The SIMDA framework maps PIP2-binding sites regulating the KCNQ1 channel

Abstract

Voltage-gated potassium channel KCNQ1 (Kv7.1) underpins cardiac repolarization, epithelial ion transport, and inner ear function. Its versatility arises from interactions with KCNE proteins, calmodulin (CaM), and the lipid phosphatidylinositol 4,5-bisphosphate (PIP2), yet the molecular basis of PIP2 regulation remains incompletely understood. Here, we present the Stepwise Integrated Multi-scale Dynamics and Advanced Analysis (SIMDA) framework, which integrates coarse-grained and all-atom molecular dynamics, well-tempered metadynamics, and clustering and energy analyses. Over 2,000 µs of simulations across eight functional states find six recurrent PIP2 sites (C0-C5), each complex populating three to four. Sites C1 and C3 agree with experimental densities, whereas C0, C4, and C5 are forward predictions. KCNE3 stabilizes the primary C1 site by amplifying a C-terminal twist motion, while CaM tunes binding at C4 and C5. Interconnected transfer pathways form a dynamic circular route among sites. SIMDA thus provides a generalizable strategy for dissecting lipid-protein dynamics.

PIP2 lipid regulates the cardiac potassium channel KCNQ1. Here, the authors use multiscale simulations to map six PIP2 binding sites and show how KCNE3 and calmodulin redistribute lipid binding to tune channel activity.

Introduction

Voltage-gated potassium channel KCNQ1 (Kv7.1) stands out for its functional versatility, playing pivotal roles in cardiac repolarization, epithelial ion transport, and inner ear function^1–4. This versatility arises from KCNQ1’s ability to interact with auxiliary KCNE proteins (e.g., KCNE1, KCNE2, KCNE3) and calmodulin (CaM), enabling tissue-specific gating properties, trafficking behavior, and activity^5–8. KCNE3, a small single-pass transmembrane protein, eliminates the channel’s voltage dependence entirely when it binds to KCNQ1, rendering the resulting complex constitutively active at resting membrane potentials, a property essential for potassium recycling in colonic epithelial cells^9–12. CaM operates differently: this ubiquitous calcium-binding protein docks onto KCNQ1’s intracellular C-terminus and fine-tunes channel gating, trafficking, and assembly^13–15. Each regulatory partnership serves distinct physiological needs, and together with PIP2, these interactors form an integrated regulatory network rather than acting independently.

Beyond these protein-protein interactions, KCNQ1 depends on a lipid cofactor: phosphatidylinositol 4,5-bisphosphate (PIP2)^16–18. This signaling lipid does more than simply bind the channel. It acts as a molecular bridge, coupling the voltage-sensing domains (VSDs) to the central pore domain (PD), thereby enabling channel opening (Fig. 1a)^19–23. PIP2’s negatively charged headgroup seeks out positively charged residues in KCNQ1’s C-terminal domain. Once bound, it stabilizes the open conformation and can even promote constitutive activity at resting voltages^9,24–26. The binding site sits close to where CaM attaches, raising the possibility of functional crosstalk or even competition between the two regulators^9,25.

Fig. 1. The Systems investigated and SIMDA pipeline.

a Superimposed conformations of 6V00 and 6V01 monomers. b Intact tetrameric conformation of 6V00. c Schematic of the SIMDA framework for studying protein-lipid binding. d, e The framework analyzes lipid dynamics, thermodynamics, and binding sites through a stepwise approach: Stepwise Integrated Multi-scale Dynamics begins with structure preparation. CG-MD simulations identify lipid binding sites (e.g., Bsid0, Bsid1, Bsidn), which are refined via AA-MD. Wt-MetaD enables enhanced sampling of dissociation pathways. Advanced Analysis includes binding site screening based on interaction stability, UMAP-Leiden clustering to identify functionally equivalent sites, interaction analysis using hydrogen bond analysis and MM/GBSA, and dissociation trajectory analysis for thermodynamic and kinetic insights.

The interaction becomes more complex when auxiliary proteins are involved. KCNE3 reshapes the VSD conformation in ways that favor PIP2 binding, strengthening constitutive activity^9,27–29. KCNE1 makes KCNQ1 about 100-fold more sensitive to PIP2 than the channel alone, dramatically amplifying ionic current^17,30,31. These observations suggest that PIP2, KCNE proteins, and CaM form an integrated regulatory network rather than acting independently.

Recent cryo-EM structures have begun to visualize this network^9,25. One structure of the KCNQ1-KCNE3 complex captured PIP2 density near the S4-S5 linker (Fig. 1a)⁹. Complementary biochemical and computational work has implicated additional regions in PIP2 binding: the S0 segment, the S2-S3 linker, various helices in the C-terminal domain (HA and HB), and the junction between the pore-lining S6 helix and HA^13,32–37. Yet these studies leave important questions unanswered. How many functional PIP2 binding sites exist on KCNQ1, and where exactly are they? Do KCNE3 and CaM simply shift which site PIP2 prefers, or do they fundamentally alter binding stability and dynamics? Once bound, can PIP2 transfer between multiple pockets? What are the specific residues involved, the conformational changes triggered, and the energetic barriers that must be overcome?

Addressing these questions requires methods that can scan broadly for binding sites, then zoom in to atomic detail, map transition routes, and handle the complexity of a tetrameric channel with multiple equivalent subunits. No single approach can fully achieve all of this. Coarse-grained MD (CG-MD) enables efficient sampling of the lipid-binding landscape but lacks the atomic detail needed to define residue-level interactions^38–41. All-atom MD (AA-MD) provides detailed interaction and conformational information, yet is limited by timescales that restrict observation of rare binding and unbinding events. Well-tempered metadynamics (Wt-metaD) can reconstruct free energy landscapes, but it depends on predefined collective variables (CVs) and prior knowledge of binding sites. Furthermore, the symmetry of KCNQ1 complicates cross-system comparison of equivalent sites, and no existing CG-MD or AA-MD workflow possesses the integrated capability to unify binding sites across systems with different auxiliary proteins and quantitatively characterize binding pathways within a single framework.

Here, we show that a Stepwise Integrated Multi-scale Dynamics and Advanced Analysis framework (SIMDA) addresses these gaps by integrating CG-MD for site sampling, AA-MD for structural refinement, and Wt-metaD for pathway exploration, together with UMAP-Leiden clustering to identify equivalent binding sites across multimeric systems and interaction, hydrogen-bond, and free-energy analyses for quantitative insight (Fig. 1b). Applying SIMDA to over 2000 µs of simulations across eight KCNQ1 functional states, we map the PIP2 binding landscape and identify six recurrent site types (C0–C5), each complex populating three to four sites according to auxiliary subunit composition and activation state. Sites C1 and C3 agree with cryo-EM densities, whereas C0, C4, and C5 correspond to predicted binding sites at the S2-S3 linker. KCNE3 stabilizes C1 through amplified C-terminal twist motions and direct hydrogen bonds; CaM exerts state-dependent regulation at C4 and C5, and Wt-metaD simulation identifies eight interconnected transfer pathways. These results provide molecular insight into tissue-specific KCNQ1 regulation and establish a generalizable framework for lipid-mediated modulation of membrane proteins.

Results

Six distinct PIP2 binding sites emerge from multi-scale simulations

CG-MD reveals how auxiliary proteins reshape the PIP2 landscape

We performed CG-MD simulations to comprehensively explore PIP2 binding to KCNQ1 across its functional landscape. All simulations were initiated from experimentally resolved cryo-EM structures in inactive (PDB: 6V00) and active (PDB: 6V01) states. We constructed eight systems by selectively including or removing KCNE3 and CaM: KCNQ1 alone, KCNQ1-KCNE3, KCNQ1-CaM, and KCNQ1-KCNE3-CaM, in both activation states (Table S1). Each channel was embedded in a membrane bilayer composed of 95% POPC and 5% PIP2. We chose 5% PIP2 because previous work showed it captures the major binding sites without overwhelming the system^42–44. Cross-system comparison of PIP2 binding site locations requires that KCNE3 and CaM do not introduce large-scale rearrangements of the KCNQ1 transmembrane domain (TMD). We verified this through systematic structural superimpositions: superimposition of all atoms of KCNE3-free KCNQ1 structures onto their respective cryo-EM references yielded RMSD values of 2.36 Å and 1.94 Å across the TMD (Fig. S1a, b), confirming that KCNE3 does not substantially alter the overall TMD conformation^9,45. For CaM, given the absence of CaM-free KCNQ1 cryo-EM structures, alignment of paired CaM-bound and CaM-free KCNQ2 all-atom structures yielded an RMSD of 0.84–1.21 Å, with structural deviations confined to the cytosolic HA/HB helix region (Fig. S1c, d)⁴⁶, consistent with CaM acting primarily through remodeling of the HA/HB helices while leaving the transmembrane core intact⁴⁷. Every system was run for 30 μs, and we performed eight independent replicates with different starting velocities to ensure robust sampling. In total, this amounted to 1920 μs of coarse-grained simulation time (see Methods).

When we mapped where PIP2 molecules spent their time in the cytoplasmic leaflet, clear hotspots emerged (Fig. 2a, b). For KCNQ1 alone, PIP2 clustered near the S2-S3 linker, with secondary accumulation around the S0-S1 region. This pattern shifted substantially when KCNE3 was present: density at S2-S3 dropped by ~40% while accumulation near the S4-S5 linker and S5 helix increased approximately threefold. CaM binding introduced yet another pattern, with PIP2 appearing prominently at the S0-S1 loop and deeper regions adjacent to the HA helix, showing ~2.5-fold enrichment compared to CaM-absent systems. In the ternary KCNQ1-KCNE3-CaM complex, two main hotspots dominated: the S4-S5/S5 interface held ~60% of all inner-leaflet PIP2, while the S0-HA region accounted for another 25%. The proteins equilibrated within 1 μs; major binding sites appeared by 5 μs and stabilized by 20 μs (Fig. S2), with density maps changing very little (Figs. S3, S4). This convergence confirmed that the observed hotspots represent genuine equilibrium distributions rather than transient fluctuations. Thus, these results indicate that auxiliary proteins reshape PIP2 binding preferences through site-specific effects rather than global changes. KCNE3 promotes PIP2 occupancy at the S4 -S5 region while reducing association at the S2-S3 linker, whereas CaM shifts binding toward the S0 HA interface, with the combined condition yielding a distinct distribution.

Fig. 2. Identification and Screening of PIP2 binding sites based on CG -MD and AA-MD simulations.

a 2D-density maps in the plane of the bilayer (XY) of the distribution of PIP2 relative to the four inactive state systems: KCNQ1, KCNQ1-KCNE3, KCNQ1-CaM, and KCNQ1-KCNE3-CaM. b Density maps of PIP2 binding across the Active state of KCNQ1, KCNQ1-KCNE3, KCNQ1-CaM, and KCNQ1-KCNE3-CaM. c Binding sites identified and screened from the AA-MD simulations in the Active KCNQ1-KCNE3-CaM system were subjected to the second screening selection. Moreover, PIP2 binding at the excluded site Bsid10 was found to be unstable within the first 100 ns, as evidenced by the 300 ns AA-MD simulations (3000 frames in total).

High-confidence binding events identified through rigorous kinetic filtering

The hotspot regions represent macroscopic zones of preferential PIP2 enrichment (Fig. 3). To identify specific discrete binding events, we turned to PyLipID⁴⁸ and LipiDens⁴⁷ and applied three filters. First, a binding event had to last at least 0.5 μs, meaning the dissociation rate constant (k_off) could not exceed 2 μs⁻¹ (Fig. S5). This ensures we are looking at stable interactions, not fleeting encounters. Second, we calculated Δk_off, which measures the difference between k_off estimated by bootstrap resampling versus direct curve fitting (Fig. S6). Large Δk_off values signal unreliable kinetics, often because the lipid never truly binds or the fitting failed⁴⁷. We excluded these cases. Third, we required high occupancy, >50% of the time the lipid had to stay within both a 0.5 nm cutoff and a 0.9 nm cutoff, and a large contact surface area, defined as >10 nm² (Figs. S5 and S6).

Fig. 3. 6 PIP2 binding sites on KCNQ1.

a The classifications of binding modes of PIP2 at the 6 binding sites. b Binding modes of PIP2 at the 6 binding sites. c Superposition of experimental structures and MD simulation results for four KCNQ channel complexes: KCNQ1 (PDB: 6V01, C1 site), KCNQ5 (PDB: 9LIZ, C2 site), KCNQ4 (PDB: 7VNP, C3 site), and KCNQ1 (PDB: 9WD8, C1 and C3 sites). Experimental protein structures are shown in white cartoon representation. PIP2 molecules resolved in the cryo-EM structures are displayed as cyan spheres. Simulation-derived channel subunits are highlighted in green (C1), purple (C2), and light blue (C3) cartoons, with corresponding simulation-derived PIP2 molecules shown in matching colors. d Different classifications and numbers of PIP2 binding sites in the Inactive and Active KCNQ1 systems.

Applying these filters across 64 trajectories (8 systems × 8 replicates) identified 72 preliminary binding sites (Fig. S5). Most sites showed stable contact areas in the later simulation stages, indicating quasi-equilibrium lipid-protein interactions (Figs. S7–S22). Reproducibility was supported by consistent binding patterns across replicates. For example, 6 of 8 in the active KCNQ1-KCNE3 system, with some expected variability due to lipid diffusion (Fig. S8). We assessed reproducibility by computing cosine similarity between every pair of trajectories: about 15% of pairs showed high similarity (above 0.8), 60% showed medium similarity (0.6–0.8), and 25% fell below 0.6, giving an average similarity of ~0.67 (Figs. S23 and S24). In other words, 75% of trajectory pairs agreed reasonably well. Each system consistently identified at least 10 sites, with most appearing in 6–8 replicates (Figs. S25 and S26). Sites that appeared in only 3–4 replicates often had questionable occupancy or residence times and were excluded.

To characterize occupancy variability, we classified sites as stable (coefficient of variation, CV < 20%), preferential (CV 20–60%), or transient (CV > 60%) based on their bimodal occupancy distributions (Fig. S27, Table S2): in some trajectories PIP2 persistently occupies the site (>70%), while in others the same site is rarely visited (<30%). This distributional heterogeneity reflects the intrinsic stochasticity of lipid-protein interactions at finite simulation timescales and is not indicative of insufficient sampling⁴⁹. Additional simulations confirmed that PIP2 binding patterns are robust to membrane composition, with similar distributions observed in a POPC: POPG: PIP2 at 18:1:1 mixture (Fig. S28). Varying PIP2 concentration from 2.5% to 10% showed that 5% provides sufficient sampling without saturation, as lower concentrations yielded fewer binding events while higher concentrations produced similar distributions (Fig. S29).

AA-MD refinement and clustering define six functional sites (C0 to C5)

CG-MD simulations are efficient, but they lack atomic detail. To verify that the identified sites are stable at full resolution, we converted representative CG structures of each preliminary site into all-atom models and ran 300 ns of AA-MD with the CHARMM36m force field^50,51. We considered a site stable if two criteria held over the final 200 ns: the protein backbone RMSD stayed below 5 Å, the PIP2 RMSD stayed below 3.5 Å, and neither fluctuated more than 2 Å in the last 50 ns (Figs. S30 and S31). Applying these filters, we retained 69 stable PIP2 binding sites across all systems.

To group equivalent sites across the tetrameric channel and eight systems, where the same functional site may appear at different spatial coordinates across subunits or systems. We used UMAP to reduce the dimensionality of contact fingerprints^52,53, then applied Leiden clustering to partition the sites into communities⁵⁴. This unsupervised approach identified six clusters, named C0-C5 (Fig. 3a). These six sites constitute a comprehensive atlas assembled by pooling binding poses across all eight simulation systems; they do not represent six simultaneously occupied sites in any single complex. In any individual system, only three to four sites are observed depending on auxiliary subunit composition and activation state (Fig. 3d). Bootstrap resampling yielded adjusted Rand index and normalized mutual information both equal to 1.00 with zero variance, confirming partition stability. Cross-system cosine similarity ranged from 0.758 (C4, weakest) to 0.865 (C1, strongest), with a mean within-cluster similarity of 0.802. Statistical tests confirmed that the six clusters are well separated from one another (Fig. S32d).

Spatially, the six sites cluster into three groups (Fig. 3b). C0 and C1 are located on one side of the C-terminal domain near the S2-S3 and S3-S4 linkers, with C0 at the S2-S3 linker and S3 segment, and C1 extending deeper to include the S4-S5 linker and KCNE3 in KCNE3-containing systems. C2 forms a distinct intermediate site at the S4-S5 linker interface. On the opposite side, C3, C4, and C5 occupy the S1 and S0-S1 regions, with C3 positioned deep near S1 and the S4-S5 linker; C4 sits at the S0-S1 loop and extends toward the HA and HB helices; it can associate with CaM. C5 is predominantly observed in CaM-bound active systems at the S0-S1 and HA interface. Per-system independent UMAP-Leiden clustering yielded three to four clusters per system (Fig. S33, Fig. 3d), fully consistent with the global six-site framework. For example, inactive-state KCNQ1 alone yields three clusters (C0, C1, C4), while inactive-state KCNQ1-KCNE3 yields four (C1, C2, C3, C4). These per-system results are fully consistent with the global six-site framework and confirm that the collective atlas of C0-C5 encompasses all distinct binding modes observed across all conditions.

Three sites match experimental structures, and three are forward predictions

We next assessed agreement between our predictions and experimental data. To validate our computational predictions, we calculated Q-scores, which measure how well an atomic model fits cryo-EM density on a scale from 0 to 1, with higher values indicating better agreement. Q-scores >0.5 generally indicate good fit, 0.4–0.5 reasonable fit, and <0.4 poor fit, though interpretation depends on local map quality⁵⁵.

Site C1 achieved a Q-score of 0.66 when aligned to the KCNQ1 structure with bound PIP2 near the S4-S5 linker (PDB: 6V00) (Fig. 3c). Two cryo-EM structures of KCNQ1 in complex with KCNE auxiliary subunits and PIP2, reported by Cui et al.⁵⁶ and Zhong et al.⁵⁷, allow direct assessment of our computational predictions against independent experimental data. We superimposed C1/C3 sites against KCNQ1-KCNE1-PIP2 (PDB: 9VEI), KCNQ1-KCNE3-PIP2 (PDB: 9WD8), and KCNQ1-KCNE1-PIP2 (PDB: 9UC8), and computed Q scores for each comparison (Fig. S34). Sites C1 and C3 showed good structural agreement with PIP2 densities in all three structures, with the strongest correspondence against 9WD8: Q values of 0.62 for C1 and 0.80 for C3, both meeting the accepted criterion for good model-to-map fit, indicating that our simulations accurately captured the location and atomic environment of these two core binding sites (Fig. 3c). We also compared our sites to structures of related KCNQ family members. Site C2 gave a Q-score of 0.39 against KCNQ5 (PDB: 9LIZ ref. ⁵⁸), and site C3 scored 0.57 against KCNQ4 (PDB: 7VNP ref. ²⁴) (Fig. 3c). These scores are lower than for C1. Notably, the Q-score of 0.39 for C2 falls below the accepted threshold for reliable model-to-map fit (Q scores > 0.4). Moreover, C2 appears in only 2 of 8 simulation systems (KCNQ1-KCNE3 inactive and active states), fully consistent with its characterization as a transient, KCNE3-dependent intermediate. We calculated Q scores for non-binding regions (regions >1.5 nm from any identified binding site) using the same PIP2 placement, which yielded Q scores of 0.15–0.22, confirming that our identified sites show substantially better fit than random placements.

Sites C0, C4, and C5 have no corresponding densities in currently available cryo-EM structures, particularly for the KCNQ1-KCNE3-CaM complex, and should be treated as conditional computational predictions. C0 appears exclusively in KCNQ1 systems lacking both KCNE3 and CaM, suggesting it represents a binding configuration distinct from the unpartnered channel. For C0, charge-adding substitutions in the S2-S3 linker (V185R/W188R, which render KCNQ1 more KCNQ2-like) and other linker mutations (R181E; R190A in KCNQ3) affect PIP2-dependent gating, lending regional plausibility to this site^47,59–61. We note that some of these effects, such as that of R181E, may arise through electrostatic or CaM-VSD-coupling mechanisms rather than direct PIP2 contact, so these residues provide regional rather than residue-specific support. Consistent with C0 being a shallow, distributed site without a single dominant anchor, PIP2 coordination here is shared among several basic residues; the positions our simulations highlight (R192, R195) are presented as predictions for direct experimental test⁶². For C4, the S0 helix has been identified as a phosphoinositide-binding locus by NMR and mutational scanning, and R109 forms persistent hydrogen bonds with membrane lipid headgroups^63,64; however, residue-specific roles of R109 and K121 in the context of PIP2 sensitivity have not been directly tested⁶⁵. For C5, prior studies demonstrate competitive interactions between PIP2 and CaM at the S6-HA junction and CaM-dependent VSD coupling via the S2-S3 linker, consistent with the existence of a CaM-active-state-exclusive PIP2 site at this binding site, but direct structural evidence is not yet available^47,66. Taken together, these results support the functional relevance of the predicted regions.

KCNE3 and CaM modulate PIP2 binding through distinct mechanisms

State-dependent redistribution of binding site occupancy

Having identified six-site types and their system-specific occurrence patterns, we next examined how PIP2 occupancy varies across the eight simulated complexes. This analysis revealed distinct distribution patterns governed by auxiliary protein composition and channel state. The results demonstrate that KCNE3 and CaM exert non-redundant regulatory effects on PIP2 binding (Figs. 3d, 4a).

Fig. 4. Contact, hydrogen bond, and MM/GBSA energy decomposition analyses of hotspot residues based on AA-MD simulations.

a Distribution of PIP2 binding sites in the Inactive and Active KCNQ1 systems, respectively. The numbers in red represent different binding sites (C0-5). b Contact, H-bond, and MM/GBSA energy decomposition occupancy between hotspot residues and PIP2 for each binding site cluster, ranging from 0% to 40%. Orange represents representative contact frequency, cyan represents representative H-bond frequency, and green represents energy decomposition frequency.

In KCNQ1-only systems, three sites showed consistent occupancy. Sites C0, C1, and C4 were present in both inactive and active states (Fig. 3d). Among these, C1 showed the highest occupancy, appearing in ~85% of CG-MD trajectories. This high occupancy suggests that C1 serves as a primary PIP2 binding site even in the absence of auxiliary proteins (Fig. 3d). Notably, C0 appeared exclusively in KCNQ1-only conditions, never appearing in systems containing KCNE3 or CaM.

KCNE3 binding induced a substantial redistribution of PIP2-binding sites. The C0 site was completely eliminated in all KCNE3-containing systems. Simultaneously, two additional sites emerged. C2 and C3 appeared on the opposite side of the C-terminal domain from where C0 had been located (Fig. 3d). C1 remained highly occupied, appearing in ~90% of CG-MD trajectories, an increase from KCNQ1-only systems. Site C4 persisted at moderate occupancy levels similar to those in KCNQ1-only systems (Fig. 3d). This redistribution suggests that KCNE3 sterically blocks the C0 binding pocket while creating favorable electrostatic or structural environments for C2 and C3.

CaM binding produced a different redistribution pattern. C1 was retained with high occupancy in both inactive and active states (Fig. 3d). A state-dependent shift occurred at the C4/C5 locus. In the active state, occupancy shifted almost entirely from C4 to C5. In the inactive state, part of the binding remains at C4. Site C3 emerged as an additional binding location in both states.

In the ternary complex with both KCNE3 and CaM, the effects combined. In the active state, C5 was highly occupied (~70% of CG-MD trajectories), C2 appeared exclusively in KCNE3’s presence (confirming its dependence on KCNE3), and C1 remained the most consistent site (~95% occupancy of CG-MD trajectories) (Fig. 3d). This makes C1 the primary anchor across all conditions, regardless of auxiliary protein composition or activation state.

Quantitative analysis across all systems revealed clear dependencies. C3 requires either KCNE3 or CaM, as it was absent in KCNQ1-only systems but appeared when either auxiliary protein was present. C5 showed absolute CaM dependence, never appearing in CaM-absent systems (Fig. 3d). C0 exhibited the opposite pattern, appearing only in the absence of auxiliary proteins (Fig. 3d). These findings demonstrate that KCNE3 and CaM regulate PIP2 interactions through distinct but complementary mechanisms. They do not simply enhance or reduce binding globally. Rather, they reshape the binding landscape, redirecting PIP2 to different sites that likely serve distinct functional roles in channel regulation.

KCNE3 stabilizes C1 binding through enhanced C-terminal twist effect

The observation that KCNE3 increases C1 occupancy from 85% to 95% raised a critical question: how does KCNE3 enhance binding at this site? We performed Wt-metaD simulations (400 ns each) starting from C1-bound configurations in KCNE3-free, KCNE3-bound, and CaM-bound systems, using the X and Y components of the PIP2-protein distance within 10 Å as CVs. Convergence of free energy profiles was confirmed by analyzing FES stability and the consistency of key interaction patterns across independent simulations (Figs. S35 and S36).

PMF barriers for PIP2 dissociation from C1 were: 58.01 kJ/mol (C1-free), 77.72 kJ/mol (C1-CaM), and 110.27 kJ/mol (C1-KCNE3) (Fig. 5a). The progressively higher barriers observed with auxiliary proteins, particularly KCNE3, are consistent with experimentally documented enhancement of PIP2 sensitivity and suggest that both KCNE3 and CaM slow PIP2 dissociation through distinct stabilizing mechanisms. We emphasize that these values represent within-system dissociation barriers rather than absolute binding free energies^9,27–29.

Fig. 5. The dissociation pathways of PIP2 at the binding sites C1-Free, C1-KCNE3, and C1-CaM based on Wt-MetaD simulations.

a–c FES during the dissociation process of PIP2 at C1-Free, C1-KCNE3, and C1-CaM Systems; Distance fluctuations between the S2-S3 linker, S2, S3, and S4 regions near the C-terminal and their positions in the initial simulation frame for C1-Free; Principal Component Analysis (PCA) of Key Regions (S2-S3 Linker, S2, S3, S4, and S4-S5 Linker) in the C1-Free, C1-KCNE3, and C1-CaM Systems; Conformational changes near the C-terminal region, including the S2-S3 linker, S2, S3, S4, and S4-S5 linker, during transitions between key stable and metastable states for C1-Free, C1-KCNE3, and C1-CaM Systems.

To understand these differences, we analyzed conformational changes during dissociation. Each trajectory was divided into four states (A-D) based on FES minima, representing progression from fully bound to fully dissociated (Figs. 5b, S37). Notably, as PIP2 left the C1 site, the S2-S3 linker moved dramatically, 8.05 Å in C1-free, 8.41 Å in C1-KCNE3, and 7.22 Å in C1-CaM, compared to minimal movement (<2 Å) during equilibrium MD with bound PIP2 (Fig. S37a). This large displacement represents the C-terminal twist, a conformational rearrangement observed experimentally with PIP2 binding. C1-KCNE3 exhibited the most extensive changes, with the twist extending beyond the S2-S3 linker to encompass the entire S2 and S3 segments (Fig. 5). PCA on the S2-S3 linker and adjacent segments revealed that C1-free showed three discrete conformational basins, while C1-KCNE3 exhibited more continuous conformational space with smoother state transitions. Because this rearrangement is energetically costly, it raises the dissociation barrier: removing PIP2 from C1 with KCNE3 present requires more extensive C-terminal domain twisting.

Crucially, interaction fingerprints revealed an additional anchoring mechanism: strong hydrogen bonds between PIP2 and KCNE3 C-terminal residues R81 and R83 (Fig. S38) provide direct stabilization that compounds the conformational coupling effect. This dual mechanism, amplified C-terminal twist plus direct hydrogen bond anchoring, explains why KCNE3 has the higher dissociation barrier (Fig. 5, S39). For CaM, the higher barrier versus C1-free stems from tighter hydrogen bonding between PIP2 and S3/S2-S3 linker residues, which increases C-terminal fluctuations and creates a more rugged energy landscape (Fig. 5, S39). In summary, PIP2 dissociation from C1 triggers C-terminal domain twisting, especially in the S2-S3 linker; KCNE3 amplifies this twist and contributes direct hydrogen bonds, while CaM stabilizes C1 through enhanced hydrogen bonding. Both auxiliary proteins ensure longer PIP2 residence at C1, important for maintaining channel activity under physiological conditions and consistent with experimental observations of enhanced PIP2 sensitivity.

CaM enables deep PIP2 binding through S6-HA junction tightening

Having understood KCNE3’s effects at C1, we examined sites C4 and C5 near the CaM-binding region to understand CaM’s operation. Free energy calculations revealed a counterintuitive pattern (Fig. S39): C4-free showed PMF ~ 100 kJ/mol; C4-CaM-active was intermediate at 79.96 kJ/mol; C4-CaM-inactive dropped to 49.16 kJ/mol. CaM’s presence lowered rather than raised the dissociation barrier, which is opposite to KCNE3’s effect at C1. The answer is competition: in C4-CaM-active, CaM residues R83 and K84 form hydrogen bonds with PIP2 (Fig. 4b), pulling it toward CaM and away from the channel S1 segment, reducing channel-lipid contact stability. In C4-CaM-inactive, CaM attracts PIP2 but fails to form stable hydrogen bonds, producing maximum destabilization at ~49 kJ/mol. This competitive mechanism prevents stable C4 occupancy and redirects PIP2 to C5, where CaM’s active conformation better coordinates PIP2, explaining why C5 appears almost exclusively with active CaM. This competitive mechanism enables CaM to dynamically tune channel activity in response to calcium signals.

Site C3, the deepest binding site, showed dissociation barriers exceeding 95 kJ/mol, with PIP2 strongly tending to rebind other sites rather than fully departing (Figs. S40 and S41). Dissociation followed mandatory C3 → C1 (Path 5) or C3 → C4 (Path 6) transitions involving the HA/HB helices and auxiliary protein regions. Community network analysis during C3 occupancy and dissociation revealed “tightening of loose links”: when PIP2 occupies C3, VSD and PD organize into three tightly coupled communities with coordinated movement (Fig. S39a); as dissociation progresses, the network reorganizes and ultimately fragments into five communities with much weaker inter-community coupling (Fig. S39). This progressive fragmentation demonstrates that C3 binding physically tightens VSD-PD coupling, crucial for coordinated voltage-dependent gating, with KCNE3 facilitating C3 access and stabilization.

Site C0 functions as a shallow site where multiple pathways converge. Dissociation follows a straightforward route along Path 3 (Fig. 6). Strong hydrogen bonds between PIP2 and S2-S3 linker/S3 residues trigger significant linker fluctuations during binding and dissociation, manifesting as the C-terminal twist effect, though less pronounced than at C1(Fig. 4b).

Fig. 6. Overview of the dynamic binding and dissociation behavior of PIP2 across multiple sites, as characterized by Wt-MetaD simulations.

a Representative dynamic circular pathways of binding-site transfers during the PIP2 binding and dissociation process. b Transparent points denote distinct categories of binding sites. In each subplot (Path1-8), dissociation trajectories obtained from metadynamics simulations are illustrated, where spheres transition from blue to red, indicating the temporal progression of PIP2 positions during unbinding.

Integrating these observations, we identify five structural domains involved in PIP2 binding dynamics: the VSD at the C-terminal, including S0–S4 and the S4–S5 linker, the PD including S5 to S6, KCNE3, CaM, and the membrane-proximal HA-HB region (Fig. 4a). Among these, VSD and PD are most directly linked to electromechanical coupling. Binding site depth determines functional roles. Deep sites such as C3 enhance VSD-PD coupling by tightening loose links, with KCNE3 promoting access and stabilization. Intermediate sites C4 and C5 are modulated by CaM through competition, shifting occupancy toward shallower regions and limiting access to deeper sites. Shallow sites C0 and C1 mainly affect local C-terminal conformations through twist effects. Together, these depth-dependent interactions and auxiliary protein effects define a multi-level regulatory system controlling PIP2-binding across sites.

Residue-level interactions reveal the molecular basis of site selectivity

To understand the molecular basis of site selectivity, we calculated residue-lipid binding probabilities and performed molecular mechanics generalized Born surface area (MM/GBSA) energy decomposition analysis. Across all sites, PIP2 primarily interacted with polar and charged residues through electrostatic attraction between its negatively charged headgroup and basic amino acids. MM/GBSA analysis revealed a binding free energy hierarchy: C4 strongest (−148.5 ± 18.4 kJ/mol), C0 (−143.5 ± 16.7 kJ/mol), C1 (−114.2 ± 14.2 kJ/mol), C3 (−111.3 ± 13.0 kJ/mol), C5 (−97.9 ± 12.6 kJ/mol), and C2 weakest (−67.4 ± 8.8 kJ/mol) (Tables S3–S8). At C0 (S2-S3 linker), R192 formed the most hydrogen bonds (31.3 ± 3.1%), and R195 showed the most contacts (17.4 ± 0.9%); the lack of a single dominant anchor explains C0’s moderate stability and sensitivity to auxiliary protein competition (Fig. 4b). Extending this decomposition to the twenty highest-ranked residues at C0 (Table S9) places W188, R192, and R195 among the top 10 contributors (−9.6 to −19.2 kJ/mol), whereas V185, R181, and S182 rank markedly lower (−1.4 to −2.7 kJ/mol). These residues have previously been probed by mutagenesis: V185R/W188R shifts deactivation kinetics toward a more KCNQ2-like phenotype⁵⁹; R181E shifts half-activation voltage, while R181Q increases current⁴⁷; and R195Q reduces current and weakens voltage-sensor to pore-domain coupling, whereas R192Q has little effect despite a comparable computed energy to R195⁶². At C1, R249 serves as a pivot residue (14.7 ± 5.4% binding energy), supporting both stable anchoring and transitions to C2 (Fig. 4b). The coordinating residues (R249, W248, Q244, K196) independently corroborate PIP2 contact residues in the KCNQ1-KCNE3-PIP2 cryo-EM structure⁹. At C2 (S4-S5 linker), R249 accounted for ~40% of hydrogen bonding and 18.8 ± 7.0% of binding energy, with weak overall energy (−67.4 kJ/mol) consistent with a transient transfer intermediate. At C3, R259 dominated (26.6 ± 2.7% H-bonds; 18.3 ± 0.9% contacts), and KCNE3 residues R81/R83 at this site imply structural coupling between C1 and C3 (Fig. 4b). Notably, both Cui et al.⁵⁶ and Zhong et al.⁵⁷ independently identified F256, R259, and K362 as key PIP2-contacting residues at the position corresponding to our predicted C3 cluster. This residue-level concordance, reached independently by two experimental groups, corroborates the functional relevance of C3 in KCNQ1 complexes. At C4 (S0-S1 loop), R109 (31.8 ± 3.2% H-bonds) and K121 (17.5 ± 0.9% contacts) dominate, but CaM competes directly through overlapping residues R83/K84 ( Fig. 4b, Table S7). At C5, CaM residue R507 showed the highest hydrogen-bonding occupancy (24.8 ± 2.5%), with T104 and R507 contributing ~90% of binding energy, reflecting the remodeled CaM-active interface (Fig. 4b, Table S8).

To complement the thermodynamic analysis, we extracted dissociation rate constants (k_off) for each site class using PyLipID⁴⁸ across all binding events. The resulting hierarchy (Table S10) shows that C1 and C3 exhibit the lowest k_off values (0.042 ± 0.032 and 0.043 ± 0.030 s⁻¹), consistent with long residence times and stable binding. C0 and C2 display intermediate k_off values (0.055 ± 0.040 and 0.052 ± 0.028 s⁻¹), in line with peripheral and transient roles. Notably, C4 shows a relatively high k_off (0.057 ± 0.035 s⁻¹) despite the most favorable binding free energy (−148.5 kJ/mol), indicating a decoupling between affinity and kinetic stability. C5 exhibits the highest k_off (0.074 ± 0.022 s⁻¹), consistent with a state-dependent, dynamically regulated site. In the context of cryo-EM, where density arises from signal averaging and thus favors interactions that are stable and consistently occupied, such kinetic differences provide a mechanistic basis for understanding which binding sites are more likely to give rise to experimentally observable lipid densities^47,67,68. Overall, the k_off hierarchy reinforces the distinction between stable anchor sites (C1, C3) and transient or regulatory sites (C2, C5), and highlights that strong binding affinity, as observed for C4, does not necessarily correspond to slow dissociation.

Dynamic pathways connect binding sites in a circular transfer network

The identification of six binding sites and their differential occupancy across conditions raised a fundamental question about PIP2 dynamics. Are these sites isolated, with PIP2 molecules binding to one site and remaining there until complete dissociation? Or do PIP2 molecules dynamically transfer between sites, sampling multiple locations during a single binding event? To address this question, we analyze the PIP2 dissociation pathways of all systems.

The spatial layout of binding sites on the C-terminal domain revealed a functional hierarchy. On one side lay C0 (shallow) and C1 (deep), while the opposite side hosted C3 (deep), C4, and C5 (outward). Bridging them was C2 at the S4-S5 linker, suggesting roles in PIP2 redistribution. Across 28 dissociation events, eight representative pathways emerged (Figs. 6 and  S41), categorized as direct, single-transfer, or multi-transfer dissociation.

Direct dissociation pathways allowed PIP2 to exit from the binding region without intermediate transfers. Path 3 enabled direct dissociation from C0, the shallowest site. Path 8 permitted direct dissociation from C5, the CaM-coordinated site in active states. These direct routes provided rapid exit mechanisms from peripheral sites, consistent with their roles in dynamic lipid exchange (Fig. 6b).

Single-transfer pathways involved transitions before exit. For instance, Path 1 described C1-to-C0 transfer, while Path 2 involved C1-to-C4 movement (Fig. 6b). Paths 4 and 7 connected C0 and C4 bidirectionally (Fig. 6b). These revealed that PI2 can traverse multiple regions before dissociation, increasing kinetic complexity.

Multi-transfer dissociation was exclusive to deep site C3. PIP2 dissociation required mandatory transitions-either C3 → C1 (Path 5) or C3 → C4 (Path 6) (Fig. 6b). These transitions involved the HA/HB helices, S5–S6, and auxiliary protein regions. C3 thus acted as a kinetic trap, enhancing VSD-PD coupling by prolonging PIP2 residence time and enforcing conformational rearrangements prior to exit.

Energetic analysis aligned with these pathways. Shallow sites (C0, C5) supported direct dissociation. C1 (−114.2 kJ/mol), centrally located, served as a hub enabling transitions to C0, C4, or direct exit. C2, with the weakest PMF (−67.4 kJ/mol), likely acted as a transient bridge. Deep C3 (−111.3 kJ/mol) served as a kinetic checkpoint, while C4 (−148.5 kJ/mol) anchored PIP2 unless displaced by CaM (Tables S3–S8).

Auxiliary proteins modulated these routes. KCNE3 blocked C0, eliminating Path 3 and stabilizing C1, favoring slower exits via Path 2. It also promoted C3 access (Path 5), enhancing coupling (Fig. 6b). CaM, via interaction with C4/C5, modulated multiple pathways (2, 6, 7), shifting PIP2 traffic in a state-dependent manner (C5 in active states, C4 in inactive) (Fig. 6b).

Combining these eight paths revealed a circular flow network (Fig. 6a). PIP2 entered via C0 or C1, transferred among sites (e.g., C1 ↔ C2, C0 ↔ C4, C1 ↔ C4), and exited directly or after multi-step transitions. From C3, mandatory transfers to C1 or C4 preceded final release. This flow allowed continuous site sampling, with residence times shaped by local energetics and available routes.

Functionally, this network enabled dynamic lipid sensing and response to PIP2 levels. Redundant routes ensured binding even under partial blockade. The C3 requirement functioned as a quality control checkpoint for VSD-PD coupling. KCNE3’s stabilization of C1 and blockade of C0 maintained high occupancy for constitutive activation, while CaM created a regulatory valve at the C4/C5 interface.

In summary, rather than isolated sites, PIP2 binding to KCNQ1 follows an interconnected circulation system governed by binding energetics, pathway architecture, and auxiliary protein state. This regulation ensures both responsiveness and stability of channel function.

Discussion

The SIMDA framework integrates coarse-grained and all-atom molecular dynamics with well-tempered metadynamics and UMAP-Leiden clustering, providing a systematic strategy for mapping lipid-protein interactions across multiple compositional and conformational states. Application to KCNQ1 identified six recurrent PIP2 binding sites and revealed a cooperative regulatory mechanism termed: Twists, Links, and Binding-Site Transfers.

Across all conditions, six recurrent PIP2 binding sites (C0–C5) are identified within three regions of the C-terminal domain. No single complex populates all sites simultaneously; instead, each auxiliary subunit combination occupies a distinct subset of three to four sites. KCNE3 and CaM modulate PIP2 interactions through local, site-specific mechanisms without altering the overall transmembrane architecture. KCNE3 stabilizes PIP2 binding at C1 via direct interactions and enhances S2–S3 linker conformational changes, while also facilitating C3-mediated tightening of VSD-PD coupling. In contrast, CaM exerts state-dependent regulation by destabilizing PIP2 at C4 or stabilizing it at C5, linking calcium signaling to channel gating. These effects are mediated by key residues, including R249 (C0–C2), R259 and KCNE3 residues (C3), R109/K121 (C4), and CaM residue R507 (C5). Beyond static binding, the C-terminal twist at C1 and the loose-link tightening at C3 represent a distinct and separate phenomenon: they are downstream consequences of PIP2 occupancy at these sites rather than determinants of the initial redistribution.

The reliability of individual predictions requires stratification. Sites C1 and C3 are well validated by consistent structural, thermodynamic, and kinetic evidence, including strong Q-scores and agreement with multiple cryo-EM structures^9,45,56,57. C2 shows weak binding, dependence on a single residue, and limited occurrence, consistent with a transient intermediate role. In contrast, C0, C4, and C5 are treated as forward predictions. Their lack of cryo-EM density likely reflects experimental and dynamic factors rather than the absence of binding. C4, despite strong computed affinity, exhibits faster dissociation, which may reduce density accumulation, further influenced by local flexibility and lower resolution. C5 is restricted to the transient CaM active state, while C0 is absent in auxiliary subunit-containing complexes typically used for structure determination. These observations underscore that computed binding thermodynamics and experimentally averaged structural density are complementary but non-equivalent observables.

Although C0, C4, and C5 lack direct structural validation, each is supported by regional functional or biochemical evidence. Mutations in the S2-S3 linker corresponding to C0 alter PIP2 sensitivity and gating in KCNQ1 and related channels^47,59–61. The S0 helix at C4 has been implicated in phosphoinositide binding by NMR, mutational studies, and simulations showing persistent lipid interactions^63,64. For C5, established competition between PIP2 and CaM at the S6-HA junction and CaM-dependent modulation of voltage sensing are consistent with the predicted binding mode^47,66. Structural data further indicate that KCNE3 and CaM do not induce large-scale rearrangements of the KCNQ1 transmembrane core, with changes confined to the cytosolic C-terminal region. Accordingly, PIP2 redistribution is driven by local electrostatic and steric effects from auxiliary subunits, including KCNE3 residues R81 and R83, occlusion of C0, and CaM-mediated repositioning at the C4 C5 region, rather than global structural remodeling. The observed conformational changes, such as the C-terminal twist at C1 and loose link tightening at C3, therefore reflect consequences of PIP2 binding rather than drivers of redistribution.

Several limitations should be noted. The lack of direct experimental structures for C0, C4, and C5 means these predictions depend on simulation accuracy and sampling completeness. While the selected CVs capture dominant motions for C1 and C3, whose free energy surfaces (FESs) are well converged and supported by structural data, alternative binding modes for C0, C4, and C5 cannot be excluded and should be regarded as hypotheses for future testing. Nevertheless, computational prediction of lipid-binding sites ahead of experimental confirmation is well established: the PIP2 site predicted for Kir channels by multiscale simulation was later confirmed, in an independent study, in the Kir2.2-PIP2 structure of Hansen et al.^43,69. More broadly, simulation-derived lipid binding sites have also shown close agreement with experimentally resolved structures in polycystin-2 and GLIC, where simulation and cryo-EM were combined within the same study^38,70. Importantly, the predictive reliability of the present approach is supported within this study itself: two of our predicted sites, C1 and C3, show strong agreement with independently determined cryo-EM PIP2 densities, providing direct internal validation within the KCNQ1 system. On this basis, we present C0, C4, and C5 as experimentally testable forward predictions. The reported PMF values represent within-system dissociation barriers rather than absolute binding free energies, so cross-system comparisons remain qualitative. Although structural comparisons of KCNQ2 suggest minimal CaM-induced changes in the TMD, the absence of CaM-free KCNQ1 structures limits direct validation⁴⁶. Subtle local rearrangements at binding interfaces may also influence PIP2 interactions beyond current sampling. In addition, kinetic estimates for C0, C2, and C5 are based on limited events and carry greater uncertainty. Membrane symmetry and leaflet distribution may further affect interaction energetics, and the coupling between C1 twisting and C3 link tightening requires future experimental clarification. Residue-level energy decomposition at C0 has limited precision, and the resulting ranking of residue contributions should be regarded as an approximate reference range rather than definitive values.

A key strength of SIMDA is its ability to resolve equivalent lipid binding sites across subunits of homo-oligomeric membrane proteins and to compare binding landscapes across different systems. Unlike existing tools such as PyLipID⁴⁸ and LipIDens⁴⁷, which analyze binding within single systems, the UMAP-Leiden strategy enables integrated comparison across multiple compositional states of symmetric complexes, allowing analysis of auxiliary subunit-dependent site redistribution. By clustering contact fingerprints across all subunits and systems, SIMDA overcomes positional degeneracy and generates a unified site atlas directly comparable across conditions. This capability addresses a gap that none of the individual constituent methods, whether CG-MD, AA-MD, or metadynamics, can bridge in isolation, and is broadly applicable to lipid-regulated channels and transporters where binding depends on subunit composition and conformational state.

Future work should prioritize experimental validation of C0, C4, and C5 through targeted mutagenesis combined with PIP2 sensitivity and electrophysiology assays. In particular, we identify W188, R192, R195, V185, R181, and S182 as candidates for mutagenesis at C0, to guide future experimental testing of their roles in PIP2 binding^47,59,62. Extension to other KCNQ family members would test the conservation of the six-site framework across subtypes. Incorporation of cryo-EM density restraints and simulation in asymmetric, physiologically realistic membrane compositions would further refine binding pose predictions and improve the quantitative accuracy of site-specific interaction estimates.

In summary, SIMDA reveals a hierarchically organized, six-site PIP2 regulatory landscape on KCNQ1 in which C1 and C3 are experimentally validated. C0, C4, and C5 represent mechanistically grounded forward predictions with indirect experimental support. The “Twists, Links, and Site Transfers” mechanism provides a unified molecular framework for tissue-specific KCNQ1 regulation and establishes SIMDA as a generalizable multi-scale strategy for dissecting lipid-protein dynamics in regulated membrane protein complexes.

Methods

The primary objective of the SIMDA computational approach is to identify lipid-binding sites within membrane proteins (ion channels) and to progressively filter out key stable binding sites. Here, we provide a detailed description of the methodology and parameter settings.

CG simulation: coarse-grained modeling in SIMDA

The SIMDA method begins by employing coarse-grained molecular dynamics (CG-MD) to explore the dynamics of the Kv7.1 (KCNQ1) channel in both open and closed states. Cryo-EM structures of Kv7.1, retrieved from the Protein Data Bank (PDB IDs: 6V00, 6V01), were used to simulate the effects of different cofactors on PIP2 binding⁹. Proteins were prepared using the Protein Preparation Wizard, which includes filling missing loops, repairing incomplete side chains, optimizing these regions, and assigning protonation states to residues in Schrödinger 2020 software⁷¹. The protein was then converted to a coarse-grained resolution using the martinize.py tool, incorporating the MARTINI version 2.2 force field. An elastic network was applied with an elastic bond force of 500 kJ/(mol nm²) and a cutoff of 0.9 nm⁷²,⁷³. A symmetric bilayer containing 5% PIP2 and POPC as the background lipid was constructed using insane.py⁷⁴. Two different membrane compositions were employed: POPC: POP5 (95:5) and POPC: POPG: POP5 (90:5:5). For the POPC: POP5 system, 236 POPC and 12 POP5 molecules were placed in both the inner (cytoplasmic) and outer (extracellular) leaflets, and 216 POPC and 11 POP5 molecules in the lower leaflet. For the POPC: POPG: POP5 system, the upper leaflet consisted of 224 POPC, 12 POPG, and 12 POP5 molecules, while the lower leaflet contained 205 POPC, 11 POPG, and 11 POP5 molecules. All systems were built in a 14 × 14 × 14 nm³ simulation box and solvated using the standard MARTINI water model. The topology of PIP2 was adapted from phosphatidylinositol 3,4-bisphosphate (PI34), excluding the dihedral potential of the headgroup⁷⁵. The system was solvated with standard MARTINI water with van der Waals radii of 0.21 nm, followed by neutralization and the addition of ions to achieve a concentration of 0.15 M KCl⁷⁶. Each replica system was initially relaxed through 5000 steps of steepest descent energy minimization, followed by the 5 ns equilibration phase. 30 µs production simulations were conducted, each with distinct random initial velocity seeds. The temperature was kept constant at 300 K using the V-rescale thermostat⁷⁷, with a time constant of τ_t = 1 ps. Pressure was maintained at 1 bar by employing a Parrinello-Rahman barostat⁷⁸, with a coupling time of τ_p = 5 ps and a compressibility of 3 × 10⁻⁴ bar⁻¹. The protein, lipids, and solvent (water and ions) were coupled individually. Electrostatic and van der Waals interactions were smoothly shifted between 0 and 1.2 nm and 0.9 and 1.2 nm, respectively, and the 20 fs integration time step was used⁷⁹. CG-MD simulations were conducted using GROMACS 2021.4⁸⁰, with each protein simulated for 30 × 8 μs in the respective bilayer composition, except for the CaM-activated system, which was simulated for 30 × 8 μs to ensure sufficient lipid-interaction sampling.

PIP2 binding site identification and initial screening in SIMDA

We employed PyLipID (http://github.com/wlsong/PyLipID)⁴⁸ to analyze the interactions between PIP2 and proteins. The occupancy of PIP2 at each binding site was calculated as the fraction of simulation time that any PIP2 bead remained in contact with a protein residue bead. Contacts were defined using cutoffs of 0.5 nm and 0.9 nm, with any distance within this range considered a valid interaction. Next, binding sites were identified by grouping residues that interacted with PIP2 consistently throughout the trajectory, using a community-based clustering approach. For each identified site, the top three binding poses were extracted and further clustered using an automatic clustering method provided by PyLipID. Additionally, the surface area of interaction between PIP2 and the protein was used to assess the significance of each binding site. The surface area for a given site was calculated based on the contact area between PIP2 and the protein residues involved in the interaction, using the accessible surface area (ASA) method⁸¹. The results from PyLipID were mapped onto the protein structures provided in the PDB input files. This included top-ranked binding poses, clusters of binding poses, and kinetics of each residue and site, which were encoded into the B-factor column of the output PDB files. These outputs, such as top-ranked binding poses (“BSidX_rank”) and clustered poses (“BSidX_clusters”), were visualized using VMD.

Since not all PyLipID-identified sites are necessarily biologically relevant, additional criteria were used to filter binding sites. Specifically, binding sites were evaluated based on three metrics: binding site occupancy, duration, and surface area. As shown in Fig. S4, binding sites were ranked according to these criteria, from highest to lowest. Sites with low lipid residence time were excluded from further analysis.

CG-to-AA transformation: transitioning from coarse-grained to all-atom models in SIMDA

We employed the backward transformation method via the initram interface to convert coarse-grained (CG) models into all-atom (AA) models^50,51. This method initiates from the CG model and progressively restores molecular details through backmapping, reconstructing a high-resolution all-atom structure. During the transformation, initram maps CG beads onto atomic coordinates following predefined mapping protocols, ensuring the retention of dynamic information and preserving the model’s geometric and physical integrity. The resulting AA model undergoes energy minimization using GROMACS 2021.4 to eliminate any structural distortions and ensure conformational stability⁸⁰. Following minimization, we performed additional MD simulations with AA force fields (CHARMM36m) to validate the accuracy of the model, ensuring that both the dynamic behavior and atomic-level detail were maintained throughout the transformation process⁸². This approach leverages the computational efficiency of CG simulations while achieving the high resolution of AA models, providing a reliable structural framework for subsequent MD simulations and detailed analysis.

AA simulation: detailed all-atom dynamics in SIMDA

The box of both all-atom protein and lipids was first solvated with TIP3P water⁸³, with counter ions added to neutralize the system and achieve a physiological ionic strength of 0.15 M KCl. The energy of the system was minimized to remove any steric clashes or unfavorable interactions. Then, the system was equilibrated in two phases: equilibration of the canonical ensemble (NVT: constant number of particles, volume, and temperature) followed by equilibration of the isothermal-isobaric ensemble (NPT: constant number of particles, pressure, and temperature), both for 1 ns.

For the production MD simulation, 300 ns simulation was performed under NPT conditions with periodic boundary conditions (PBC) applied in all directions⁸⁴. The temperature was maintained at 300 K using a V-rescale thermostat with a time constant of 1.0 ps⁸⁵, while pressure was maintained at 1 bar using a Parrinello-Rahman barostat⁸⁶ with a coupling constant of 2.0 ps. The integration time step was set to 2 fs, and long-range electrostatic interactions were handled using the Particle Mesh Ewald (PME) method with a cutoff of 1.0 nm for both Coulomb and van der Waals interactions⁸⁷. The hydrogen bond lengths were constrained using the LINCS algorithm⁸⁸. Trajectories were saved every 10 ps for subsequent analysis.

Secondary screening: refining binding sites in SIMDA

After completing the AA-MD simulations for each system and binding site, the root-mean-square-deviation (RMSD) between the lipid molecules at each binding site and the corresponding protein was calculated to assess the stability of lipid-protein interactions. Low and stable RMSD indicated strong lipid binding, whereas significant fluctuations suggested unstable or transient interactions. Based on the RMSD, stable lipid-binding sites were selected, and sites with large fluctuations were excluded. This screening process allows for further refinement, focusing on key binding sites that demonstrate high stability during the simulation, enabling more detailed analysis.

UMAP-Leiden clustering in SIMDA

The identification and clustering of binding sites within individual systems alone do not readily allow for the comparison of binding differences across different protein states or cofactor modulation. To address this limitation, after two-step screening, we first normalized the occupancy data of binding site residues from all systems using the natural logarithm transformation (Log1p)^89,90. This transformation reduces the magnitude differences across the data, minimizes noise, and preserves the sparsity of the dataset. The transformation is defined as formula 1:

$$x^{\prime }=\log \left(1+x\right)$$ 1

where $x$ represents the original occupancy value, and $x^{\prime }$ is the transformed value after applying the log transformation. This approach ensures that large variations in occupancy are compressed while maintaining the relative differences between low occupancy values⁸⁹.

Next, the Leiden clustering algorithm was employed to group binding sites based on their occupancy profiles⁵⁴. The Leiden algorithm optimizes modularity to maximize internal connectivity within clusters⁹¹. The modularity Q is shown in formula 2:

$$\mathrm{Q}=\frac{1}{2m}{\sum }_{i,j}\left[A_{ij}-\gamma \frac{k_i{k}_j}{2m}\right]\delta \left(c_i{c}_j\right)$$ 2

where $A_{ij}$ represents the adjacency matrix (occupancy similarity between binding sites), $k_i$ and $k_j$ are the degrees of nodes $\mathrm{i}$ and $\mathrm{j}$, $\gamma$ > 0 is a resolution parameter, and $m$ is the total number of edges in the graph⁵⁴. The resolution parameter for clustering was set to 2.0 to control the granularity of the clusters.

To further visualize the relationships between the binding site clusters, we applied Umap (Uniform Manifold Approximation and Projection) for dimensionality reduction^52,53. Umap projects the high-dimensional occupancy data into a lower-dimensional space while preserving the local neighborhood structures of the binding sites. The objective of Umap is to minimize the following cost function:

$$\mathrm{L}={\sum }_{i,j}{w}_{ij}{\parallel f\left(x_i\right)-f\left(x_j\right)\parallel }^2$$ 3

where $w_{ij}$ is the weight based on the local proximity of data points nodes $i$ and nodes $\mathrm{j}$ in high-dimensional space, and $f\left(x\right)$ is the low-dimensional embedding function. This method allows for a clear visualization of clusters in lower dimensions, helping to capture the underlying structure of the data⁹². Finally, through the application of Leiden clustering and Umap, we identified clusters of binding sites that share similar residue occupancy profiles. These clusters revealed the heterogeneity in binding site occupancy across different protein systems and cofactor modulations. The combination of clustering and dimensionality reduction highlighted key differences in the interactions of PIP2 with various protein states and cofactors, providing insights into the differential modulation of these binding sites.

FES calculation based on Wt-metaD simulation in SIMDA

WT-metaD has been integrated into the SIMDA framework to investigate lipid binding and dissociation dynamics at protein binding sites. The Wt-metaD method enhances sampling efficiency by moderating the growth of the bias potential over time, preventing overfilling of the energy landscape⁹³. This is crucial for accurately capturing events such as lipid binding and dissociation. To effectively model these processes, critical collective variables (CVs) are selected, including the distance between lipid head groups and binding site residues, as well as the angular orientation of the lipid relative to the protein surface⁹⁴. These CVs were carefully chosen to ensure comprehensive sampling and to accurately capture the binding dynamics across the free energy landscape^94,95.

The calculation of the FES was based on the Wt-metaD simulation, which provided a detailed mapping of the energy landscape of the system under study. By moderating the growth of the bias potential, Wt-metaD improved upon traditional metadynamics, ensuring that the energy landscape is explored without excessive bias, leading to more reliable free energy profiles. In

Wt-metaD, the bias potential $V\left(s,t\right)$ was progressively reduced using the formula4:

$$V\left(s,t\right)={\sum }_{\tau <t}{\omega e}^{V\left(s,t\right)/\mathrm{\Delta }T}\exp \left(-\frac{{\left(s-s\left(\tau \right)\right)}^2}{2{\sigma }^2}\right)$$ 4

where $\omega$ was the height of Gaussian potentials, $\mathrm{\Delta }T$ was a tempering parameter affecting the decay rate of the potentials, $\sigma$ was the width of the Gaussian, and $\tau$ represented the simulation time steps^93,96. This setup ensured controlled exploration of the energy landscape to prevent potential bias⁹⁷. This setup prevented the potential from growing uncontrollably, allowing for a more accurate reconstruction of the FES⁹⁷. Upon convergence of the simulation, the accumulated bias potential was used to reconstruct the FES, which revealed the energetic barriers and stable states associated with lipid binding and dissociation⁹⁸. This free energy profile was essential for understanding the underlying mechanisms of lipid–protein interactions, offering valuable insights into the dynamics of lipid binding at protein sites.

Here, the X and Y components of the distance between each PIP2 and residues within 10 Å were chosen as the collective variables CV1 and CV2, respectively. For both CVs, the bin size was set to 0.1 Å, the Gaussian width (σ) to 0.05 Å, the simulation temperature (T) to 300 K, and deposition time (τ) to 2 ps for the Wt-metaD simulations. The restraints were implemented as half-harmonic walls with force constants of 200 kJ/(mol Ų). The Wt-metaD simulation was terminated when the lipid dissociated from the binding site, typically after observing several binding and dissociation events. The duration of these simulations ranged were ~400 ns based on the rate of PIP2 dissociation events.

Calculation of MFEP and PMF for lipid binding and dissociation in SIMDA

The minimum free energy path (MFEP) was crucial for understanding the most likely transition path during events like lipid dissociation or binding. It was defined as the trajectory connecting two stable minima via the lowest possible energy barriers on the FES^99,100. The MFEP was obtained using the nudged elastic band (NEB) method within the Metadynminer package^101,102. This method interpolated a series of optimized states along the energy landscape. The process involved several key steps: (1) fixing two local minima as the starting and ending points of the pathway; (2) generating an initial coarse path on the FES, represented by discrete images corresponding to different states along the pathway; and (3) applying the NEB method to refine the path by connecting images with springs of constant k, simulating an elastic band, and minimizing forces along the band to converge to the MFEP. Sensitivity parameters were sequentially adjusted to ensure precision and convergence, with settings of 10,000 steps (nsteps), a step size of 0.01, and a spring constant (k) of 1. The optimization process further refined the pathway to accurately depict the free energy change from start to end states. Each dissociation pathway was represented by 100 points, proportional to the MFEP distance between the two basins.

MM/GBSA energy decomposition calculations

Molecular mechanics generalized Born surface area (MM/GBSA) analysis was performed to quantify PIP2 binding free energies and identify key residue contributions at each binding site. Calculations were conducted on the last 200 ns of AA-MD trajectories for all stable binding sites across all systems and replicates. The MMGBSA approach employed the generalized Born implicit solvent model with parameters optimized for membrane protein systems.

Energy calculations included three components: molecular mechanics energy (electrostatic and van der Waals interactions), polar solvation energy (calculated using the generalized Born model with igb=2), and nonpolar solvation energy (estimated from solvent-ASA)¹⁰³. Salt concentration was set to 0.15 M to approximate physiological ionic strength. The binding free energy (ΔG_bind) was calculated as the difference between the complex energy and the sum of individual component energies (protein and PIP2). All MMGBSA calculations were performed using the MMPBSA.py module in AmberTools, with results averaged over frames extracted at 10 ps intervals from the final 200 ns of each trajectory.

MapQ score calculations

To quantitatively assess the consistency between the binding poses sampled in molecular dynamics (MD) simulations and experimentally resolved structures, we performed model-to-map fitting using cryo-electron microscopy (cryo-EM) density maps as references. Representative MD-derived conformations were extracted via clustering based on backbone RMSD of binding site residues and subsequently aligned to the corresponding cryo-EM maps using UCSF Chimera (version 1.16)¹⁰⁴. The following experimentally resolved protein-ligand complexes were used as references: PDB ID: 6V01/EMDB ID: EMD-20967 (3.9 Å resolution), PDB ID: 9LIZ / EMDB ID: EMD-63128 (3.1 Å), and PDB ID: 7VNP/EMDB ID: EMD-32044 (2.79 Å). For all density maps, a resolution setting of 3.0 Å and a sigma value of 0.496 were applied to standardize fitting conditions.

Following alignment, Q-scores (MapQ scores) were computed to evaluate the degree of local agreement between atomic coordinates and the electron density. The Q-score provides a quantitative measure based on the correlation between the model-derived atomic density and the experimental map, with higher scores indicating better model-to-map fit and more reliable atomic placement^55,105. To dissect model quality across different structural regions, Q-scores were calculated separately for the lipid, binding pocket residues, and the overall complex.

Statistics and reproducibility

No statistical method was used to predetermine sample size. Sampling adequacy was instead established by convergence of the PIP2 density maps by 20 µs and by reproducibility across eight independent replicates per system (64 trajectories in total), assessed by pairwise cosine similarity of binding-site fingerprints. Binding sites were retained or excluded using predefined kinetic and structural criteria (residence time >0.5 µs, occupancy >50%, contact area >10 nm², and all-atom stability thresholds); sites failing these criteria were excluded, and the exclusion criteria are stated in the Methods and Results. The experiments were not randomized; randomization is not applicable because all systems were built and simulated under identical, defined protocols. The investigators were not blinded to allocation during experiments and outcome assessment, as blinding is not applicable to deterministic molecular dynamics simulations analyzed by automated, criteria-based pipelines.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Author contributions

L.W. and H. Liu conceived the study. L.W. developed the SIMDA framework, wrote the analysis code, performed the simulations and data analysis, and wrote the manuscript. S.L. performed the initial modeling and data analysis and contributed to manuscript revision. Y.Z. and H.S. reviewed the manuscript and provided suggestions for revision. Q.L., W.Z., X.L., X. Yan, X. Yao, and H.H.Y.T. participated in discussions on method implementation. H.Liu provided computational resources, the initial concept, and manuscript revision. L.W. and S.L. contributed equally to this work. All authors read and approved the final manuscript.

Peer review

Peer review information

Nature Communications thanks Ben Corry, Indra Sahu and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. A peer review file is available.

Funding

This work was supported by the Science and Technology Development Fund, Macau, SAR (No. 0091/2022/A2) and Macao Polytechnic University (No. RP/FCA-02/2023).

Data availability

The atomic coordinates used to initiate the simulations were obtained from the RCSB Protein Data Bank: 6V00 and 6V01. Experimental structures used to validate the predicted PIP2 sites were obtained from the Protein Data Bank: 9WD8, 9VEI, 9UC8, 9LIZ, and 7VNP. The simulation input files, force-field parameters, and processed data supporting the findings of this study have been deposited in Zenodo [https://doi.org/10.5281/zenodo.14885900].  Source data are provided with this paper.

Code availability

The SIMDA framework is freely available under an MIT license on GitHub [https://github.com/xiaoxiaowang1201/SIMDA-framework] and archived on Zenodo [https://doi.org/10.5281/zenodo.14885900]. The version used in this study is citable through this DOI.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s41467-026-76627-9.