Uncovering the biological mechanisms of TREM2 with molecular simulations: A comprehensive review and perspective.

Aggregation of membrane-bound TREM2 monomers into non-covalently bound dimers or trimers via antibodies is thought to potentiate signaling, thereby amplifying the downstream effects on microglia function 11 , 13 . For TREM2 multimers, this signal amplification may arise from increased recruitment and clustering of DAP12 ( Fig. 3B ), promoting cooperative cross-phosphorylation and enhanced Syk phosphorylation/activation, though direct validation of this hypothesis is needed. While the canonical 1:2 TREM2:DAP12 relationship holds for the monomeric receptor (see Section 5 ), how this stoichiometry scales within a crosslinked multimer—and how it influences signal amplification—remains an open question. Defining the molecular determinants of TREM2 multimerization, and how disease-associated variants disrupt these assemblies, is therefore of direct relevance to the design of multimerization-modulating therapeutics. An important caveat at the outset is that the TREM2 ectodomain does not appear to self-associate in solution: biophysical experiments have detected only monomeric sTREM2 under soluble conditions 104 . This means that the PS-stabilized hexameric and trimeric assemblies observed crystallographically 18 , and used as starting structures in the simulations discussed below, likely reflect a membrane- and crystal-packing-facilitated state rather than constitutive solution-phase multimerization. The computational studies in this section should therefore be interpreted as models for understanding the structural consequences of forced or antibody-induced oligomerization, rather than as descriptions of spontaneous self-assembly. Structural insights from X-ray crystallography provided the starting point for these simulations: Sudom et al. 18 resolved PS-stabilized TREM2 trimers (PDB 6B8O), in which two alternating trimers stack to form a hexamer. The assembly is anchored by electrostatic interactions between negatively charged PS headgroups and a positively charged surface depression formed by CDR2 residues, establishing PS as both a ligand and a structural mediator of TREM2 oligomerization in the crystalline context. Whether these crystallized orientations reflect membrane-bound states in vivo remains uncertain, as steric constraints from the flexible stalk domain and associated coreceptors could substantially alter the geometry of the multimer interface. Nonetheless, this structure has served as the primary template for subsequent computational investigations of TREM2 multimerization. Dean et al. (2024) 105 used the 6B8O structure to examine how AD-risk mutations affect the stability and energetics of both dimeric and trimeric TREM2 assemblies. Importantly, these simulations were conducted with AMBER ff14SB—a force field well-suited to protein backbone and side-chain interactions but lacking dedicated parameters for lipid acyl chains and phospholipid headgroup electrostatics 106 . For a system where PS-mediated contacts are central to the multimerization interface, this is a meaningful limitation: protein-lipid electrostatic interactions at the CDR2 binding surface may be misrepresented, potentially biasing conclusions about whether PS stabilizes or destabilizes specific multimeric configurations. This should be borne in mind when interpreting the PS-related findings below. The simulations revealed that tail-to-tail TREM2 WT dimers isolated from the hexameric assembly dissociated rapidly in early simulation regardless of whether PS or mutations were present, suggesting that this specific dimer geometry is only stable as part of the larger hexameric assembly rather than as an independent species—consistent with the solution-phase monomer observation noted above 104 . After initial dissociation events, a stable WT dimer was eventually achieved via a R76–D87 interfacial salt bridge, and WT trimers remained stable across replicates. By contrast, trimers containing the R47H mutation exhibited heightened dynamics, CDR2 flexibility, and secondary structure loss—notably, in the opposite direction from the CDR2 rigidity observed for R47H monomers in Lietzke et al. 31 This apparent discrepancy is most likely attributable to the different starting structures used: Lietzke et al. 31 built from the TREM2 R47H crystal structure (PDB 5UD8 18 ), where CDR2 is resolved as a tight helix, whereas the WT hexameric structure 6B8O 18 used by Dean et al. (2024) 105 already exhibits reduced CDR2 helicity before any mutation is introduced. Computational insertion of R47H into 6B8O thus begins from a different structural baseline, and the oligomeric packing environment itself may further amplify CDR2 disorder in ways that are not present in the monomer. Together, these differences suggest that CDR2 behavior in R47H is sensitive to both starting conformation and oligomeric context—a point worth bearing in mind as the field moves toward modeling full-length, membrane-embedded multimers. The role of PS across these multimeric configurations proved difficult to interpret cleanly. In R47H-containing trimers, the presence of PS caused dissociation that was not observed in its absence, while in WT tail-to-tail dimers, PS appeared to prevent stable assembly via the R76–D87 salt bridge—a finding that contrasts with PS's established role as a stabilizing ligand at the CDR2 interface in the hexameric crystal structure. The force field caveat above is particularly relevant here: without reliable lipid–protein electrostatics, it is unclear whether PS-associated dissociation or instability in these simulations reflects genuine biology or a computational artifact. What can be said with confidence is that the relationship between PS, multimer geometry, and variant background is more complex than a simple "PS stabilizes multimers" model—and that resolving it will require force-field-appropriate simulations with longer timescales and systematic PS inclusion across all variant and assembly combinations. The D87N, R98W, P37D, and G90D variants provided additional mechanistic insight into the interfacial salt bridge network ( Table 4 ). D87N directly ablated the R76–D87 salt bridge, predictably destabilizing trimers and leading to dissociation in some simulations. R98W trimers did not dissociate, but like R47H, showed modestly reduced R76–D87 occupancy and slightly weaker inter-monomer binding affinity, suggesting that AD-risk mutations broadly—and through distinct mechanisms—weaken the same interfacial contact. The engineered P37D and G90D variants introduced new salt bridges that enhanced interfacial strength in some configurations but at the cost of reduced R76–D87 occupancy, raising the question of whether compensatory electrostatic contacts can substitute for the native interface without adverse effects on CDR2 dynamics or ligand accessibility. Steered MD provided orthogonal support for these conclusions: greater mechanical force was required to dissociate a monomer from WT trimers than from R47H or D87N trimers, offering direct evidence that AD-risk mutations reduce the mechanical stability of the multimerization interface. Dantas et al. (2025, Chem Phys Lipids) 107 examined a subset of these questions with a complementary focus on how distinct lipid species—PS versus cardiolipin—influence dimer stability, using 100 ns simulations without replicates on the same tail-to-tail dimer geometry. Consistent with Dean et al. (2024) 105 , unliganded dimers were unstable. Dimers bound to PS or cardiolipin showed modestly greater stability over the 100 ns window; however, given that Dean et al.'s replicated, longer simulations still showed PS-associated dimer dissociation, this apparent stabilization is better interpreted as a short-lived effect that does not persist at longer timescales. The two studies also differ in force field: Dantas et al. 107 used CHARMM36 with CGenFF lipid parameters, which more accurately represents protein–lipid electrostatics than the AMBER ff14SB used by Dean et al. (2024) 105 . The finding that cardiolipin—an anionic lipid enriched in mitochondrial membranes but present at low levels in the plasma membrane—showed similar stabilizing effects to PS suggests TREM2 may interact promiscuously with negatively charged lipid headgroups at the CDR2 surface, independent of specific lipid acyl chain identity. Whether this reflects a meaningful biological interaction or a non-specific electrostatic effect is worth examining in longer, replicated simulations with physiologically representative lipid compositions. In summary, simulation studies of TREM2 multimerization have established that the R76–D87 interfacial salt bridge is a central determinant of trimer stability, that AD-risk mutations consistently weaken this contact through both direct (D87N) and indirect (R47H, R98W) mechanisms, and that PS exerts complex, context-dependent effects on multimer geometry whose interpretation is currently limited by force field and sampling constraints. From a therapeutic standpoint, these findings highlight both an opportunity and a caveat: compounds designed to reinforce interfacial salt bridges or otherwise stabilize TREM2 multimers could potentiate signaling in disease contexts, but the sensitivity of multimer stability to lipid environment and simulation methodology means that in silico predictions in this space will require careful experimental validation. Given that isolated sTREM2 ectodomain is monomeric in solution, such stabilization strategies are likely most relevant in the membrane-proximal context where antibody crosslinking or lateral receptor clustering can drive assembly. As summarized in Table 4 , the current computational picture provides a mechanistic foundation for variant-specific differences in multimerization, but systematic, long-timescale simulations with appropriate lipid force fields and biologically realistic membrane compositions remain essential next steps.

MD simulations provide a powerful framework for probing TREM2 mechanisms at atomic resolution—offering insights that are difficult or even impossible to obtain experimentally. Over the past decade, advances in sampling algorithms, force fields, and computational power have enabled increasingly realistic representations of TREM2 behavior, illuminating how structural rearrangements within overlapping binding surfaces, in both TREM2 WT and disease-associated variants, influence ligand recognition and downstream signaling. Yet substantial gaps remain, and addressing them in a targeted, prioritized manner will be essential for the field to advance beyond the current generation of models. A critical near-term priority is the incorporation of N-linked glycosylation into TREM2 simulations. TREM2's Ig-like domain contains two confirmed N-linked glycosylation sites—N20 and N79—that are present in all mammalian-expressed TREM2 structures resolved to date 176 and are known to influence folding, stability, and ligand accessibility. N79 sits directly at the end of the strand connected to CDR2, adjacent to the PLIR/Surface 1—the very region most implicated in ligand recognition and most perturbed by disease-associated mutations. Given this proximity, glycans at N79 may directly modulate CDR2 loop dynamics and shield or reshape the ligand-binding surface in ways that no current simulation captures. The R47H mutation additionally alters TREM2's glycosylation profile relative to WT 18 , 177 , and—critically—the TREM2 R47H crystal structure (PDB 5UD8) was produced in E. coli and is therefore entirely aglycosylated, unlike all other resolved TREM2 structures. This raises the specific concern that the open CDR2 conformation observed in 5UD8, and reproduced in simulations built from it, may partly reflect the absence of glycans rather than the mutation alone—an uncertainty that cannot be resolved without glycan-inclusive simulations of R47H. Glycan force fields capable of addressing this are available: GLYCAM06 and the CHARMM36 carbohydrate parameter set have been validated for glycoprotein MD studies and are compatible with existing TREM2 simulation workflows. The primary barriers are the added complexity of modeling branched, heterogeneous glycan chains and the extended sampling required to equilibrate glycan conformational ensembles—real but surmountable challenges, particularly as high-mannose glycan compositions at N20 and N79 have already been characterized experimentally 21 , 176 , providing a defined starting point. Incorporating even a single high-mannose N-glycan at N79 in near-term simulations of both WT and R47H would represent a meaningful and tractable advance. Another critical factor in MD simulations is the choice of the initial TREM2 structure. Although existing TREM2 structures have been determined using X-ray diffraction or NMR, these static models cannot fully capture dynamic features of the receptor, such as conformational fluctuations within CDR2. They nonetheless provide essential starting points for simulation, but interpretation of resulting dynamics requires careful consideration of stabilizing factors, construct design, and expression systems used to generate the experimental structures. In parallel, emerging computational tools, including AlphaFold 24 or MODELLER 178 , have expanded the scope of MD simulations, enabling studies of full-length proteins or disease-associated variants that have not yet been structurally resolved. While these approaches allow access to otherwise intractable systems, transparency in model construction is critical. MD simulations are inherently sensitive to initial conformations, particularly when missing residues are modeled or mutations are introduced into existing structures. Given MD’s strength in characterizing conformational ensembles rather than single states, future studies would benefit from sampling multiple TREM2 starting structures to ensure that observed behaviors reflect intrinsic receptor dynamics rather than artifacts of specific initial models or simulation parameters. Finally, a major frontier is simulating full-length, membrane-embedded TREM2. Most structural and computational studies to date have focused on the isolated Ig-like domain, implicitly assuming that its conformational dynamics and ligand-binding patterns represent those of the intact receptor. However, this simplification neglects the electrostatic and steric influences of the stalk, TMD, and the membrane itself. The long, flexible stalk likely modulates the position and orientation of the Ig-like domain relative to the membrane, while TMD interactions at the TREM2-DAP12 interface are crucial for initiating intracellular signaling. Understanding how ligand binding propagates through these domains to generate differential signaling outcomes remains a central challenge. Integrating these missing components—glycosylation, the full-length receptor with its signaling partners, and realistic membrane environments—will be critical for developing predictive, biologically relevant models of TREM2 function. The next generation of TREM2 simulations should therefore aim to combine atomistic and coarse-grained approaches with experimental validation through rationally designed, simulation-guided assays that directly test predicted binding or signaling mechanisms. This synergistic framework will not only deepen mechanistic understanding but also accelerate the rational design of therapeutics that precisely modulate TREM2 activity across diverse physiological and pathological contexts.

TREM2 signals through its coreceptors DNAX-activating proteins (DAP; typically DAP12 or DAP10 of 12/10 kDa, also known as TYROsine Kinase Binding Protein (TYROBP)), which are required to propagate downstream signaling cascades critical to microglial function 86 , 87 . DAP12 forms a homodimer via a cysteine-mediated disulfide bond between its two short extracellular domains 88 . Because these domains are too short to independently bind ligands, DAP12 cannot signal alone and instead depends on its association with TREM2. The current structural model posits a 1:2 TREM2:DAP12 stoichiometry, in which a single TREM2 transmembrane helix associates with the DAP12 homodimer via a salt bridge between TREM2 K186 and D50 residues of each DAP12 chain 23 , 89 , 90 ( Fig. 3B , bottom ). This electrostatic interaction enables the TREM2–DAP12 complex to signal through the immunoreceptor tyrosine-based activation motif (ITAM) in the intracellular domain of DAP12 86 , 87 , 91 . Upon ligand binding to TREM2, tyrosine residues in ITAM are phosphorylated by spleen tyrosine kinase (Syk), triggering a variety of downstream signaling cascades. Despite this structural framework, many questions remain unresolved: how different ligand interactions with TREM2 lead to distinct cellular signaling outcomes 32 , 92 , how signal transduction through the TREM2–DAP12 salt bridge produces differential downstream effects, and how disease-associated mutations in the ECD are mechanistically communicated across the full-length receptor to alter TMD conformation and DAP12 coupling. Addressing these questions is a prerequisite for effectively targeting TREM2 signaling with small molecules, peptides, or antibodies in disease. The foundation for MD simulations of TREM2's TMD was laid by Steiner et al. 23 , who used NMR spectroscopy to resolve the structure of the TREM2 TMD, revealing a 'kinked' conformation within the membrane (PDB 6Z0G). This kink was lost when K186 was mutated to alanine (PDB 6Z0H) or when TREM2 was complexed with DAP12, leading to the proposal that DAP12 binding stabilizes a straighter, 'unkinked' helix. In this model, the unkinked conformation represents the active, signaling-competent state, while the kinked conformation reflects the unbound, membrane-sequestered and likely inactive state. Supporting this interpretation, atomistic MD simulations of TREM2 WT TMD in a POPC/POPG bilayer—in the absence of DAP12—showed that the positively charged K186 is drawn toward the negatively charged phosphate headgroups of the surrounding lipids, providing a physical driving force for helix kinking. Simulations of TREM2 K186A confirmed that removing this charged residue abolished the kink, consistent with NMR data, though this mutant remains signaling-incompetent as it cannot form the K186–DAP12 salt bridge. These simulations spanned "up to 150 ns" with no replicates, and because MD was a secondary rather than primary focus of the study, force field details were not fully reported. The absence of DAP12 also limits direct applicability to the signaling complex. These caveats notwithstanding, Steiner et al. 23 established the essential conceptual framework—kinked/unkinked TMD states as proxies for inactive/active signaling—that shaped all subsequent TREM2 TMD simulations. Zhong et al. (2024) 93 substantially extended this framework using a multiscale approach, combining 200 μs coarse-grained MD (MARTINI 2.2P force field) with 300 ns atomistic MD (CHARMM36), in triplicate, to probe TREM2–DAP12 complex dynamics. Starting from HADDOCK-docked complexes of both the kinked and unkinked TREM2 TMD conformations (from Steiner et al. 23 ) with DAP12, as well as an AlphaFold-predicted unkinked structure, all models were embedded in an 80:20 POPC:cholesterol bilayer—a composition more representative of microglial membranes than the POPC/POPG bilayer used by Steiner et al. Across all starting models, both kinked and unkinked conformations remained stable in complex with DAP12, converging to distinct but persistent states rather than interconverting. The canonical K186–DAP12 salt bridge was maintained in both, demonstrating that salt bridge formation alone is insufficient to define a signaling-competent state. A more discriminating structural feature emerged: the unkinked complex formed hydrogen bonds with both DAP12 chains, whereas the kinked complex engaged only one . This dual-chain H-bonding, in addition to the shared salt bridge, may constitute the minimal structural requirement for productive downstream signaling—a hypothesis that, if validated experimentally, would meaningfully sharpen our understanding of what makes the TREM2–DAP12 interface signaling-competent versus merely assembled. Zhong et al. (2025) 94 built directly on this model to examine how loss-of-function TMD variants perturb complex integrity. Simulations of K186A and W194A mutations in full TREM2 230 in complex with DAP12 confirmed that K186A reduces TREM2–DAP12 contacts and decreases TMD stability, reinforcing the central role of K186. Intriguingly, both K186A and W194A abolished H-bonding at residue 194 with DAP12 while simultaneously promoting hydrophobic interactions there, with K186A additionally driving π–π stacking at W194. This suggests that W194 acts as a secondary interaction node that can compensate—partly and imperfectly—when the primary K186 salt bridge is disrupted, by shifting toward hydrophobic rather than electrostatic engagement. The broader implication is that signaling fidelity at the TMD interface depends not just on K186, but on a coordinated network of contacts across multiple residues, where perturbation at one node propagates allosterically to reshape interactions elsewhere in the TMD. Zhong et al. (2025) 94 also simulated the alternatively spliced TREM2 219 isoform in complex with DAP12; however, as discussed in Section 4 , experimental evidence indicates this isoform is primarily expressed as a soluble protein lacking the amphipathic helix needed for membrane embedding 85 , limiting the biological relevance of those particular simulations. Two additional studies round out the published TREM2–DAP12 computational literature. Garcia-Alberca et al. 95 reported MD simulations of the TREM2–DAP12 complex and included trajectory videos, but the absence of reported methods, quantitative analyses, and substantive discussion makes it difficult to assess the validity or significance of their findings. Cobas-Carreño et al. 96 introduced a drug-screening pipeline targeting the TREM2–DAP12 TMD interface; while no new MD simulations are presented, the work leverages Steiner et al. 23 and Zhong et al. 93 , 94 to identify candidate small molecules that insert between TREM2–DAP12 transmembrane helices and engage hydrophobic contacts. The prospect of a transmembrane agonist that stabilizes the TREM2–DAP12 interface—potentially in combination with an ECD-targeting compound—represents a plausible synergistic therapeutic strategy, though it awaits mechanistic validation. Taken together, these studies converge on a model in which the TREM2 TMD exists in at least two stable conformational states—kinked (inactive) and unkinked (potentially active)—whose interconversion is regulated by K186-mediated electrostatic interactions with DAP12 and by the broader lipid environment. Yet this model is still far from complete. Biologically relevant membrane composition remains a critical and underexplored variable: while the inclusion of cholesterol in Zhong et al. 93 , 94 is an advance over purely POPC bilayers, tissue-specific microglial membrane compositions in both homeostatic and disease states differ considerably and are known to influence TMD behavior in related receptor systems 97 - 100 . Membrane-embedded cofactors beyond DAP12—including heparan sulfate proteoglycans 44 and integrins 101 —have not yet been incorporated into any TREM2 TMD simulation, despite experimental evidence for their functional roles. Most importantly, all existing TMD simulations exclude the extracellular and intracellular domains of both TREM2 and DAP12, which are known to influence TMD dynamics in other type I receptors systems 102 , 103 . Since TMD conformational changes can propagate to reposition ITAM motifs and thereby modulate Syk recruitment and activation, connecting ECD ligand-binding events to intracellular signaling outcomes will require simulations that span the full-length receptor. As summarized in Table 3 , the current picture provides a mechanistic foundation for understanding how K186 and its interaction partners govern TREM2–DAP12 complex assembly, but translating this into a structure-to-signaling framework capable of guiding therapeutic design will depend critically on addressing these remaining gaps.

Molecular Dynamics (MD) simulations are a widely adopted tool for studying biological macromolecules, providing atomistic insights into their structure, dynamics, and interactions. By numerically integrating Newton’s equations of motion using empirically parameterized force fields, MD simulations generate time-resolved trajectories that capture conformational fluctuations, biochemical mechanisms underlying behavior at protein-ligand interfaces, and associated energetic landscapes—insights that are often difficult or even impossible to obtain from traditional static experimental methods alone. As such, MD has become an essential complement to experimental approaches for understanding biological function, probing the effects of mutations, and informing the rational design and engineering of therapeutic candidates. In practice, MD simulations are initiated from a static starting structure, which may be derived from experimental techniques such as X-ray crystallography or nuclear magnetic resonance (NMR) spectroscopy, or predicted computationally using structure-prediction tools such as AlphaFold 24 . Experimentally-determined 3D structures are archived on the RCSB Protein Data Bank (PDB). For multicomponent systems, including protein–protein or protein–ligand complexes, molecular docking methods are often used to generate plausible starting configurations for MD by identifying energetically favorable binding orientations based on electrostatic and geometric complementarity. Subsequent, physics-based MD simulations are critical for refining these initial docked models, correcting local structural artifacts, and moving beyond static representations to capture the dynamic behavior of the system. MD simulations are typically performed using established simulation engines such as GROMACS 25 , LAMMPS 26 , or DESMOND 27 . Within these frameworks, an appropriate force field is selected to define the functional forms and parameters governing interatomic interactions. Force-field choice is a key determinant of simulation accuracy, as most widely used force fields are optimized for specific biomolecular classes and thus may not reliably describe systems outside their intended scope. Standard MD workflows generally include an initial energy minimization (EM) step, followed by equilibration under constant temperature and pressure, and a production phase—most often conducted in the NPT ensemble—during which statistically meaningful data are collected. For fully atomistic simulations of biomolecules, accessible timescales typically extend up to several microseconds, with continued advances in high-performance computing and sampling strategies increasingly enabling exploration of longer-timescale biological processes. MD trajectories can be analyzed using both qualitative and quantitative approaches to extract structural and dynamic information ( Table 1 ). Software tools such as Visual Molecular Dynamics 28 allow trajectories to be inspected frame-by-frame to visualize molecular motion, while standard quantitative metrics, such as root-mean-square deviation (RMSD) and root-mean-square fluctuation (RMSF), enable robust assessment of conformational stability and flexibility. More specialized analyses, including interatomic distance measurements, clustering, correlated motion analyses, and ligand occupancy calculations, can be used to interrogate system-specific behaviors. In addition, free-energy estimation approaches can provide insight into relative binding affinities and energetic trends, albeit with important trade-offs between computational cost and accuracy. Together, these analyses allow MD simulations to move beyond visual inspection toward statistically grounded mechanistic interpretation. Against this methodological backdrop, an increasing number of studies 21 , 29 - 31 have leveraged MD—often in tandem with experimental approaches—to probe TREM2 biology. These efforts include foundational analyses on AD- and NHD-associated mutations and their effects on TREM2 structure, dynamics, and ligand recognition, as well as investigations of TREM2 interactions with other proteins and the structural features underlying receptor function. Given the expanding MD-based literature on TREM2 and its growing therapeutic relevance, this Review synthesizes recent in silico-derived findings, critically evaluates them within a broader biological context, proposes best practices for future simulation studies, and highlights remaining gaps in the field.

MD simulations are now guiding the rational design of TREM2-targeted therapeutics at the forefront of neurodegenerative disease research. Although TREM2 has been widely studied, key molecular mechanisms linking its structure to ligand binding and downstream signaling remain incompletely understood. Increasing computational power and machine-learning-based structure prediction tools such as AlphaFold have accelerated progress—indeed, the number of TREM2 simulation studies has doubled in the past two years compared with 2023. To date, simulations have explored both the structural effects of point mutations and TREM2’s interactions with diverse ligands, including small molecules, lipids, large proteins such as ApoE, and even self-association through multimerization. Structure-based studies confirm that TREM2 contains multiple binding regions with distinct conformations and biochemical properties that support broad ligand recognition. Across these, the CDR loops—particularly CDR2, which overlaps with the PLIR and hydrophobic tip—emerge as crucial determinants of ligand accessibility and specificity. Simulations show that LOF variants like TREM2 R47H can lock CDR2 into specific states (e.g., a more open conformation in the case of TREM2 R47H ), reducing flexibility and conformational sampling and impairing TREM2’s ability to adapt to diverse ligands. Collectively, mutation studies suggest that a given residue’s impact on binding depends more on its structural context and spatial location within TREM2 than on the associated disease type. Advances in simulation methodologies have also enabled increasingly realistic models of TREM2, including studies of cleaved, canonical sTREM2 and of the TMD embedded in lipid bilayers that mimic cellular membranes. Together with investigations of the Ig-like domain, these efforts have substantially advanced understanding of how TREM2 initiates downstream signaling upon ligand binding. However, major questions remain. Integrating findings across these domains will be essential to uncover how different ligands promote distinct signaling outcomes—insights that will ultimately support the development of therapeutics capable of selectively tuning tissue-specific macrophage responses. Future simulations of small molecules and peptides bound to full-length, membrane-embedded TREM2 will likely drive the next stage of drug discovery and design, not only for TREM2-related neurodegenerative diseases such as AD and NHD, but also for broader TREM2-linked conditions including cancer, metabolic disorders, and chronic inflammation.

TREM2 therapeutic development is further complicated by AD subtype, as early-onset (EOAD) and late-onset (LOAD) forms may involve distinct triggers and kinetics of TREM2 activation. While it is well established that TREM2 signaling drives the microglial transition from homeostatic to DAM phenotypes 157 , the precise spatiotemporal cues that regulate this process remain incompletely understood. Although no experimental or computational studies have directly examined whether TREM2 WT behavior differs mechanistically between EOAD and LOAD, existing evidence suggests that the surrounding molecular environment (e.g., fibrillar Aβ plaques 37 vs. metabolic and inflammatory stressors 58 , 157 ) modulates microglial responses. These temporal and molecular-context differences suggest that the efficacy of TREM2-targeted drugs may depend strongly on disease subtype, underscoring the need for stage- and context-specific therapeutic design and dosing strategies. The development of translatable, high-efficacy drugs remains a formidable challenge, as evidenced by the billions of dollars invested annually by the pharmaceutical industry. A TREM2-targeted therapeutic occupies a dauntingly expansive yet opportunity-rich design space, with multiple mechanisms through which TREM2 activity could be modulated. For example, a drug designed to bind the stalk domain could either recruit ADAM10/17 to promote cleavage or block the metalloproteinase domain to inhibit it, thereby altering sTREM2:TREM2 ratios. Beyond cleavage, the contribution of the flexible stalk domain to the propagation of extracellular ligand sensing into intracellular signaling remains uncertain. One hypothesis, which has precedence throughout biology, is allosteric propagation 163 , in which ligand binding induces conformational changes that transmit across the unstructured stalk region to generate differential signaling outputs. Designing stalk-binding small molecules or peptides that shift the sampling of stalk conformations associated with either allosteric propagation (activation) or allosteric inhibition could therefore represent a promising strategy for modulating TREM2 activity. Alternatively, ligands that agonize or antagonize binding sites on the Ig-like domain could influence TREM2 signaling more directly, though their design depends on both the kinetics and thermodynamics of the receptor-ligand complex. MD simulations can play a central role in this design process by identifying conformational states of the Ig-like domain that correlate with distinct functional outcomes (e.g., states with altered CDR2 flexibility/rigidity or PLIR accessibility). Several additional approaches could be leveraged to modulate TREM2 signaling: modulation of the TREM2 C-terminal fragment, which enables signaling activation without direct engagement of ligands with Ig-like domain 164 ; promoting or inhibiting multimer formation; or modulating transmembrane salt-bridge formation with DAP12 (see Sections 5 and 6 for further discussion). Collectively, these mechanisms highlight the diverse strategies through which TREM2 activity could be engineered across structural domains and pathological contexts. Among the most immediately actionable therapeutic insights to emerge from TREM2 simulations is the consistent identification of W44—located at CDR1 within the hydrophobic tip—as a recurrent small-molecule and lipid binding hotspot 52 , 53 , 107 . This residue occupies a structurally central position at the core of TREM2's ligand-recognition surface, and its spatial proximity to R47 makes it a mechanistically compelling anchor for compounds designed to modulate CDR2 conformational behavior. The simulation-derived evidence that some small-molecule binders at this site induce CDR2 fluctuations 53 resembling those seen in TREM2 R47H serves as an important cautionary note: ligand engagement at W44 is not inherently therapeutic, and the functional outcome will depend critically on whether binding stabilizes a closed, WT-like CDR2 state or displaces the loop toward an open, disease-associated geometry. Future drug-design campaigns targeting the Ig-like domain should therefore prioritize not just binding affinity at W44 but the conformational consequence of binding—explicitly screening for compounds that anchor to W44 while constraining CDR2 in the closed conformation. MD simulations are well-positioned to drive this selection, through long-timescale, replicated trajectories of candidate compounds paired with quantitative CDR2 state classification along the CDR1–CDR2 distance coordinate. Finally, a growing body of literature is exploring the therapeutic potential of TREM2-expressing tissue-associated macrophages across a wide range of immunometabolic diseases, including cancer, osteoporosis, and atherosclerosis. As discussed throughout this review, regulation of TREM2 activity is highly nuanced, with the consequences of its activation or inhibition being strongly tissue-, context-, and disease-dependent. For example, modulating TREM2 activity can remodel the immune microenvironment surrounding tumors: TREM2 inhibition or knockdown has been shown to enhance antitumor immunity, supporting the ongoing development of TREM2-targeting antibodies and small molecules in oncology 165 - 168 . TREM2 also plays a critical role in osteoclast differentiation and bone remodeling, making it a potential target for inhibition in the prevention of osteoporosis and other bone-loss disorders 169 , 170 . Likewise, in atherosclerosis, TREM2 promotes lipid uptake by cholesterol-laden “foamy” macrophages within arterial plaques, suggesting that its inhibition may mitigate atherosclerotic plaque accumulation and progression 171 . Beyond these examples, TREM2 dysregulation contributes to inflammatory and metabolic dysfunction in conditions such as inflammatory bowel disease, endometriosis, and fibrotic liver disease 158 , 172 , 173 . Collectively, these findings emphasize that effective TREM2 agonist and antagonist development must be highly contextual and tissue-specific. Achieving this precision will depend on a detailed molecular-level understanding of TREM2’s conformational dynamics—insights that MD simulations are uniquely positioned to provide with their ability to dynamically sample conformational ensembles across ligand types. An increasingly compelling direction in TREM2-targeted drug development is the design of small molecules or peptides tailored to specific TREM2 variants. Such variant-focused therapeutics could enable a form of personalized medicine, guided by genetic testing, to improve efficacy in individuals carrying TREM2 LOF mutations that exacerbate neurodegenerative or metabolic disease progression. Two general design principles emerge from this approach. The first involves designing therapeutics that selectively target either TREM2 WT or specific variants, depending on the desired modulation of TREM2-mediated signaling. This strategy is particularly relevant in individuals with heterozygous TREM2 expression, where selective engagement of the “healthy” TREM2 WT (or WT-like) allele could restore—or compensate for—lost signaling capacity resulting from the expressed disease-associated allele. Conversely, in cases where overactivation of TREM2-mediated signaling may be detrimental, variant-specific inhibition could be therapeutically beneficial. Rational structure-based design of selectively tuned ligands for WT versus variant protein forms—a foundational application of MD simulations—could also serve as a powerful means to dissect structure-function relationships underlying distinct TREM2-mediated signaling outcomes in both health and disease. One prominent example of this concerns the TREM2 R47H –ApoE4 interaction. Lietzke et al. 31 observed synergistic interfacial rigidification when TREM2 R47H and ApoE4 were simulated together—a structural phenotype with no parallel in the TREM2 WT –ApoE4 or TREM2 R47H –ApoE2/3 combinations. This finding suggests that individuals carrying both risk alleles may experience a qualitatively distinct receptor–ligand interaction that compounds the effects of either variant alone, potentially explaining why their co-occurrence synergistically elevates AD risk 131 . From a therapeutic standpoint, this raises the possibility that a single-agent strategy—either restoring TREM2 CDR2 flexibility or reducing ApoE4's contribution to interfacial rigidity—may be insufficient in double-risk carriers, and that dual-targeting approaches warrant exploration. While speculative, this hypothesis is directly testable in simulation by examining whether compounds that restore CDR2 plasticity to TREM2 R47H can also disrupt the synergistic rigidification observed in the R47H–ApoE4 complex. The second design principle centers on identifying shared structural vulnerabilities or biochemical motifs among pathogenic TREM2 variants that can be therapeutically targeted in a broader, cross-variant manner. Rather than allele-specific modulation, this approach aims to restore key conformational features common to TREM2 WT —particularly the geometry and electrostatic landscape of the CDR loops and PLIR—that together govern ligand accessibility and receptor plasticity. The collective simulation data suggest a loose hierarchy of variant mechanism classes that may inform this cross-variant strategy: variants that lock CDR2 into the open state (R47H) are mechanistically suited to conformational restoration approaches, where the goal is to re-enable stochastic CDR2 loop-closing; variants that primarily alter PLIR electrostatics (W50S, R52C) may be better addressed by compounds that compensate for the changed surface chemistry without requiring large-scale CDR2 repositioning; and variants that cause broad structural destabilization or misfolding (T66M, V126G, Y38C) may require pharmacological chaperone strategies acting upstream of the mature receptor surface rather than at the ligand-binding interface itself. This framework, while necessarily provisional given the limited number of MD studies for most variants, illustrates how simulation-derived mechanistic classifications can guide the rational matching of therapeutic modality to variant-specific pathophysiology—a principle that will become more powerful as the variant simulation dataset expands. Prior studies analyzing multiple missense variants suggest that the severity of a mutation’s impact on TREM2 function reflects not only its biochemical nature (and consequent structural effects) but also its spatial location relative to functional sites within the ECD. Yet most variants other than TREM2 R47H have been examined in only 1-2 MD studies to date, limiting the generalizability of this principle, which likely does not act in isolation in any case. Nonetheless, it remains an informative organizing idea. For example, multiple simulations 31 , 49 of TREM2 R47H propose that it limits CDR2 loop “openness” and rigidifies the Ig-like domain, reducing TREM2’s capacity to adapt to ligands of different size, structure, and chemistry. Variants within or adjacent to CDR2, particularly those that overlap the PLIR/Surface 1, may therefore produce greater LOF than variants distal to primary ECD binding regions. The V126 mutation, for instance, overlaps instead with Surface 2 found on sTREM2 and thus largely does not perturb canonical ligand-binding sites, likely allowing it to retain partial functionality. A useful counterexample is the R47C mutation, observed in FTD and potentially AD 174 , 175 . Despite sharing the same spatial location as R47H, substituting arginine with cysteine produces little change in TREM2 function relative to the pathogenic outcomes associated with the arginine-to-histidine substitution. Rather, TREM2 R47C ’s pathogenicity may be more directly related to altered surface expression 175 , instead of changes in binding affinity. This distinction is therefore biochemical rather than spatial in nature: the histidine in R47H can engage in π-π stacking interactions with nearby CDR2 residues, altering local H-bonding networks compared to the native arginine 18 , 31 , whereas the cysteine in R47C lacks an aromatic ring and cannot support such stacking, reducing the likelihood of disruptive rearrangements in this region. The broader lesson is that location alone is insufficient to broadly predict mutational effects. Biochemical factors such as side-chain chemistry (charge state, protonation, aromaticity, H-bonding and salt-bridge propensity, steric bulk, etc.) and local structural context together likely determine functional outcomes. MD simulations are uniquely suited to disentangle these effects by quantifying how specific mutations reshape TREM2’s CDR2 structural ensemble, PLIR accessibility, and interfacial electrostatics. Integrating these datasets with machine learning could reveal generalizable principles for variant-aware therapeutic design—for instance, small molecules or peptides that restore flexible, WT-like CDR2 loop conformational sampling or stabilize productive ligand-binding poses. As such, to comprehensively understand the effects of missense variants, altered binding, and domain interactions within TREM2 itself, and in the presence of potential therapeutics, rigorous MD simulations utilizing best practices must be employed.

Triggering receptor expressed on myeloid cells 2 (TREM2) is a single-pass, type I transmembrane receptor primarily expressed by myeloid cells throughout the body, including macrophages, dendritic cells, osteoclasts, neutrophils, and microglia 1 . TREM2 is constitutively expressed by tissue-resident macrophages, most notably brain-resident microglia, where it plays a central role in lipid metabolism and microglial activation. Numerous TREM2 variants alter receptor function and increase risk for neurodegenerative diseases, including Alzheimer’s disease (AD) 1 - 3 , frontotemporal dementia (FTD) 4 - 6 , and Nasu-Hakola disease (NHD) 7 . As a result, TREM2 has emerged as a promising therapeutic target in recent drug development efforts. Accordingly, a growing number of TREM2-targeting strategies have been designed to modulate or rescue receptor function 8 - 10 , enhance signaling 11 - 13 , or regulate expression levels 14 , 15 . Notably, several of these approaches rely on antibodies 16 , 17 that engage extracellular regions of TREM2 11 , 13 , underscoring the importance of a detailed understanding of TREM2’s structure-function relationships across both homeostatic and disease contexts. Experimental structural studies have yielded substantial static information on TREM2, with 15 structures collectively defining its extracellular domain (ECD) and membrane-proximal architecture, including six structures of its immunoglobulin (Ig)-like domain with ligands 13 , 18 - 20 and three without ligands 18 , 21 , three of its stalk fragments 20 , 22 , and three of its transmembrane domain (TMD) 23 . Despite this wealth of structural data, key dynamic aspects of TREM2’s behavior remain poorly understood—particularly how its ECD engages diverse ligands, how disease-associated mutations reshape binding interfaces, and how these processes ultimately influence downstream signaling. Traditional experimental techniques for structure determination, such as X-ray crystallography and nuclear magnetic resonance (NMR), are inherently limited in their ability to capture conformational dynamics at atomistic resolution, which is crucial to fully elucidate the function of TREM2. Computational approaches, particularly molecular dynamics (MD) simulations, therefore provide a complementary framework for probing TREM2’s structure, dynamics, and interactions with endogenous ligands, small molecules, coreceptors, and cell membranes. As computational power and methodological sophistication have advanced, MD simulations have become increasingly valuable for uncovering molecular mechanisms governing TREM2’s behavior in both health and disease. With over 20 TREM2-focused MD studies now published, the field has reached a critical point at which a comprehensive synthesis is needed to assess how computational findings align with experimental data, reveal emergent principles of TREM2 dynamics, and establish rigor and best practices for simulation design and analysis to inform next-generation TREM2 studies. In this review, we summarize and rigorously assess key findings from recently published TREM2-focused MD studies. We begin with a brief overview of MD terminology and commonly employed analyses, followed by insights into the structure and dynamics of TREM2’s ECD in both wild-type and disease-associated variants. We then review studies of more complex TREM2 states, including soluble, membrane-bound, and multimeric forms, as well as TREM2–ligand interactions involving proteins, antibodies, small molecules, and lipids. Throughout, we emphasize the importance of rigorous simulation design—including sufficient replicates and timescales—to ensure reliable conclusions. Finally, we integrate mechanistic insights across these studies and discuss their implications for the rational engineering of TREM2-targeted therapeutics.

As highlighted throughout this review, the methods used in simulations and post-simulation analyses strongly influence molecular-level outcomes. Transparent reporting of software, parameters, and analysis pipelines is therefore essential. Numerous specifications—such as cutoff distances, integration settings, and convergence criteria—can vary across in silico workflows; however, our focus here is on assessing experimental reproducibility and adequate simulation length. To facilitate cross-study comparison, Supplementary Table 1 includes key methodological parameters from the papers discussed above. These include software specifications (docking server, MD engine, force field), reproducibility measures (production sampling time, number of replicates, details of individual component simulations prior to protein-ligand docking), and descriptions of simulated systems (PDB codes, residues modeled, and variant identifiers). To improve consistency, reproducibility, and rigor across TREM2 simulation studies, several best practices are recommended. First, with modern computational efficiency and resource accessibility, simulations should sample on long timescales of at least 1 μs, since multiple studies report substantial changes in CDR2 behavior only after approximately 500 ns. Importantly, extended sampling does not replace the need for replicates; multiple independent trajectories—ideally initiated from distinct configurational regions of the underlying free-energy landscape—remain essential for statistically distinguishing genuine structural or dynamical transitions from stochastic fluctuations. Second, for TREM2-ligand complexes, both receptor and ligand should undergo full MD equilibration (following energy minimization) before molecular docking. This ensures that docked conformations reflect biologically relevant interfaces rather than artifacts from crystallization or structure-prediction algorithms. Third, trajectory analyses would be greatly strengthened—and more easily validated—if trajectory snapshots were explicitly referenced in the text/figures and, ideally, if compressed trajectories were deposited in public repositories. Visual observations should be corroborated by quantifiable metrics derived directly from the trajectory. Finally, as in many computational disciplines, limited methodological reproducibility constrains collective progress. Transparent collaboration and detailed methodological reporting—including input structures, parameter files, and analysis scripts—are essential to enable cumulative advances in understanding TREM2 structure, dynamics, and function. Considering the limitations of experimental methods to capture TREM2 structures and computational tools, like AlphaFold 24 , that can predict structure from sequence, transparency surrounding starting structures is especially important. With adoption of these best practices, in silico studies will be more directly comparable, accelerating the development of well-validated, predictive, and mechanistically grounded models of TREM2 behavior in both health and disease.

Structurally, TREM2 contains an extracellular domain (ECD; 19–130 amino acids (aa)) that mediates ligand binding, an extracellular stalk (131–174 aa), a transmembrane domain (TMD; 175–195 aa), and an intracellular cytosolic tail (196–230 aa) ( Fig. 1A ). TREM2’s ECD is its most extensively studied region due to its capacity to engage diverse ligands and its resulting promise for targeted binding therapies 32 . The immunoglobulin (Ig)-like ECD contains three complementarity-determining regions (CDRs) that mediate ligand interactions with high specificity 18 , 21 . The conformations of these CDRs influence the accessibility of experimentally observed TREM2 ligand-binding patches, including a hydrophobic tip and a putative ligand-interacting region (PLIR), also referred to as “Surface 1” 33 ( Fig. 1B ). Among the CDRs, CDR2 is believed to play a dominant role in ligand binding 18 , 21 , as PLIR availability is highly dependent on displacement of the CDR2 loop away from adjacent residues of the hydrophobic tip. Owing to the variable shape and chemistry of these regions, TREM2 is capable of binding to a broad range of anionic and lipophilic ligands, including apolipoproteins 34 - 36 , amyloid beta (Aβ) 37 - 39 , anionic and zwitterionic lipids 34 , 40 - 42 , nucleic acids 43 , and proteoglycans 21 , 44 . The diversity of these ligands reflects the broad and essential roles TREM2 plays in microglial function. Because many of these ligands participate in lipid metabolism and the regulation of inflammatory and anti-inflammatory pathways, TREM2 is critical for maintaining microglial homeostasis and appropriate immune responses. The prevalence of neurodegenerative disease-associated mutations within TREM2’s ECD further underscores the importance of this receptor in brain health ( Fig. 1C ). The most extensively studied of these mutations is TREM2 R47H , one of the strongest known genetic risk factors for late-onset AD 45 . This point mutation occurs at the tail end of CDR1 but is spatially positioned between CDR1 and CDR2 at the interface of the hydrophobic tip and PLIR, markedly reducing ligand binding 18 , 21 , 29 , 34 - 36 , 40 . While numerous disease-associated mutations have been identified, TREM2 R47H illustrates how relatively subtle changes in local chemistry and regional conformation can produce loss-of-function (LOF) TREM2 hypomorphs ( Fig. 1D - E ). Detailed studies of additional mutations at comparable resolution will be critical for determining more broadly how distinct chemical perturbations alter TREM2 structure and function, and whether they induce conformational states associated with pathogenicity. Such mechanistic insights may ultimately inform therapeutic strategies aimed at rescuing activity in LOF variants expressed in both the brain and peripheral tissues. Before discussing simulation findings, it is important to define three related but conceptually distinct terms used throughout this section. Stability refers to the maintenance of overall protein fold, typically assessed by RMSD convergence and the absence of large structural deviations over a trajectory. Flexibility refers to the magnitude of positional fluctuations of specific residues or regions, as quantified by RMSF. Conformational plasticity , as used here, refers specifically to the ability of a region to reversibly sample multiple distinct functional states—for CDR2, the capacity to transition between open and closed loop conformations. These distinctions matter: a protein region can be flexible (high RMSF) without being plastic. As discussed below, TREM2 R47H exemplifies this: its CDR2 exhibits elevated RMSF relative to WT yet is less plastic, because it cannot reversibly transition between loop conformational states. This distinction is central to understanding R47H pathogenicity and to interpreting MD findings across variant studies. It is also important to note that all experimentally resolved TREM2 crystal structures used as starting points for simulation (including the PDB 5ELI 21 and 5UD7 18 structures) were generated using TREM2 expressed in mammalian cells and therefore originally contain N-linked glycosylation at residues N20 and N79 ( Fig. 1C , E ). The notable exception is the TREM2 R47H structure (PDB 5UD8 18 ), which was produced in E. coli and is therefore aglycosylated. This difference may contribute to the distinct CDR2 conformation observed in 5UD8 relative to mammalian-expressed (wild-type) TREM2 WT structures, and should be considered an additional caveat when using 5UD8 as a starting structure for MD simulations of TREM2 R47H . The significance of glycosylation for TREM2 dynamics is discussed further in Section 10 . Despite these differences though, none of the simulations discussed in this review actually incorporated N-linked glycans (e.g., removing them in the starting crystal structure prior to energy minimization), but their potential effect on the starting conformation of TREM2 should be considered when evaluating the conformational landscapes accessed across studies. The conformational dynamics of CDR2 in TREM2 WT have emerged as the central organizing theme across multiple independent simulation studies, revealing a picture of the wild-type receptor as an inherently dynamic entity capable of sampling rare, functionally relevant conformational states that are inaccessible on short timescales. Early foundational work established the baseline dynamic properties of the CDR2 loop, while more recent microsecond-scale studies revealed its capacity for stochastic loop-opening—a behavior previously considered pathognomonic of disease-associated variants. The first MD simulations of TREM2 were published in 2014 by Abduljaleel et al. 46 , using a structure prediction model in the absence of any experimentally resolved structure. While novel at the time, these results have limited interpretability in the context of subsequent structural and computational advances and are primarily of historical interest. The first crystal structure of TREM2's Ig-like domain (PDB 5ELI) was resolved by Kober et al. (2016) 21 , followed in 2018 by the hexameric structure 5UD7 from Sudom et al. 18 , enabling the structure-based MD simulations that form the backbone of subsequent work. Dean et al. (2019) 30 performed the first MD study of TREM2 using an experimentally resolved structure (5UD7), conducting 250 ns simulations of the TREM2 WT Ig-like domain without replicates. Despite the absence of replicates limiting statistical confidence, this study established that TREM2 WT maintains relative conformational stability across its Ig-like domain (lower RMSF relative to tested variants) and noted reduced α-helical character within CDR2 compared to variants. Dash et al. (2020) 47 subsequently corroborated these findings using the 5ELI structure across 100 ns simulations (no replicates), similarly identifying TREM2 WT as more stable than tested NHD-risk variants and characterizing intramolecular hydrogen-bonding networks within CDR1 and CDR2 that stabilize the CDR loops. Together, these two studies established a consistent picture: CDR2 in TREM2 WT is less α-helical and less rigid than in disease-associated variants, and this reduced helicity is associated with a more dynamic, fluctuating loop. The subsequent Dash et al. (2022) 48 study reproduced the WT baseline using a longer 500 ns simulation with triplicate replicates (using 5UD7), though it reported a discrepancy from the 2020 work: while Dash et al. (2020) 47 observed reduced α-helicity in TREM2 WT CDR2 relative to tested variants, the 2022 48 study found the opposite. This inconsistency most likely reflects differences in the specific variant sets compared in each study (Y38C, W50C, T66M, V126G in 2020 vs. W50S, R52C, D104G in 2022), the absence of replicates in the earlier work amplifying stochastic variation, or both. Importantly, neither study's conclusions about variant-induced CDR changes are invalidated by this discrepancy—rather, it reinforces that CDR2 helicity changes are mutation-dependent rather than universal, a point discussed further in Section 3.3 . The most significant advance in understanding WT CDR2 dynamics came from microsecond-scale, multi-replicate simulations by Saeb et al. 49 and Lietzke et al. 31 . Saeb et al. 49 employed six independent 1 μs simulations using an AlphaFold-predicted TREM2 structure, while Lietzke et al. 31 used the experimentally derived 5ELI structure with three replicates. In two of six Saeb et al. 49 replicates, a transition of the CDR2 loop was observed around 500 ns, shifting from an initial "closed" conformation to a more "open" state operationally defined by an increased inter-residue distance between positions 45 (CDR1) and 70 (CDR2) ( Fig. 2 ). During this transition, the CDR2 loop shifted away from the hydrophobic tip, sampling a distinct spatial configuration characterized by elevated RMSF values and increased CDR1–CDR2 separation. Lietzke et al. 31 subsequently reproduced this stochastic CDR2 loop-opening behavior in TREM2 WT using a different starting structure, reinforcing both the reproducibility and likely biological relevance of the observation. Critically, in open conformations, CDR2 of TREM2 WT retained little—and in one replicate, no—α-helical character, consistent with the earlier findings from Dean et al. (2019) 30 and Dash et al. (2020) 47 regarding reduced CDR2 helicity in WT. Those studies further proposed that CDR1–CDR2 hydrogen-bonding, present only when α-helical structure is retained, contributes to overall Ig-like domain stability—a mechanism that explains why partial loop-opening reduces helicity and simultaneously loosens the CDR1–CDR2 interaction network. The occurrence of these events only after ~500 ns—and in a minority of replicates—has two important implications. First, it demonstrates that capturing CDR2 loop-opening in TREM2 WT requires both long timescales and sufficient replicate sampling to distinguish genuine conformational transitions from stochastic noise; short simulations without replicates will systematically miss this behavior. Second, the stochastic, low-frequency nature of loop-opening in WT (vs. its constitutive nature in R47H, discussed in Section 3.2 ) is itself functionally informative, as discussed below. Collectively, these findings establish that TREM2 WT 's CDR2 loop does not occupy a single fixed conformation but instead samples a dynamic ensemble that includes both closed and open states. This conformational plasticity—the ability to reversibly transition between states rather than remain trapped in one—may be critical for TREM2 WT 's ability to engage structurally and chemically diverse ligands under physiological conditions, as discussed further in Section 3.2 and Section 8 . Both Saeb et al. 49 and Lietzke et al. 31 relied primarily on qualitative visual inspection of trajectories to characterize these transitions, highlighting the need for future studies employing quantitative state-classification metrics (e.g., clustering, free-energy surfaces along the CDR1–CDR2 distance coordinate) and biochemical or structural validation to independently confirm the physiological relevance of these conformational states. Finally, several ligand-focused studies that include TREM2 WT simulations gave limited attention to the intrinsic dynamics of the unbound receptor—an important context for interpreting ligand-induced effects. Alrouji et al. 50 commented only briefly on Ig-like domain stability in ligand-free simulations. Greer et al. 51 and Dean et al. (2026) 52 simulated the unbound receptor only to confirm CDR stability prior to docking. Mishra et al. (2025) 53 identified CDR2 conformational perturbations upon small-molecule binding but did not provide a systematic comparison to the unbound TREM2 ensemble. As ligand-bound simulations become increasingly prevalent, rigorous characterization of unbound TREM2 dynamics will be essential for properly contextualizing ligand-induced conformational changes. TREM2 R47H has become a paradigmatic example of how a single missense mutation can fundamentally restructure the conformational ensemble of a ligand-binding domain—not by destabilizing the fold globally, but by selectively eliminating the conformational plasticity of CDR2. Across multiple simulation studies, a consistent mechanistic picture has emerged: whereas TREM2 WT dynamically samples both open and closed CDR2 loop states ( Section 3.1 ), TREM2 R47H adopts a persistently open CDR2 conformation, effectively trading dynamic adaptability for structural rigidity in the ligand-binding region. The R47H mutation, identified in 2013 via two independent population studies 1 , 2 , is one of the strongest known genetic risk factors for late-onset AD 45 . Both in vitro and in vivo studies have shown that R47H impairs TREM2-mediated microglial activity, driving dysfunctional microglial states that exacerbate AD pathology 54 - 56 . Mechanistic studies further show that R47H reduces both TREM2 ligand-binding affinity 21 , 36 and subsequent downstream signaling 57 , 58 , thereby perturbing homeostatic microglial responses 59 , 60 . The crystal structure of TREM2 R47H (PDB 5UD8), published in 2018 by Sudom et al. 18 , was the first to capture the “open” CDR2 loop conformation, in which the loop extends away from the rest of the Ig-like domain. This structure also revealed substantial α-helical character within CDR2. Structural analysis suggests that loop opening in TREM2 R47H is driven by disruption of a stabilizing hydrogen-bond network following the Arg-to-His substitution, together with π–π stacking between H47 and H67 enabled by a ~180° conformational shift of H67. Notably, CDR2 loop residues 76–81 (numbered 58–63 in the PDB) were unresolved in the 5UD8 18 structure, consistent with intrinsic disorder or enhanced flexibility in this region—these residues must be explicitly modeled via homology modeling or alignment prior to simulation. An important caveat regarding the R47H starting structure applies across all studies that used 5UD8: because this structure was produced in E. coli and is therefore aglycosylated 18 (unlike all other resolved TREM2 structures, which were expressed in mammalian cells), the open CDR2 conformation may partly reflect the absence of N-linked glycans at N20 and/or N79 rather than the mutation alone. This possibility has not yet been computationally explored and should be considered when interpreting results derived from 5UD8. Dean et al. (2019) 30 first examined TREM2 R47H in silico by introducing the mutation directly into the TREM2 WT hexameric structure (PDB 5UD7) and performing a single 250 ns simulation. The variant exhibited increased RMSD relative to WT, and elevated CDR1 fluctuations were observed; however, pronounced CDR2 perturbations were largely absent—likely reflecting incomplete sampling given the limited timescale and lack of replicates, rather than a genuine mechanistic finding. Increased α-helicity in CDR2 of TREM2 R47H was reported relative to WT, consistent with the crystallographic observation, though minimal electrostatic differences were detected at the PLIR—a finding that contrasts with experimental evidence of substantial disruption in ligand binding at this site, which is widely attributed at least partly to local electrostatic changes 18 , 21 . This contrast underscores how short, single-trajectory simulations may fail to capture the electrostatic consequences of mutation-induced conformational change. The most mechanistically informative R47H simulations were performed by Saeb et al. (2024) 49 and Lietzke et al. (2025) 31 , using the 5UD8 crystal structure with unresolved residues reconstructed via homology modeling, and conducting 1 μs simulations with six and three replicates, respectively. Both studies converged on a consistent finding: TREM2 R47H adopts a persistently open CDR2 loop conformation across all replicates, in stark contrast to the stochastic, reversible loop-opening observed for TREM2 WT . This behavioral difference is the mechanistic core of R47H pathogenicity as revealed by MD: the mutation does not simply open CDR2—it eliminates the receptor's capacity to close it. Intriguingly, despite this positional locking, Saeb et al. 49 observed greater average CDR2 RMSF for TREM2 R47H than for TREM2 WT , emphasizing the distinction between flexibility and plasticity introduced above: the CDR2 loop in R47H is locally fluctuating within the open state but unable to escape it. Lietzke et al. 31 further confirmed that TREM2 R47H retains greater α-helical content within CDR2 and maintains stable contacts between H47 and H67 consistent with π–π stacking throughout all simulations, in agreement with prior crystallographic and experimental reports 18 . The convergence of these results across two independent research groups, two different starting structures, and differing numbers of replicates strongly supports the biological relevance of the observed CDR2 dynamics. That said, a methodological note is warranted: Lietzke et al. 31 used the experimentally resolved 5UD8 structure of R47H, while Dean et al. 30 introduced the mutation into the WT structure (PDB 5UD7). These approaches capture different aspects of the mutational effect: the former reflects the conformational state actually adopted by R47H after folding (subject to the aglycosylation caveat noted above), while the latter captures the trajectory away from the WT conformation. With sufficient timescale and replicates, both approaches should converge on the same conformational ensemble—and to the extent they do, this convergence strengthens confidence in the findings. The mechanistic picture that emerges from these studies has a direct therapeutic implication: because R47H's pathogenicity arises from loss of plasticity rather than from gross structural instability, the therapeutic goal is not to stabilize a particular CDR2 conformation but to restore the receptor's capacity to transition between them. Compounds that disrupt the H47–H67 π–π stacking interaction—which appears to be the primary driver of CDR2 locking—or that provide an alternative conformational anchor allowing CDR2 to re-access closed states represent the most mechanistically grounded design hypothesis to emerge from simulation work to date. Beyond R47H, a growing body of simulation work has examined how other disease-associated missense mutations perturb TREM2 structure and dynamics ( Fig. 1C - E , Table 2 ). While each variant ultimately impairs TREM2 function, the mechanisms by which they do so are heterogeneous—reflecting both the diverse spatial locations of mutation sites and the distinct biochemical consequences of each substitution. Four mechanistic classes emerge from the collective data, though it is important to note that most variants other than R47H have been examined in only one to two simulation studies, limiting the generalizability of these groupings. Before reviewing these studies, a critical limitation applies broadly across all variant simulations, particularly those of NHD-associated mutations. All MD simulations begin from a folded TREM2 structure into which the mutation has been computationally introduced, or from a pre-existing variant crystal structure. In reality, several NHD-associated variants—including Y38C, W50C, and V126G—are believed to cause LOF primarily through misfolding, impaired ER-to-Golgi trafficking, and reduced cell surface expression, rather than through altered ligand-binding affinity at a correctly folded receptor 21 , 61 . Because MD simulations of these variants begin from a stable, folded conformation that may never be attained in vivo, the dynamics they capture reflect how a hypothetically folded version of the variant would behave—a useful structural proxy, but one that should not be conflated with the actual pathogenic mechanism, which likely operates at the level of co-translational folding and quality control rather than mature receptor function. The most striking domain-wide perturbations across all variant simulations were observed for T66M. Dean et al. (2019) 30 reported substantially increased RMSD, elevated flexibility in both CDR2 and CDR3 extending into the adjacent βC" strand (which lost β-sheet character relative to WT), and increased anti-correlated motion between CDR2 and the rest of the Ig-like domain in DCCM analysis—a pattern not observed for other variants. Electrostatic mapping indicated that these deviations were driven primarily by structural rather than electrostatic changes, and inter-CDR distances were greatest for T66M among all variants tested. Dash et al. (2020) 47 independently confirmed large conformational shifts for T66M, including elevated CDR2 fluctuations, broadly consistent with Dean et al. (2019) 30 . These findings align with experimental evidence that T66M disrupts a key hydrogen bond between T66's hydroxyl and K48's backbone amide, leading to destabilization and impaired surface transport 21 , 62 . As summarized in Table 2 , T66M thus stands out for inducing the broadest structural consequences among simulated variants—consistent with its tendency to cause folding-dependent LOF rather than a pure binding-site effect. For R62H, T96K, and N68K (Dean et al., 2019) 30 , MD simulations revealed modest increases in CDR flexibility and decreased fluctuations in β-sheets adjacent to CDR2, without the large RMSD changes or CDR uncoupling observed for T66M. N68K, used as a predicted-benign control, behaved most similarly to WT, providing an important internal validation of the simulation approach—if the benign variant produced behavior indistinguishable from disease-associated variants, confidence in the methodology's discriminatory power would be substantially reduced. For Y38C (Dash et al., 2020) 47 , RMSD indicated behavior broadly similar to WT, with selectively elevated CDR1 flexibility. Collectively, these variants appear to perturb TREM2 function through more subtle, localized effects on CDR dynamics rather than global destabilization—a distinction relevant to predicting their relative clinical severity. Dash et al. (2022) 48 examined W50S, R52C, and D104G—computationally predicted high-risk variants—in 500 ns triplicate simulations. All three exhibited increased RMSD relative to WT. Most notably, W50S and R52C displayed increased electrostatic potential at the PLIR, whereas D104G did not. R52C additionally showed a substantial increase in CDR2 3 10 helicity—a secondary structure change without a clear parallel in other variants. The observation of PLIR electrostatic changes for W50S and R52C contrasts with Dean et al. (2019) 30 , who attributed variant instability primarily to structural rather than electrostatic effects. This discrepancy is unlikely to reflect a methodological artifact given that Dash et al. (2022) 48 used longer simulations and replicates; instead, it more likely reflects genuine mechanistic diversity: variants in or beneath the PLIR (W50, R52) may directly alter PLIR surface chemistry, while variants at more distal positions (such as those studied by Dean et al. 30 ) may exert their primary effects through structural rearrangements that only indirectly perturb the PLIR. As summarized in Table 2 , this distinction suggests that mutation location relative to the PLIR partially predicts the type of functional perturbation, though biochemical factors also play a key role (see Section 8.2.1 ). V126G (Dash et al., 2020) 47 stands out for increased SASA—consistent with the exposure of a normally buried hydrophobic surface following removal of the bulky valine side chain. This mechanism is physically distinct from CDR-loop dynamics and more closely resembles a folding defect: the loss of a buried hydrophobic residue likely reduces the enthalpic driving force for maintaining the local fold, consistent with experimental evidence of V126G-induced misfolding and impaired surface expression 21 . Notably, V126G overlaps spatially with Surface 2 of sTREM2 rather than with the primary CDR-based ligand-binding sites, potentially explaining why it can cause severe LOF (via misfolding and reduced surface expression) while sparing the primary binding interface—a point with implications for variant-specific therapeutic targeting. In summary, MD simulations reveal that disease-associated missense variants in TREM2's Ig-like domain impair function through at least four mechanistically distinct routes: loss of CDR2 plasticity (R47H), broad CDR uncoupling from adjacent structural elements (T66M), subtle increases in CDR flexibility (R62H, T96K, Y38C), and direct PLIR electrostatic perturbation (W50S, R52C). As summarized in Table 2 , these mechanisms are not strictly disease-type specific—T66M, for instance, is associated with both FTD and AD, and R47H with AD alone, yet both ultimately perturb CDR2 dynamics, albeit through distinct mechanisms. This argues against the hypothesis that AD- and FTD-associated variants destabilize TREM2 through a single shared structural mechanism, as originally proposed by Dean et al. (2019) 30 , though the data remain consistent with a broader principle: that mutations proximal to CDR2 and the PLIR produce greater functional impact than those at distal positions. Resolving shared versus disease-specific pathways will require integrated in silico, in vitro, and in vivo studies examining TREM2-ligand interactions in the presence of both AD- and FTD-associated mutations across biologically relevant timescales. The mechanistic diversity documented across these variant simulations maps naturally onto distinct therapeutic strategies. Variants that lock CDR2 into the open state—R47H being the clearest example—are mechanistically suited to conformational restoration approaches, where the goal is to re-enable the stochastic loop-closing that TREM2 WT normally performs. Variants that destabilize the fold more broadly, including T66M and the misfolding-prone NHD variants, likely require intervention upstream of the mature receptor surface—pharmacological chaperones, gene therapy, or cell therapy approaches that act on TREM2 biogenesis rather than on its extracellular function. Variants that primarily alter PLIR electrostatics, such as W50S and R52C, occupy a middle ground: the fold is intact, but the binding surface chemistry is changed, suggesting that compounds designed to compensate for the altered electrostatic environment at the PLIR—rather than to reposition CDR2—may be most effective. This framework is necessarily provisional given the limited simulation data available for most variants beyond R47H, but it provides a principled starting point for matching therapeutic modality to mechanism, which is developed further in Section 8.2 .

TREM2 engages a wide range of ligands, as documented across numerous experimental studies and methods 21 , 34 - 37 , 40 , 41 , 43 , 58 , 108 - 116 . While most endogenous ligands tend to be anionic and/or lipidic, mounting evidence indicates that TREM2 can also bind an expanded array of structurally and chemically diverse partners. These interactions are thought to underlie TREM2’s ability to initiate distinct downstream signaling cascades, thereby enabling a broad spectrum of cellular functions. An important caveat that applies across all TREM2–ligand simulations discussed in this section is that endogenous TREM2 ligands are rarely monomeric in their physiological presentation. ApoE exists as part of lipoproteins; Aβ forms oligomers and fibrils; TDP-43 aggregates under pathological conditions; complement protein C1q is itself a hexameric complex; IL-34 is a homodimer; and even phospholipids are likely presented to TREM2 as components of membranes, membrane blebs, or micelles rather than as isolated molecules. The 1:1 TREM2:ligand models used in all current MD studies therefore represent a significant simplification of the physiological interaction—one that may systematically underestimate avidity effects driving TREM2 clustering and signaling at the cell surface. Future simulations incorporating oligomeric or membrane-embedded ligand forms will be essential for capturing the full mechanistic complexity of TREM2 ligand engagement. With this in mind, computational studies over the past five years have characterized TREM2 binding across multiple ligand classes. In this section, we discuss findings on TREM2–protein interactions, with a focus on apolipoproteins; TREM2–lipid interactions, particularly sphingo- and phospholipids; and TREM2–small-molecule interactions, highlighting computational drug-design approaches targeting TREM2's ECD. Xie et al. 116 were the first to simulate TREM2-protein interactions, investigating TREM2 binding with TAR DNA-binding protein 43 (TDP-43) using a transgenic mouse model followed by confirmatory computational experiments. TDP-43 primarily binds DNA and RNA, thereby regulating gene expression, transcription, and RNA processing 117 . While largely implicated in amyotrophic lateral sclerosis (ALS) 118 , TDP-43 is also found within insoluble aggregates in frontotemporal dementia and AD 119 . After identifying residues on human TDP-43 (hTDP-43) that bind TREM2 WT in vitro, the authors conducted simulations to mechanistically evaluate these interactions in silico. Four short peptide fragments (6-11 residues) from the low-complexity domain (LCD) of hTDP-43 were selected for testing, each named after the first three residues within their respective portion of hTDP-43 (NFG, AMM, SWG, and GFN; see Supplementary Table 1 for full names). These fragments were docked to TREM2’s Ig-like domain—between CDR2 and CDR1/3—and simulated for 316 ns with 20 replicates per fragment, representing one of the more statistically robust sampling approaches among the studies reviewed here. The GFN and NFG fragments remained bound between TREM2’s CDRs, engaging similar residues consistent with their shared chemistries. In contrast, the AMM and SWG fragments dissociated from the initial docking site and rebound to alternate positions on or near CDR2, distinct both from the other fragments and from one another. SWG bound with an interaction energy nearly double that of the other fragments, though its greater length may partly contribute to this effect. Notably, RMSD profiles were not reported, which would have provided a useful quantitative readout of complex stability and clarified whether dissociation and rebinding events were accompanied by conformational changes in the CDRs themselves. The diversity of binding sites—some spanning multiple CDRs, others localizing to CDR2 alone—raises an intriguing concept for therapeutic development. These fragments demonstrate that even short peptides with distinct chemistries can engage TREM2's CDRs in different ways, potentially tuning receptor activity. Of particular note, both NFG and GFN were observed to interact directly with the R47 residue, raising the question of whether such peptides could rescue TREM2 R47H activity by compensating for the lost arginine-mediated contacts or by physically restoring closed CDR2 loop conformations. The differing behaviors of AMM and SWG—which modulate CDR2 rather than anchor between CDRs—may also influence endogenous ligand accessibility, warranting further investigation. Dantas et al. (2025, J Mol Graph Model) 120 built on this work by simulating full-length TDP-43 in complex with TREM2 WT , providing a first look at how the intact protein engages the receptor. Using an AlphaFold-predicted TDP-43 structure, they conducted a 100 ns production simulation (no replicates) and compared RMSD profiles of TREM2 WT alone versus in complex. The complex displayed lower RMSD values, which the authors interpreted as stabilization, and on this basis ranked TDP-43 among the most stable TREM2 WT binders out of 14 proteins tested. However, as discussed in Section 7.1.4 , RMSD convergence rather than absolute magnitude is the appropriate stability metric, and reporting RMSD of the full complex—rather than TREM2 alone—makes it difficult to attribute the apparent stabilization to the receptor specifically. The RMSD profile did appear to converge early in the simulation, which makes it all the more notable that subsequent analyses of binding energies and interfacial hotspots were performed on the static pre-MD docking models rather than on the equilibrated trajectories. Together, these studies establish TDP-43 as a biologically relevant TREM2 ligand, particularly in neurological disease states where TDP-43 aggregates accumulate. Direct comparison across the two studies is complicated by the stark differences in ligand form and size—short peptides versus a full-length ~414-residue protein with distinct electrostatic and globular properties—and by the use of non-comparable analytical metrics. Whether short LCD fragments and the intact protein engage TREM2 through conserved or distinct mechanisms remains an open and important question, especially given the experimental validation from Greven et al. 104 confirming TREM2 binding to TDP-43 at the hydrophobic site. This intersection of in silico predictions and experimental confirmation provides a compelling foundation for further investigation into TREM2-mediated clearance of TDP-43 aggregates and the therapeutic potential of TDP-43-derived peptide modulators. Five studies to date have simulated TREM2-Apolipoprotein E (ApoE) interactions: Mai et al. 121 , Mishra et al. (2024) 122 , Lietzke et al. 31 , Dantas et al. (2025, J Mol Graph Model) 120 , and Greer et al. 51 . ApoE, a major lipid transport protein, forms lipoproteins that package phospholipids, triglycerides, and cholesterol for extracellular trafficking 123 . Beyond lipid transport, ApoE binds to multiple cell-surface receptors 124 - 127 , including TREM2 on microglia 34 - 36 , facilitating lipid uptake for intracellular metabolism and activating downstream signaling pathways that support diverse cellular functions. ApoE is a focal point of AD research, given that the APOE ε4 allele is the strongest genetic risk factor for late-onset AD 45 , while ApoE3 is considered neutral and ApoE2 potentially neuroprotective 128 - 130 . The co-occurrence of TREM2 R47H and ApoE4 appears to synergistically heighten AD risk 131 , suggesting that the two proteins may directly influence each other’s structural dynamics in complex. ApoE is structurally diverse 127 , 132 - 134 and can exist in both unlipidated (monomeric or aggregated 127 , 135 , 136 ) and lipidated states 137 - 141 , the latter promoting lipoprotein formation, with TREM2 shown to bind to each 29 . All five in silico studies, however, examined only unlipidated, monomeric ApoE, primarily because no high-resolution structure of lipidated ApoE has been resolved. The implications of this shared limitation are discussed at the end of this section. Table 5 provides a methodological overview of the five studies to facilitate comparison. For all simulation details, see Supplementary Table 1 . Despite substantial differences in simulation length, replication strategy, ApoE construct, and input structure preparation, all five studies converge on a consistent structural picture: TREM2's CDRs—most prominently CDR2—mediate the primary binding contact with ApoE, and ApoE's hinge region constitutes the principal interface on the ligand side. This is well-aligned with biophysical experiments from Kober et al. (2021) 29 , and its consistent recapitulation across methodologically diverse simulations provides meaningful support for the predictive power of MD in identifying conserved features of TREM2–ligand recognition. Additional experimental validation from Greven et al. 104 further confirms binding at the TREM2 hydrophobic site, reinforcing the computational consensus. Studies using full-length ApoE also identified secondary contacts outside the CDR–hinge core: Lietzke et al. 31 described contacts mediated by distinct CDR2 loop states, and Greer et al. 51 identified additional ApoE N-terminal and C-terminal contributions to the hydrophobic and basic patches on TREM2, respectively. Together, these results suggest that the CDR–hinge interface is the dominant and most reproducible contact, while peripheral interactions likely vary with ApoE conformation and TREM2 CDR2 state. Where these studies diverge most sharply is on two biologically critical questions: whether TREM2 differentially engages ApoE isoforms, and how R47H alters these interactions. On isoform selectivity, Mai et al. 121 and Lietzke et al. 31 both report stronger TREM2–ApoE4 binding relative to other isoforms, consistent with experimental observations 29 —while Greer et al. 51 found no isoform-dependent affinity differences. The most likely explanation for this discrepancy is sampling depth: Greer et al. 51 's models lack replicates for the ApoE3 and ApoE4 comparisons, making statistical discrimination of subtle affinity differences impossible, whereas Lietzke et al. 31 's replicated, long-timescale simulations provide the greatest statistical power. The observation by Lietzke et al. 31 that ApoE2 undergoes dynamic dissociation and rebinding at an alternative TREM2 site—rather than simply failing to bind—further illustrates how limited sampling can give the misleading impression that ApoE2 interactions are categorically unstable when they are in fact dynamic and site-promiscuous. On R47H effects, the studies tell a more nuanced story than a simple reduction in binding. Mai et al. 121 reported fewer H-bonds and weaker binding energies for TREM2 R47H , consistent with experimental LOF data. Lietzke et al. 31 further observed that TREM2 R47H and ApoE4 together produced synergistic rigidification at the binding interface—a finding that resonates with the observed synergistic increase in AD risk among carriers of both variants, and raises the possibility that these two mutations may directly amplify each other's structural effects in complex. Greer et al. 51 offered a different angle: rather than a reduction in affinity, R47H appeared to shift ApoE3 binding away from the CDRs and toward TREM2's multimerization sites. If confirmed with adequate replication, this would represent a mechanistically distinct LOF route—altered binding geometry rather than reduced affinity—with different implications for therapeutic rescue strategies. However, this finding rests on a single-replicate simulation and requires independent validation before it can be considered biologically meaningful. Comparing Mai et al. 121 and Lietzke et al. 31 also reveals an important methodological fork in how R47H is modeled: Mai et al. 121 introduced the mutation computationally into the TREM2 WT structure (PDB 5ELI 21 ), while Lietzke et al. 31 used the TREM2 R47H crystal structure (PDB 5UD8 18 ). As discussed in Section 3.2 , the two starting conformations differ substantially in CDR2 loop geometry, and each carries its own caveats—particularly the aglycosylated nature of 5UD8. With sufficient timescales and replicates, both approaches should converge on the same conformational ensemble, but this remains to be demonstrated directly. The most striking result from Mishra et al. (2024) 122 came from their membrane-adjacent simulations: TREM2–ApoE3 binding affinity and H-bond count both decreased when the complex was modeled on top of a POPC bilayer, and ApoE3 underwent substantial helical loss and tilting motions at the lipid interface. These observations support a model in which the Ig-like domain can transiently sample the membrane surface—mediated by the flexible stalk—and in doing so may compete with protein ligand engagement, potentially displacing ApoE under certain conditions. This remains speculative given the analytical limitations noted for this study, but it raises a broader question applicable to all large lipophilic TREM2 ligands: whether membrane proximity systematically competes with protein ligand engagement, and how this balance is resolved in vivo. A limitation shared by all five studies is the use of unlipidated, monomeric ApoE. While unlipidated ApoE accumulates under neurodegenerative conditions 142 - 146 and therefore represents a physiologically relevant species in disease, ApoE likely exists predominantly in a lipidated state in the healthy brain 137 - 141 , and experimental data indicate that lipidation alters TREM2 binding affinity 29 . In silico models suggest lipidated ApoE adopts an extended, horseshoe-like conformation within lipoproteins 147 - 153 —structurally quite different from the compact monomeric forms used in current simulations—but this conformation has not been experimentally validated at high resolution. Until it is, simulating lipidated ApoE–TREM2 interactions remains technically intractable in any rigorously validated sense. This means that despite five published studies, we still lack a mechanistic understanding of how TREM2 engages its most physiologically relevant ApoE form—a gap that is directly pertinent to therapeutic strategies targeting this interaction in the brain. Despite these challenges, the collective body of ApoE-focused simulations offers real mechanistic insight and demonstrates that MD can reliably identify conserved features of TREM2–ligand recognition even across methodologically variable studies. The field is now at a point where the next meaningful advance is not more studies of unlipidated ApoE under identical conditions, but rather efforts to push toward membrane-embedded, glycosylated, and ultimately lipidated ApoE models—guided by the convergent CDR–hinge interface that all current work points to as the structural foundation of this interaction. While most MD studies of TREM2 to date have focused on interactions with endogenous ligands such as TDP-43 and ApoE, recent work has begun to explore how therapeutic antibodies engage TREM2 at its extracellular stalk domain 11 , 13 . These studies offer a complementary perspective on how TREM2 structural flexibility enables tight molecular recognition. Hsiao et al. 22 determined the crystal structures of humanized anti-TREM2 Fab fragments bound to a short peptide segment of WT TREM2’s flexible stalk domain (aa148-165) and subsequently performed MD simulations to probe the molecular basis of differential antibody affinity. The anti-TREM2 antibodies were originally generated by immunizing rats with the TREM2 ECD; among the resulting lineages, the 3.10C2 antibody bound the stalk at the ADAM10/17 cleavage site (H157–S158), the same region targeted by humanized variants hu3.10C2 and the lineage-mined huPara.09. Notably, huPara.09 bound with significantly higher affinity than both hu3.10C2 and the parental rat antibody. To investigate this affinity difference, Hsiao et al. 22 performed 1 μs MD simulations of each Fab–TREM2 complex and of the unbound stalk peptide, in triplicate—a rigor that distinguishes this study from many others in the TREM2 MD literature. The simulations revealed that it was the TREM2 peptide rather than the antibody that differed most in conformational behavior: when bound to huPara.09, the stalk peptide sampled greater conformational diversity in its C-terminal region (residues I159–S160), resembling the range of conformations adopted by the unbound peptide. This increased local flexibility—rather than increased rigidity—underlies the higher binding affinity, because it enables the peptide to adopt more complementary geometries within the antibody-binding pocket. Site-directed mutations of the peptide (S158A, H154A) confirmed that C-terminal anchoring at I159–S160 is essential for stable binding to both antibodies, while S158 specifically stabilizes hu3.10C2 but not huPara.09. These findings extend the theme established in Sections 3.1 and 4 : TREM2's conformational flexibility, whether in the CDR2 loop of the Ig-like domain or in the stalk, is not merely a structural idiosyncrasy but a functionally exploitable property. High-affinity binding at the stalk—as with high-affinity endogenous ligand engagement—appears to depend on maintaining, rather than restricting, local conformational sampling. Because the stalk peptide in these simulations was studied in isolation from the rest of the receptor, this flexibility is likely more constrained in the full-length membrane-bound context where the stalk is tethered; this is an important caveat when translating these insights to intact receptor behavior. This work was also the first to employ MD simulations of isolated TREM2 peptides, highlighting a potential design space that has not yet been fully exploited. Stalk mimetics positioned near the H157–S158 cleavage site could, in principle, either block protease access and maintain higher surface TREM2 levels, or crosslink adjacent ECDs to promote multimerization—both potentially signaling-enhancing strategies, though the feasibility and tissue-specificity of each would require careful evaluation. Stapled or cyclic peptide designs could further bias the membrane-bound stalk toward conformations that favor intermolecular contacts between adjacent TREM2 ECDs. These remain speculative but offer concrete, testable hypotheses for future work. Dantas et al. (2025, J Mol Graph Model) 120 conducted the most expansive survey of TREM2–protein interactions to date, simulating 14 complexes in total—including the ApoE and TDP-43 results discussed above—alongside other apolipoproteins (ApoA1, ApoA2, ApoJ), cytokines (IL-4, IL-34), immune-response proteins (Cyclophilin A, C1QA), stress or signaling proteins (Galectin-3, HSP60), and disease-associated proteins (cholera toxin subunit B, SARS-CoV-2 membrane protein). All were simulated for 100 ns without replicates, and complex stability was assessed from RMSD profiles—with the analytical limitations regarding absolute RMSD values versus convergence applying throughout, as discussed previously. Cyclophilin A, Galectin-3, and C1QA yielded the lowest RMSD values and highest apparent stability; ApoA1 and HSP60 exhibited more dynamic profiles suggesting less stable interactions. Trajectory visualization of the less stable complexes would have been valuable for determining whether variability reflected ligand dissociation, conformational rearrangements at the interface, or simply large-scale protein motions elsewhere in the complex—distinctions with different mechanistic implications. Consistent with the studies discussed above, CDR2 emerged as the dominant binding site across ligands, with minor CDR1 contributions. This trend also aligned with TREM2–IL-34 work from Greven et al. 104 , which identified these interfaces in docked complexes (using equilibrated TREM2, but did not simulate said complexes). Intriguingly, only Galectin-3 and ApoJ engaged CDR3, and Galectin-3 was unique in interacting with all three CDRs simultaneously—an engagement pattern the authors linked to its relatively high apparent stability. As one of the few studies to assess TREM2 interactions across ligands spanning a broad size range (103–528 residues) and diverse functional classes, this work establishes a useful initial landscape for TREM2's protein-binding repertoire. The recurring centrality of CDR2—now evident across TDP-43 peptides, full-length TDP-43, ApoE isoforms, apolipoproteins, cytokines, and immune proteins—further reinforces its role as the primary recognition hub of the Ig-like domain. Systematically characterizing what structural and chemical features distinguish tight from weak CDR2 binders across this diverse ligand set, and relating these features to downstream signaling outcomes, represents a tractable and high-value direction for future computational work. TREM2 plays a major role in sensing lipids under both homeostatic and disease conditions, thereby regulating lipid metabolism in macrophages. Lipids are presented to TREM2 in diverse biological contexts, including when bound to ApoE and other lipoproteins 34 - 36 , complexed with Aβ 144 , 154 or myelin 40 , 58 , 155 , 156 , or as components of other damage-associated molecular patterns (DAMPs), such as lipids exposed on apoptotic neurons 40 , 114 . Lipid binding to TREM2 can promote the transition of macrophages into lipid-sensing, disease-associated states—such as disease-associated microglia (DAMs 157 ) in the brain or lipid-associated macrophages (LAMs 158 ) in peripheral tissues—which play adaptive roles in debris clearance and lipid metabolism but may contribute to pathology (e.g., altered lipid metabolism, chronic inflammation, neurodegeneration) when chronically activated 123 . Three MD studies have examined TREM2–lipid interactions, each with a distinct focus. Two—Dean et al. (2024) 105 and Dantas et al. (2025, Chem Phys Lipids) 107 —were discussed in Section 6 for their treatment of lipid-mediated multimer stability; here we focus on what they reveal about lipid-binding mechanisms themselves. The third, by Saeb et al. 49 , was introduced in Section 4 in the context of sTREM2 stalk dynamics; here we highlight its lipid-centric findings. The most mechanistically detailed lipid-binding analysis comes from Saeb et al. 49 , who systematically mapped phospholipid–TREM2 interactions using molecular docking followed by 150 ns replicate simulations of PS and PC bound in diverse poses to WT and R47H Ig-like domain and sTREM2 models. In TREM2 WT , both PS and PC bound primarily to the hydrophobic tip and PLIR/Surface 1—consistent with the crystallographically defined PS-binding sites on CDR1 and CDR2 18 , 159 —though PS showed greater occupancy of positively charged Surface 1 residues than PC, consistent with its anionic headgroup chemistry. In sTREM2, the stalk's presence reduced lipid occupancy at Surface 1 and increased it at an 'Expanded Surface 2' and along the stalk itself, suggesting that the stalk sterically or electrostatically competes with phospholipids at the primary binding face. These Surface 2 contacts align with prior experimental data on sTREM2–Aβ interactions 33 , providing an additional point of cross-study validation. The most functionally suggestive result from Saeb et al. 49 was the near-complete loss of PS/PC discrimination in TREM2 R47H : whereas WT TREM2 showed markedly different interaction patterns and binding site preferences for the two lipids, R47H models showed little-to-no difference in either binding energetics or contact patterns. This represents a distinct, lipid-centric LOF mechanism—impaired capacity to discriminate between anionic and zwitterionic lipids—that is mechanistically separate from the CDR2 loop-locking described in Section 3.2 . Whether these two deficits are independent consequences of the same structural change or are causally linked through CDR2 dynamics remains an open question. Notably, rapid CDR2 loop conformational changes were observed in several phospholipid-bound WT simulations, suggesting that lipid binding itself can influence CDR2 loop behavior—an underappreciated form of ligand-induced plasticity that may fine-tune TREM2's accessibility to other binding partners. The Dantas et al. (2025, Chem Phys Lipids) 107 docking screen across twelve lipidic ligands—PS, cardiolipin, phosphatidic acid, sphingomyelin, sphingosine-1-phosphate, sulfatide, chrysin, and additional phospholipids (PC, PG, PE, PI)—consistently identified the hydrophobic CDR1–CDR2 region as the preferred docking site, with W44 emerging as a recurrent interaction hotspot across multiple chemically distinct ligands. W44 sits at the CDR1 face of the hydrophobic tip, precisely at the interface most directly disrupted by R47H and most frequently identified as an energetically important contact in small-molecule docking (see Section 7.3 ). Its repeated appearance across diverse lipid chemistries suggests it may represent a fundamental anchor point for hydrophobic engagement at the CDR surface—a property worth exploiting in structure-based ligand design. MM-GBSA calculations ranked cardiolipin as the strongest binder and PS as the weakest, which is notable given PS's privileged role in crystallographic TREM2 multimer assembly; this discrepancy likely reflects differences between isolated protein–lipid energetics and the context-dependent, multivalent nature of PS binding in the hexameric crystal structure. Collectively, these studies establish an early computational framework for TREM2–lipid recognition, with consistent identification of the CDR1–CDR2 hydrophobic region—particularly W44—as a central contact point, and with Saeb et al. 49 providing the most mechanistically grounded picture of lipid selectivity and its loss in R47H. Key gaps remaining include longer, replicated simulations with appropriate lipid force fields, comparisons across the full diversity of TREM2-relevant lipid species, and ultimately simulations of TREM2 in a realistic membrane environment where lipid binding competes or cooperates with protein ligand engagement in real time. There is a growing list of TREM2-targeting small molecules in various stages of development, including the clinical candidate VG-3927 ( NCT06343636 159 , 160 ) and preclinical compounds such as hecubine 161 . Advancing this pipeline depends on a detailed molecular understanding of how small molecules engage TREM2's ECD—precisely the kind of insight MD is positioned to provide. Four computational studies have now addressed this, and together they reveal two pharmacologically distinct surfaces on the Ig-like domain that can be targeted independently. Table 6 provides a methodological overview of the four studies to facilitate comparison. For all simulation details, see Supplementary Table 1 . Alrouji et al. 50 screened FDA-approved antipsychotics against TREM2's CDR region, identifying carpipramine, clocapramine, and pimozide as candidate binders. Carpipramine and clocapramine engaged T85, T88, T92, and T94 with similar CDR binding modes, while pimozide formed a single H-bond with D87—a residue that participates in the R76–D87 interfacial salt bridge during multimerization ( Section 6 ), though downstream consequences of this contact were not explored. Binding affinities reported in the attomolar range are almost certainly MM-PBSA artifacts rather than genuine values; relative rankings among the compounds may nonetheless hold qualitatively. Mishra et al. (2025) 53 examined a chemically broader ligand panel against two putative binding sites identified by machine-learning pocket prediction: a CDR-proximal site centered on W44, L71–F74, S81, and T82, and a C-terminal site comprising W50, R62, V63, and H103. Per-residue MM-PBSA decomposition identified W44 as the dominant energetic hotspot at the CDR site, with predicted micromolar affinities—biophysically more plausible than the Alrouji et al. 50 values. A notable cautionary finding was that bexarotene, the highest-affinity predicted ligand, induced CDR2 fluctuations resembling those of unbound TREM2 R47H , illustrating that binding at the CDR surface does not automatically produce a therapeutically desirable conformational outcome. Dean et al. (2026) 52 addressed the CDR hydrophobic site with greater rigor, applying a rank-by-rank consensus approach across three docking programs to screen ~2,400 endogenous human metabolites, then directly validating hits by biolayer interferometry. The top endogenous hit, cholecalciferol (unhydroxylated vitamin D3) was confirmed by replicated MD simulations to engage the TREM2’s apical CDR loops through van der Waals contacts—particularly with W44, L71, F74, T88, and L89—consistent with the neutrally charged, hydrophobic character of this surface. Notably, R47H and an engineered F74D mutant both caused cholecalciferol to dissociate before simulation completion, with R47H acting allosterically rather than directly occluding the pocket. The most structurally interesting finding is that cholecalciferol does not compete with ApoE isoforms at the CDR surface; biophysical experiments showed it instead enhances TREM2–ApoE binding cooperatively across all three isoforms—an effect attenuated in R47H and L69D—suggesting that small molecules can potentiate endogenous ligand engagement at this surface rather than simply occupying it. However, these cholecalciferol–TREM2–ApoE complexes were not tested computationally, so the precise molecular basis underlying enhanced binding is yet to be determined. Such findings motivate the need to continue exploring how small molecules, especially those developed as therapeutics, may be useful in cases of altering TREM2 binding with AD-associated variants or ligands and thus downstream cellular functions. Cho et al. 162 identified an entirely different binding site using PyRod-derived pharmacophores from short MD replicas, revealing a hydrophobic cavity on the face of the Ig-like domain opposite the CDR loops, formed by β-sheets C, C', C", D, E, and G, with key hotspots at W50, L97, Y108, V63, and R52. The top screening hit from this site, EN020, confirmed selectivity for TREM2 over TREM1. Allosteric path analysis on a full-length membrane-embedded TREM2 model further identified communication pathways flanking this cavity that traverse the TMD to the intracellular tail—suggesting, speculatively, that compounds engaging this site could influence DAP12 coupling through a mechanism that does not directly compete with CDR-surface ligands. Experimental validation of this allosteric pathway remains an important next step, especially when pairing this potential mechanism with molecular-level findings from TREM2 multimers and TREM2–DAP12 interactions embedded in membranes. Across all four studies, W44 at the CDR1 face of the hydrophobic tip emerges as the most consistently implicated contact residue, appearing also in lipid docking results ( Section 7.2 ), reinforcing its role as a structurally central anchor point for hydrophobic engagement at the CDR surface. The newly described allosteric site from Cho et al. 162 , if validated, would expand the druggable landscape of the Ig-like domain considerably and raise the possibility of combining CDR-targeting and allosteric-site-targeting compounds without direct competition. Most importantly, the bexarotene finding from Mishra et al. 53 and the cooperative ApoE enhancement from Dean et al. 52 together highlight that the functional consequence of binding at W44—whether it restores or disrupts WT-like CDR2 plasticity—is as critical a design criterion as binding affinity itself.