by an explicit Soret coefficient in the core formulation but is captured implicitly through state-dependent mobility and the divergence drift (Jasuja et al., 28 Jun 2025). The reported applications include protein positioning, thermal gradient sensing, and hot Brownian motion in energy wells. In the “Protein Positioning” studies, the relative magnitudes of ${\rm L_d}$06 and ${\rm L_d}$07 determine whether the protein moves toward the initial concentration peak or whether the concentration field redistributes toward the protein. In “Hot Brownian Motion,” increased ${\rm L_d}$08 raises ${\rm L_d}$09 and sharply decreases first-passage escape times from energy wells as heating amplitude increases (Jasuja et al., 28 Jun 2025).

5. Hydrodynamic, surface, and curvature-mediated extensions

Beyond heterogeneous composition and non-equilibrium fields, MP2D-type transport models incorporate hydrodynamic interactions, long-range electrostatics, and membrane-shape coupling. For rigid-body proteins near surfaces, Brownian Dynamics in the SDA package propagates solutes by an Ermak–McCammon step with configuration-dependent diffusion and force ${\rm L_d}$10, while practical many-molecule simulations replace explicit mobility tensors with scalar short-time diffusion coefficients reduced by a mean-field hydrodynamic-interaction model based on the local occupied volume fraction ${\rm L_d}$11 (Reinhardt et al., 2020). Near a no-slip plane, anisotropic wall corrections are introduced through the Lorentz–Faxén reduction factor ${\rm L_d}$12 and the Brenner–Goldman factor ${\rm L_d}$13, so that ${\rm L_d}$14 and ${\rm L_d}$15 (Reinhardt et al., 2020).

The same surface framework combines Poisson–Boltzmann-derived interaction grids with Debye–Hückel tails, including a smooth grid–DH transition in the “partial” region for electrostatic force continuity (Reinhardt et al., 2020). Applied to hen egg-white lysozyme adsorption, it recovered several experimental observables. For HEWL on mica at ${\rm L_d}$16, the total surface coverage after ${\rm L_d}$17 was approximately ${\rm L_d}$18 and the first layer approximately ${\rm L_d}$19, consistent with reported saturated first-layer coverage. Although wall hydrodynamics strongly reduced local short-time diffusion near the surface, global adsorption curves were only minimally affected at both ${\rm L_d}$20 and ${\rm L_d}$21 (Reinhardt et al., 2020).

A distinct extension concerns diffusing proteins on fluctuating membranes. There, a single inclusion of radius ${\rm L_d}$22 couples to membrane shape ${\rm L_d}$23 through spontaneous curvature ${\rm L_d}$24 and bending rigidity contrast ${\rm L_d}$25 in a Helfrich-type Hamiltonian (Reister-Gottfried et al., 2010). The coupled overdamped equations for the protein and membrane modes yield a rigorously derived reduction of the effective lateral diffusion coefficient, with ${\rm L_d}$26 in equilibrium (Reister-Gottfried et al., 2010). The mechanism is not a simple static barrier; simulations identify the dominant reduction as arising from correlations between the stochastic force on the protein and the delayed response of membrane shape, which effectively pulls the protein back toward the instantaneous energy-minimizing deformation (Reister-Gottfried et al., 2010).

The same theory predicts two time scales in membrane height correlations:

${\rm L_d}$27

with the latter typically much longer than the former (Reister-Gottfried et al., 2010). The long-time decay of height correlations therefore provides an indirect route to determine ${\rm L_d}$28. An important qualification reported in that work is that stiffness contrast alone, ${\rm L_d}$29, does not change ${\rm L_d}$30 unless there is nonzero spontaneous curvature ${\rm L_d}$31; curvature coupling is the crucial ingredient for mobility reduction in that model (Reister-Gottfried et al., 2010).

6. Experimental signatures in crowded protein solutions

An experimental realization of several MP2D themes is provided by MHz-XPCS measurements of ferritin diffusion in crowded solution (Girelli et al., 2024). The observables are the two-time correlation

${\rm L_d}$32

the intensity autocorrelation

${\rm L_d}$33

and the Siegert relation

${\rm L_d}$34

At low and moderate concentration, ${\rm L_d}$35 is described by a stretched-exponential KWW form; at the highest concentration, ${\rm L_d}$36, clear two-step relaxation requires a double-exponential intermediate scattering function (Girelli et al., 2024).

The static structure factor ${\rm L_d}$37 exhibits a correlation peak whose position shifts with concentration: ${\rm L_d}$38 and ${\rm L_d}$39 for ${\rm L_d}$40 and ${\rm L_d}$41, corresponding to correlation lengths ${\rm L_d}$42 and ${\rm L_d}$43 (Girelli et al., 2024). The dynamic counterpart is De Gennes narrowing: ${\rm L_d}$44 has a minimum near ${\rm L_d}$45, while the hydrodynamic function ${\rm L_d}$46 has a peak near the same wave number (Girelli et al., 2024). This is a direct structure–dynamics coupling at the inter-protein length scale.

At high concentration, anomalous dynamics are manifested not through directly extracted MSD exponents but through non-exponential ${\rm L_d}$47, KWW exponents ${\rm L_d}$48, and a resolved fast/slow mode decomposition (Girelli et al., 2024). The double-exponential analysis yields ${\rm L_d}$49, with the two branches sharing the same ${\rm L_d}$50-dependence. The cage amplitude follows

${\rm L_d}$51

with ${\rm L_d}$52 and ${\rm L_d}$53, indicating that approximately ${\rm L_d}$54 of proteins participate in cages with sub-nanometric to nanometric rattling (Girelli et al., 2024). Using ${\rm L_d}$55 at ${\rm L_d}$56 gives ${\rm L_d}$57 and an interaction time

${\rm L_d}$58

which coincides with the observed crossover between fast in-cage and slow escape dynamics (Girelli et al., 2024).

Hydrodynamic modeling with the ${\rm L_d}$59-theory reproduces the ${\rm L_d}$60-shape of ${\rm L_d}$61, but quantitative agreement over the experimental microsecond window requires a global scaling ${\rm L_d}$62 associated with long-time reduction by direct interactions (Girelli et al., 2024). At ${\rm L_d}$63, the measured ratio ${\rm L_d}$64 is much smaller than the short-time prediction ${\rm L_d}$65, implying an effective long-time reduction factor of about ${\rm L_d}$66 in that regime (Girelli et al., 2024). The study therefore gives experimental support to an MP2D picture in which crowding, hydrodynamics, and direct interactions jointly determine transport.

7. Sequence-space usage, limitations, and broader interpretation

A separate use of the acronym appears in “MP2D: Constrained Monte Carlo Tree-Guided Diffusion for Multi-Objective Protein Sequence Design” (Kong et al., 7 May 2026). Here the object of diffusion is a protein sequence ${\rm L_d}$67 over an amino-acid vocabulary, and the framework combines conditional discrete diffusion, constrained MCTS, dynamic Pareto constraints, and global iterative refinement for multi-objective optimization (Kong et al., 7 May 2026). The reverse denoising model uses classifier-free guidance, denoising is cast as a constrained sequential decision process, and candidate branches are filtered by angular alignment with pre-specified optimization directions. The method was evaluated on antimicrobial peptide and protein binder optimization tasks involving four to five conflicting properties and was reported to outperform existing multi-objective baselines without retraining the generator (Kong et al., 7 May 2026).

CFP-Gen occupies a nearby conceptual position while explicitly stating that it does not use the phrase MP2D. It composes functional annotations, sequence-level controls, and structural constraints in a single diffusion LLM through Annotation-Guided Feature Modulation, Residue-Controlled Functional Encoding, and a GVP-Transformer structural adapter (Yin et al., 28 May 2025). In that literature, “multi-property diffusion” therefore denotes multimodal conditioning during sequence denoising rather than transport in physical space (Yin et al., 28 May 2025). A plausible implication is that the acronym now functions more as a structural descriptor—multiple properties coupled through a diffusion process—than as the name of one standardized model family.

Across the physical transport papers, the principal limitations are consistent. The membrane models are coarse-grained, often approximately flat, and omit cytoskeletal remodeling, motor activity, active flows, or detailed multicomponent lipid chemistry (Sakamoto et al., 2023, Jasuja et al., 28 Jun 2025). The surface Brownian Dynamics framework assumes rigid-body solutes, continuum electrostatics, implicit solvent, and mean-field hydrodynamics rather than explicit many-body mobility tensors (Reinhardt et al., 2020). The fluctuating-membrane theory is developed for a single inclusion in the small ${\rm L_d}$68 regime and treats the core analytical case as tensionless (Reister-Gottfried et al., 2010). The ferritin XPCS study does not report pH, MSD-based subdiffusive exponents, or non-Gaussian parameters directly, and its interpretation of long-time slowdown relies on model-based decomposition of short- and long-time effects (Girelli et al., 2024). In sequence design, the main constraints are predictor noise, direction-vector and threshold selection, incomplete coverage of annotation vocabularies, and one-way structure conditioning rather than full sequence–structure co-design (Kong et al., 7 May 2026, Yin et al., 28 May 2025).

The generalizability claims are correspondingly broad but specific. The LDPF methodology is stated to apply to phase-separating polymer blends, porous or granular media, and cytoplasmic condensates when a scalar field controls local mobility (Sakamoto et al., 2023). The non-equilibrium hybrid framework is proposed for related biological systems and soft materials (Jasuja et al., 28 Jun 2025). The surface Brownian Dynamics framework extends to arbitrary proteins and surfaces given structural, electrostatic, hydrodynamic, and grid-based interaction inputs (Reinhardt et al., 2020). Taken together, these works support a unifying interpretation of MP2D as an approach in which protein behavior is governed by multiple coupled property fields, with the specific meaning of “diffusion” determined by whether the state space is physical space, membrane configuration space, or protein sequence space.