Introduction
Phosphorylation is a fundamental mechanism by which signaling pathways are regulated in cells. Protein kinases catalyze the phosphorylation reaction by transferring the γ-phosphate from an ATP molecule to Ser, Thr, or Tyr residues of the substrate (Ahiri et al., 2019). Given their crucial role in cellular function, protein kinases are tightly regulated, and their dysregulation can lead to a variety of disorders, including cancer (Attwood et al., 2021).
The active site of a protein kinase (p-kinase) contains several structural elements that are critical for regulating its enzymatic activity: The activation segment, beginning with a conserved DFG motif (typically Asp-Phe-Gly) and extending to an APE motif (typically Ala-Pro-Glu), passing through the activation loop (A-loop); the αC-helix, which carries a conserved glutamic acid (Glu) residue (Brinkworth et al., 2003; Cowan-Jacob, 2006).
The active state of a p-kinase, leading to phosphorylation of a protein substrate by ATP, depends on these kinase elements in the following ways: The A-loop forms a cleft that binds the substrate through specific interactions with the conserved HRD motif (typically His- Arg-Asp) of the kinase catalytic loop. This configuration is designated A-loop-out (active) (Chiu et al., 2013); the loop is displaced from the substrate-binding cleft, allowing substrate access. The DFG Asp residue is positioned and oriented to coordinate a magnesium ion that interacts with the ATP phosphates. This is referred to as the active DFG-in configuration (Vijayan et al., 2015). The αC-helix, oriented inward toward the active site, positions its conserved glutamate to form a salt bridge with the lysine residue of the β3 strand. This is the αC-helix-in active configuration and is also characterized by proximity between the Cα atoms of DFG-Phe and αC-Glu (Kooistra et al., 2016). In addition, the conserved GxGxxG motif in the glycine-rich loop of the N-lobe helps stabilize the phosphates of bound ATP during catalysis (Alam, 2017; Kemp & Pearson, 1990).
Inactive kinase states lack one or more of the structural constraints required for catalysis and therefore encompass multiple conformations. In an A-loop-IN configuration, the activation loop deviates from its active trajectory and can occlude substrate binding, although the nucleotide-binding site may remain accessible (Steichen et al., 2012; Taylor & Kornev, 2011). In a DFGout configuration, DFG-Asp is displaced from the catalytically competent Mg2+-coordinating position while DFG-Phe occupies the adjacent hydrophobic pocket; this arrangement is incompatible with the canonical catalytic geometry, although ATP-site ligands may still bind (Vijayan et al., 2015). Retraction of the N-terminal end of the αC-helix disrupts the Lysβ3–GluαC salt bridge and defines the αC-helix-out configuration.
Unlike the αC-helix-in configuration, the Cα atoms of DFG-Phe and αC-Glu are not proximal in the αC-helix-out state (Huang et al., 2012; Taylor & Kornev, 2011).
Beyond these key kinase elements, the P+1 loop, located at the C-terminal end immediately after the A-loop, plays an essential role in substrate binding by interacting with the residue following the residue to be phosphorylated (the P+1 residue) (Nolen et al., 2004).
Several attempts have been made to classify the conformational states of protein kinases in the Protein Data Bank (PDB) and to study inhibitor interactions. Notably, Möbitz (2015) performed a quantitative classification of all mammalian kinases using pseudo-dihedral angles of four consecutive Cα atoms from the DFG motif and its neighbors, along with their distance to the αC-helix, resulting in 12 conformational categories. Ung et al. (2018) used a similar approach employing two directional vectors for DFG residues and the αC-helix distance, yielding five categories. These geometric schemes resolve many states, and more recent frameworks, Modi & Dunbrack (2019) and the KLIFS resource (Kooistra et al., 2016), further distinguish inactive DFG-in from active structures using spine- and distance-based criteria. Our contribution is therefore complementary rather than a replacement: We add a quantitative interaction-energy axis and a secondary-structure subclassification of the activation segment, while noting that fully automated classification of newly deposited structures remains an open goal.
Furthermore, previous classification studies suggest that conformational changes in the A-loop are related to changes in its local environment, including the αC-helix and adjacent structural elements, without fully accounting for the specific secondary-structure features of the activation segment (Cheek et al., 2002; Martin et al., 2010). Therefore, the present study classifies conformational variation in human and murine kinase structures while explicitly considering how secondary structures within the activation segment contribute to stabilization and conformational change.
Materials and methods
Preliminary categorization of activation segment configurations
An active kinase generally requires an activation segment compatible with substrate access. We classified activation segment configurations as IN, OUT, or SWAPPED. IN denotes an A-loop that occludes the substrate-binding cleft; OUT denotes an A-loop displaced from that cleft; and SWAPPED denotes a dimeric arrangement in which the activation segment of one protomer extends into or across the partner protomer and can serve as a trans-autophosphorylation substrate. In the latter arrangement, the segment is detached from its own catalytic cleft (Pike et al., 2008).
Differentiation was guided by DSSP secondary-structure assignment together with geometric criteria (proximity of the A-loop to the γ-phosphate site and its position relative to the substrate cleft). To verify that these assignments are reproducible and not idiosyncratic, the DFG and αC-helix states were cross-validated against the independent KLIFS annotations, giving 99.0% agreement for the DFG state (n=192) and 97.6% for the αC-helix state (n=208). Catalytic-motif integrity (HRD-Asp, DFG-Asp) was screened to flag pseudokinases, and activation-segment B-factors were recorded as a structure-quality descriptor. The dataset comprised human and murine typical (eukaryotic) protein kinase (ePK) structures available in the RCSB PDB: 1,699 tyrosine kinases and 2,971 serine/threonine kinases (Kouranov et al., 2006).
Conformational states were assigned by explicit, reproducible geometric criteria and cross-validated against KLIFS. The A-loop was labeled IN when any activation-loop Cα lies within 8 Å of the ATP γ-phosphate-equivalent position (occluding the substrate cleft) and OUT otherwise; the DFG state was taken from the DFG pseudo-dihedral and the Phe position relative to the R-spine/αC (KLIFS agreement 99.0%, n = 192); the αC state from the Lys(β3)– Glu(αC) salt bridge and the DFG-Phe(Cα)–αC-Glu(Cα) distance (KLIFS agreement 97.6%, n = 208). Each structure therefore carries a single, unambiguous IN/OUT/SWP label.
Association analysis between conformations (DFG and αC-helix) and the A-loop
To evaluate associations between the regulatory states of the A-loop (A-loop-IN or A-loop-OUT), DFG motif (DFG-IN or DFG-OUT), and αC-helix (αC-helix-IN or αC-helix-OUT), we constructed association diagrams for the best-resolved human and murine typical protein kinases: 104 tyrosine kinases and 110 serine/threonine kinases. The diagram counts (104 TK and 110 STK) refer to unique kinases. By contrast, N in Supplementary Table S5 denotes the eligible PDB structure observations retained for each pairwise Fisher test; multiple PDB structures may represent the same kinase, and test-specific exclusions yield N = 76 for TK and N = 122 or 121 for STK. Here, “typical” protein kinases refer to the eukaryotic protein kinase (ePK) superfamily (Manning et al., 2002), as distinct from atypical protein kinases (aPKs). A small number of divergent ePKs and catalytically active kinases bearing variant catalytic-loop or DFG motifs (e.g., Haspin) were retained when a fully resolved activation segment was present; their variant motifs are annotated in Supplementary Table S4.
Energetic classification and heat map
For quantitative analysis, we retained 68 catalytic domains with fully resolved activation segments (21 from the set of 104 tyrosine kinases and 47 from the set of 110 serine/threonine kinases). Catalytic domains were delimited using the Pfam server. This refined dataset was subjected to K-means clustering with scikit-learn (Pedregosa et al., 2011) using four criteria: (1) the mean interaction energy of the activation segment with its spatial environment in the catalytic domain (Galgonek et al., 2017), (2) the DFG conformation, (3) the αC-helix conformation, and (4) the A-loop conformation. The INTAA server was used to evaluate interaction energies with the AMBER parm03 force field (Duan et al., 2003). Because X-ray structures generally lack hydrogen atoms, hydrogens were added using Reduce to obtain an all-atom model (Word et al., 1999). Before clustering, all four features were standardized to unit variance (z-score) so that the continuous and categorical descriptors contributed equally to the distance metric. The number of clusters (K = 7) was assessed using elbow and silhouette criteria (Supplementary Figure S1) and is presented as a descriptive partition of a continuous energetic landscape rather than as a set of discrete archetypes. As a robustness check, a representative subset spanning the clusters and the active, inactive, and SWAPPED states was reexamined under an implicit-solvation model: protonation states were assigned at pH 7.4 with PROPKA (PDB2PQR), and activation-segment interactions were recomputed using a Poisson–Boltzmann continuum solvent (APBS; solute/solvent dielectric constants 2/78.5; 0.15 M ionic strength) with a nonpolar surface-area term (Supplementary Table S7).
For reproducibility, the exact clustering input is provided in full. All 68 PDB IDs, with their per-structure DFG/αC/A-loop states, activation-segment interaction energies, and cluster assignments, are listed in Supplementary Table S2B, and a per-cluster PDB roster is appended below Table A1. K-means was run on the four features standardized to unit variance (z-score) with scikit-learn (K = 7, n_init = 20, random_state = 42); the standardized input matrix is released as the machine-readable file Supplementary_Data_KMeans_zscore_matrix_R2.csv so that the z-score partition can be reproduced exactly.
From the K-means classification, representative p-kinases were selected to illustrate the principal conformational states of the activation segment (active OUT, inactive IN, and SWAPPED) sampled across the clusters. To visualize the distribution of interaction energies of activation segment residues with their spatial environment in the catalytic domain, a two-dimensional spatial heat map was produced showing the magnitude of each residue’s interaction through color, which can vary in hue and/or intensity (Galgonek et al., 2017).
Differential structural analysis (RMSD and per-residue Cα displacement)
To quantify the structural evolution between p-kinases from the classification, a differential structural analysis was performed by evaluating RMSD and the per-residue Cα displacement between superimposed crystal structures using VMD (Humphrey et al., 1996), comparing active (OUT) with inactive (IN) or SWAPPED (SWP) forms. This per-residue Cα displacement measures the coordinate difference between two static end-state structures; it is not a root-mean-square fluctuation (RMSF), which requires a molecular-dynamics trajectory or a conformational ensemble. Representative kinases from the obtained clusters were chosen to best represent the transition from IN to OUT and SWP forms.
Results and discussion
Preliminary categorization of activation segment configurations
Analysis of the kinase structures yielded three categories according to activation-segment conformation (Figure 1). IN conformations, in which the activation segment occludes the substrate-binding cleft and is generally catalytically inactive, accounted for 38% of the structures (Modi & Dunbrack, 2019). IN subclasses were distinguished by the presence of a β-sheet (18%), a helix (11%), or neither secondary-structure element within the activation segment. OUT conformations, in which the segment is displaced from the substrate-binding cleft and the catalytic motifs can adopt an active arrangement, were the most abundant (55%) (Nolen et al., 2004; Modi & Dunbrack, 2019). OUT subclasses were defined by the β10–β11 sheet (A-loop/αEF– αF-loop interaction) or the β6–β9 sheet (catalytic-loop/activation-segment interaction), present in 77% of OUT structures. SWAPPED conformations accounted for 6% of the broad survey and occurred in dimeric arrangements compatible with trans-autophosphorylation, in which the activation segment of one protomer extends toward the partner kinase (Pike et al., 2008; Beenstock et al., 2016).
Figure 1. (A) Activation segment conformations of protein kinase structures and their secondary structures and (B) proportions in the studied database shown as pie charts.
Association between DFG and αC-helix conformations and the activation segment
The central question is to situate the conformations of key kinase regulatory elements other than the A-loop, namely, the DFG motif and the αC-helix, when the A-loop is active (A-loop-OUT) or inactive (A-loop-IN). This is a two-dimensional exploration of DFG and the αC-helix, once the A-loop (and therefore activation segment) configuration is fixed.
For serine/threonine kinases, the A-loop-OUT conformation was associated with both DFG-IN and αC-helix-IN conformations (Figure 2). These configurations are consistent with proper substrate binding and stabilization in established kinase-activation models (Modi & Dunbrack, 2019; Nolen et al., 2004).
Figure 2. Association diagram for serine/threonine kinases between the A-loop-OUT conformation and the conformations of the DFG motif and αC-helix. Fisher tests for the STK dataset: A-loop × DFG, p = 0.006; A-loop × αC-helix, p = 4.34×10–9.
For serine/threonine kinases, the A-loop-IN conformation was associated primarily with αC-helix-OUT. Both DFG states occurred, although DFG-IN was more frequent (25 DFG-IN versus 19 DFG-OUT cases; Figure 3 and Supplementary Table S5). This coexistence is consistent with partial uncoupling of A-loop position from the DFG/regulatory-spine state; disruption of the regulatory spine permits displacement of DFG-Phe without imposing a single inactive geometry (Hu et al., 2015; Modi & Dunbrack, 2019).
Figure 3. Association diagram for serine/threonine kinases between the activation segment IN conformation and the conformations of the DFG motif and αC-helix. Fisher tests for the STK dataset: A-loop × DFG, p = 0.006; A-loop × αC-helix, p = 4.34×10–9.
For tyrosine kinases, the A-loop-OUT conformation was associated primarily with DFG-IN and tended to co-occur with αC-helix-IN (Figure 4).
Figure 4. Association diagram for tyrosine kinases between the activation segment OUT conformation, the DFG motif conformations, and the αC-helix. Fisher tests for the TK dataset: A-loop × DFG, p = 0.011; A-loop × αC-helix, p = 0.058.
The A-loop-IN conformation was most frequently observed with αC-helix-IN and DFGOUT; however, αC-helix-OUT and DFG-IN configurations also occurred (Figure 5).
Figure 5. Association diagram for tyrosine kinases between the activation segment IN conformation, the DFG motif conformations, and the αC-helix. Fisher tests for the TK dataset: A-loop × DFG, p = 0.011; A-loop × αC-helix, p = 0.058.
Examining the few SWAPPED active conformations in tyrosine kinases, we found they are associated with the DFG-IN conformation, while the αC-helix can be in either OUT or IN configuration, with a slight predominance of the former (Figure 6).
Figure 6. Association diagram for tyrosine kinases between the SWAPPED activation segment conformation, the DFG motif conformations, and the αC-helix. SWAPPED cases are descriptive only; Fisher tests for TK IN/OUT associations are reported in Supplementary Table S5.
Energetic classification of protein kinase activation segments
K-means clustering was applied to the four classification criteria (DFG-in/out, A-loop-in/out, αC-helix-in/out, and activation-segment interaction energy with the catalytic domain) after standardizing all four features to unit variance (z-score), so that the categorical conformational states and the continuous energy contribute equally to the Euclidean distance. This yielded seven clusters, designated Cluster 0 through Cluster 6 (mean silhouette 0.527 at K = 7), listed in Table A1; the cluster indices follow this standardized re-clustering, and the full membership is given in Supplementary Table S2. With the balanced standardization, the partition is DFG-homogeneous (all seven clusters pure on the DFG state), whereas the A-loop IN/OUT state is only partially aligned with the clusters (four of seven A-loop-pure), reflecting a genuine partial decoupling between A-loop geometry and the DFG/αC/energy axis rather than a clustering failure.
The standardized partition is DFG-homogeneous in all seven clusters and A-loop-homogeneous in four of them (Clusters 1, 2, 4, and 5). The three A-loop-heterogeneous clusters are Cluster 0 (6 A-loop-IN/3 OUT), Cluster 3 (4 IN/3 OUT), and Cluster 6 (4 IN/2 OUT) (Supplementary Table S8). Here “heterogeneous” describes cluster composition—structures that individually carry different, unambiguous A-loop labels—not a thermodynamic equilibrium and not an unresolved K-means substate, since the A-loop labels are assigned from geometry and KLIFS independently of, and prior to, clustering. This heterogeneity expresses the study’s premise directly: the DFG flip dominates the R-spine/energetic state, whereas the activation-loop tip can occupy several positions within a given DFG/αC configuration, so interaction energy and A-loop geometry are partially orthogonal. The seven clusters, with their mean ± standard deviation activation segment interaction energies (kJ/mol) and dominant conformational character, are: Cluster 0: −74.6 ± 18.2 (n = 9); DFG-out, predominantly inactive (A-loop-in). Cluster 1: −79.1 ± 10.0 (n = 30); DFG-in/A-loop-out (active), the largest cluster. Cluster 2: −63.7 ± 5.1 (n = 2); SWAPPED activation segment. Cluster 3: −89.8 ± 20.0 (n = 7); DFG-out, mixed A-loop. Cluster 4: −110.2 ± 12.0 (n = 9); DFG-in/A-loop-out (active), the most stabilized. Cluster 5: −98.5 ± 13.6 (n = 5); DFG-in/A-loop-in (inactive DFG-in). Cluster 6: −80.8 ± 21.6 (n = 6); DFG-in, mixed A-loop.
The SWAPPED A-loop configuration was identified in only two structures (n = 2), both assigned to Cluster 2: MAP4K1/HPK1 (6CQD) and AURA (4C3P). Both structures were associated with DFG-IN and αC-helix-IN and are treated as observations rather than as a validated class.
The underrepresentation of SWAPPED structures in this subset (2/68 = 2.9% versus 6% in the broad survey) reflects two compounding factors rather than a modeling artifact. First, the SWAPPED mechanism is intrinsically rare and family-restricted (here Aurora A and MAP4K1/ HPK1). Second, the requirement for a fully resolved activation segment preferentially excludes dimeric swapped segments, which are often disordered; the AURA SWAPPED and inactive structures have the highest activation-segment B-factors in the representative set (79–95 Å2; Table S1). The observed 2/68 proportion is not statistically distinguishable from 6% (exact binomial p ≈ 0.44; Wilson 95% CI 0.8–10.1%). This depletion is therefore attributable to crystallo-graphic selection before energy calculation, not to the AMBER parm03 force field, which was used only after structural selection and neither models nor penalizes the dimeric quaternary state. Both SWAPPED structures were retained and scored (4C3P, −60.14 kJ/mol; 6CQD, −67.32 kJ/mol).
The one-dimensional impact of each fragment’s configuration on interaction energy is kinase-dependent and cannot be generalized: αC-helix effects ranged from −4.88 to −7.19 kJ/mol (CASK/BMX), while DFG effects varied from −5.29 to −34.58 kJ/mol across EPHA2, FGFR4, MET, and RPS6KA3.
Mean interaction energies per cluster are reported with their standard deviations in Table A1 (within-cluster SD 5–22 kJ/mol). The substantial within-class spread reflects that interaction energy carries information beyond the geometric IN/OUT label: when all four features are standardized, the partition separates the DFG state cleanly (all seven clusters DFG-homogeneous), whereas the A-loop IN/OUT state remains only partially aligned with the clusters. Energy and A-loop geometry are therefore partially independent, and clusters sharing an energy range may contain different activation-segment geometries, an expected consequence of the energetic axis rather than a failure of the partition.
Clusters 1 and 4 share the DFG-in/A-loop-out active label yet differ by ~31 kJ/mol in mean activation-segment interaction energy (−79.1 vs −110.2; Welch t = 7.1, p < 0.001). Coordinate inspection identifies the structural basis. First, activation-loop phosphorylation: 5 of 9 Cluster 4 members carry a phosphorylated activation-loop residue (pThr/pSer/pTyr between the DFG and APE motifs) versus 4 of 30 in Cluster 1 (Supplementary Table S9), and phosphorylated structures are ≈20 kJ/mol more stabilized (−101.6 vs −81.6 kJ/mol pooled). The phosphate forms a multivalent salt-bridge network with the catalytic domain; in PKA (3FJQ, the most stabilized structure at −137.1 kJ/mol), pThr197 simultaneously contacts Arg165 (the catalytic HRDarginine, 2.77 Å), Lys189 (2.71 Å), and His87 of αC (3.74 Å). Second, family-specific packing: Cluster 4 is AGC/CMGC-enriched and gains further activation-segment-body coupling from the AGC C-terminal tail, acidic or pre-ordered A-loops (e.g., CK2α) and dense basic catalytic-loop environments, so even its nonphosphorylated members remain strongly stabilized (−105 vs −78 kJ/mol). These contributions are captured by the interaction-energy axis but are invisible to the geometric IN/OUT label—the quantitative rationale for adding an energetic dimension.
Table 1. Per-cluster activation-segment interaction energies after all-four-feature z-score standardization (4C3P energy corrected; full 68-structure membership in Supplementary Table S2).
| Cluster | Dominant conformational character | Mean (kJ/mol) | SD | n |
|---|---|---|---|---|
| 0 | DFG-out, predominantly inactive (6 A-loop-IN/3 OUT) | −74.6 | 18.2 | 9 |
| 1 | DFG-in/A-loop-out (active), largest | −79.1 | 10.0 | 30 |
| 2 | SWAPPED | −63.7 | 5.1 | 2 |
| 3 | DFG-out, mixed A-loop | −89.8 | 20.0 | 7 |
| 4 | DFG-in/A-loop-out (active), most stabilized | −110.2 | 12.0 | 9 |
| 5 | DFG-in/A-loop-in (inactive DFG-in) | −98.5 | 13.6 | 5 |
| 6 | DFG-in, mixed A-loop | −80.8 | 21.6 | 6 |
Per-cluster PDB IDs. Cluster 0: 1G3N, 1IRK, 3KMU, 3RHK, 3UOH, 6BXI, 6HHJ, 6I99, and 6PDJ. Cluster 1: 1FIN, 1J1B, 1JKK, 1WBP, 2BIY, 2F4J, 2HK5, 2QR7, 2VAG, 2X4F, 2Y7J, 3C0I, 4FR4, 4IIR, 4NFN, 4NW6, 4QPM, 4W9W, 4WKQ, 4WSQ, 4XCU, 5O0Y, 5ORL, 5TQ8, 5UKL, 5WNF, 6BDL, 6DC0, 6NO9, and 6RCG. Cluster 2: 4C3P and 6CQD. Cluster 3: 3NPC, 3SXS, 4NUS, 4QRC, 5IA5, 6GQQ, and 6VG3. Cluster 4: 1XJD, 2WTK, 3FJQ, 3IQ7, 3OWJ, 3QKL, 4OTD, 4WNO, and 4ZZN. Cluster 5: 3PIX, 5CEN, 5FXS, 5NKF, and 6BS0. Cluster 6: 2G1T, 3C0G, 4R1V, 5H0B, 5J0A, and 6Q4K.
Interaction energy distribution and spatial heat maps
To further characterize activation-segment interaction patterns, we constructed spatial heat maps for eleven representative structures spanning all seven clusters (Clusters 0–6) and the OUT, IN, and SWAPPED conformational states. Clusters 3 (DFG-out, mixed A-loop) and 5 (DFG-in/A-loop-in) are illustrated by VEGFR2 (PDB ID: 6GQQ) and DLK (PDB ID: 5CEN), respectively. The largest cluster (Cluster 1, n = 30) is illustrated by members from several kinase families (CDK2, AURA, HCK, and AAK1) to convey within-cluster diversity rather than a single archetype. These maps highlight contacts between the activation segment and the catalytic domain and show their relative intensities.
Inactive AKT1 kinase (Cluster 0)
This kinase (PDB ID: 6HHJ, Cluster 0, IN) has a mean energy of −76.48 kJ/mol (Figure 7a). Its heat map shows low-amplitude interactions between the activation segment and the β1 strand (Figure 7b). –210 –200 –150 –100 –50 0 3 5 10 20 40
Figure 7. (A) Structure of protein kinase AKT1 (PDB ID: 6HHJ) and (B) spatial heat map of the interaction between the activation segment and the catalytic domain. 58
Inactive AURA kinase (Cluster 0)
This kinase (PDB ID: 3UOH, Cluster 0, IN) has a mean energy of −34.00 kJ/mol and an activation segment forming a small sheet with β1 (Figure 8a). Its heat map shows pronounced catalytic loop interactions, moderate internal segment interactions, and weak β1/αEF/αF interactions (Figure 8b).
Figure 8. (A) Structure of protein kinase AURA (PDB ID: 3UOH) and (B) spatial heat map of the interaction between the activation segment and the catalytic domain.
Active CDK2 kinase (Cluster 1)
This kinase (PDB ID: 1FIN, Cluster 1, OUT) has a mean activation segment energy of −64.53 kJ/mol (Figure 9a). Its heat map reveals predominantly intra-activation segment interactions and interactions with the C-loop and αEF/αF loop, with the strongest interactions within the segment itself (Figure 9b).
Figure 9. (A) Structure of protein kinase CDK2 (PDB ID: 1FIN) and (B) spatial heat map of the interaction between the activation segment and the catalytic domain; color and intensity reflect interaction strength.
Active AURA kinase (Cluster 1)
This kinase (PDB ID: 5ORL, Cluster 1, OUT) has a mean energy of −82.37 kJ/mol (Figure 10a). Its heat map shows moderate interactions at the catalytic loop, activation segment, and αEF/αF loop (Figure 10b).
Figure 10. (A) Structure of protein kinase AURA (PDB ID: 5ORL) and (B) spatial heat map of the interaction between the activation segment and the catalytic domain.
Active HCK kinase (Cluster 1)
This kinase (PDB ID: 2HK5, Cluster 1, OUT) has a mean energy of −76.58 kJ/mol and a well-structured activation segment with β-sheets (Figure 11a). Its heat map shows internal segment interactions and interactions with the αEF/αF loop, with weaker catalytic loop contacts (Figure 11b).
Figure 11. (A) Structure of protein kinase HCK (PDB ID: 2HK5) and (B) spatial heat map of the interaction between the activation segment and the catalytic domain.
AAK1 kinase (Cluster 1, activation segment OUT)
This kinase (PDB ID: 4WSQ, Cluster 1, OUT) has a mean energy of −94.35 kJ/mol and a helix within the activation segment (Figure 12a). Its heat map shows moderate-to-strong internal interactions, moderate catalytic loop contacts, and weak αEF/αF interactions (Figure 12b).
Figure 12. (A) Structure of protein kinase AAK1 (PDB ID: 4WSQ) and (B) spatial heat map of the interaction between the activation segment and the catalytic domain.
SWAPPED dimeric form of AURA kinase (Cluster 2)
This kinase (PDB ID: 4C3P, Cluster 2, SWP) has an activation-segment interaction energy of −60.14 kJ/mol. The activation segment adopts a swapped arrangement across a kinase dimer that is consistent with a trans-autophosphorylation assembly (Figure 13a). Its heat map shows weak interactions with the catalytic loop, within the activation segment, and with the αEF/αF loop (Figure 13b).
Figure 13. (A) Structure of protein kinase AURA (PDB ID: 4C3P) and (B) spatial heat map of the interaction between the activation segment and the catalytic domain.
Inactive VEGFR2 kinase (Cluster 3, DFG-out)
This kinase (PDB ID: 6GQQ, Cluster 3, DFG-out) has a mean energy of −98.10 kJ/mol and adopts the catalytically inactive DFG-out conformation, in which the DFG motif is flipped and the activation segment displaced (Figure 14a). Its heat map is dominated by internal interactions along the activation segment, together with moderate contacts with the catalytic loop and the αEF/αF region (Figure 14b).
Figure 14. (A) Structure of protein kinase VEGFR2 (PDB ID: 6GQQ) and (B) spatial heat map of the interaction between the activation segment and the catalytic domain.
Active AKT1 kinase (Cluster 4)
This kinase (PDB ID: 3QKL, Cluster 4, OUT) has a mean energy of −112.43 kJ/mol (Figure 15a). Its heat map shows stronger interactions at the catalytic loop and within the activation segment compared to the αEF/αF loop or β1 strand (Figure 15b).
Figure 15. (A) Structure of protein kinase AKT1 (PDB ID: 3QKL) and (B) spatial heat map of the interaction between the activation segment and the catalytic domain.
Inactive DLK kinase (Cluster 5, DFG-in)
This kinase (PDB ID: 5CEN, Cluster 5, IN) has a mean energy of −92.69 kJ/mol and represents the inactive DFG-in state, with the activation segment folded inward (A-loop-in) (Figure 16a). Its heat map is dominated by strong internal interactions within the activation segment, including a prominent favorable contact, together with contacts with the catalytic loop and the αEF/ αF region (Figure 16b).
Figure 16. (A) Structure of protein kinase DLK (PDB ID: 5CEN) and (B) spatial heat map of the interaction between the activation segment and the catalytic domain.
Inactive HCK kinase (Cluster 6)
This kinase (PDB ID: 5H0B, Cluster 6, IN) has a mean energy of −93.15 kJ/mol and an incipient helix within the activation segment (Figure 17a). Its heat map shows moderate internal and catalytic loop interactions, with weak αEF/αF and particularly weak β1 contacts (Figure 17b).
Figure 17. (A) Structure of protein kinase HCK (PDB ID: 5H0B) and (B) spatial heat map of the interaction between the activation segment and the catalytic domain.
Overall, the spatial heat maps provide a residue-level representation of interactions involving the activation segment, both internally and with the remainder of the catalytic domain (Galgonek et al., 2017). Interactions with the catalytic loop, β1 strand, and αEF/αF loop occurred at varying intensities and were generally moderate to weak, with a few exceptions. The activation segment–catalytic loop interaction can form the β6–β9 sheet (Nolen et al., 2004). Contacts with β1 or the αEF/αF region, mainly in active forms, can form a β10–β11 sheet, as observed in HCK. Numerous internal contacts reflect the multiple secondary-structure elements present within the activation segment.
Differential analysis and per-residue Cα displacement calculation between IN/OUT conformations
Having observed that certain representative kinases from different energetic clusters possess both active and inactive conformations, we investigated the switching between these states by identifying the conformational changes accompanying this transition. This was achieved through a differential analysis of per-residue Cα displacements between the active and inactive forms.
Inactive-to-active conformational switch of HCK kinase
Comparing the two heat maps constructed previously, we found that the two energy maps differ only slightly, particularly in terms of internal activation segment interactions which become weaker, while the activation segment–catalytic loop interactions become relatively more visible. During the IN–OUT switch, the largest conformational change (per-residue Cα displacement ~14 Å) involves the A-loop, reflecting its large conformational displacement and central role in the switch; smaller changes were observed in the G-loop (glycine-rich), β2/β3, αC, and β4/β5, with per-residue Cα displacement values ranging from 3 to 4 Å (Figure 18).
Figure 18 (A) Superposition of the two HCK kinase conformations (PDB IDs: 2HK5 and 5H0B, shown in brown and blue, respectively; black arrow indicates the direction of displacement) and (B) per-residue Cα displacement comparison between chain A of PDB 2HK5 (activation segment OUT) and chain A of PDB 5H0B (activation segment IN) for the HCK catalytic domain.
Inactive-to-active conformational switch of AKT1 kinase
Comparison of the IN and OUT heat maps shows stronger internal activation-segment and activation segment–catalytic loop interactions in the OUT form. Per-residue Cα displacement analysis identifies the A-loop as the region with the largest displacement (~19 Å), followed by the αC-helix (~11 Å), G-loop (~5 Å), and β4/β5 region (~4 Å) (Figure 19). The global RMSD is ~8.1 Å. A localized ~5 Å feature occurs near phosphorylated Thr308 at the base of the A-loop peak.
Figure 19. (A) Superposition of the two AKT1 kinase conformations (PDB IDs: 3QKL and 6HHJ, shown in brown and blue, respectively) and (B) per-residue Cα displacement comparison between chain A of PDB 3QKL (OUT) and chain A of PDB 6HHJ (IN) for the AKT1 catalytic domain.
Inactive-to-active conformational switch of AURA kinase
Because AURA is represented by three forms—one inactive and two active or activation- compatible conformations (Chiu et al., 2013; Pike et al., 2008)—we analyzed the structural differences between the inactive IN form and both the OUT and SWAPPED forms.
IN/OUT switch
The largest displacement remains in the A-loop, reaching 18 Å. Other marginal variations not exceeding 3 Å were also detected (Figure 20). The global RMSD of the catalytic domain during this switch is approximately 5.99 Å.
Figure 20. (A) Superposition of the two AURA kinase conformations (PDB IDs: 5ORL and 3UOH, shown in brown and blue, respectively) and (B) per-residue Cα displacement comparison between chain A of PDB 5ORL (OUT) and chain A of PDB 3UOH (IN) for the AURA catalytic domain.
IN/SWP switch
The IN/SWP comparison exhibits the largest structural differences: the A-loop displacement reaches 23 Å, with broader changes of ~5 Å across residues Ala273–Lys309. The global RMSD is ~6.99 Å. A local displacement of ~8 Å is centered near Thr288 (Figure 21b); however, 4C3P is a dephosphorylated Aurora A structure, so this feature should not be attributed to Thr288 phosphorylation.
Figure 21. (A) Superposition of the two AURA kinase conformations (PDB IDs: 4C3P and 3UOH, shown in brown and blue, respectively) and (B) per-residue Cα displacement comparison between chain A of PDB 4C3P (SWP) and chain A of PDB 3UOH (IN) for the AURA catalytic domain.
Inactive-to-active conformational switch of CDK2 kinase
In the CDK2 comparison, the A-loop shows the largest displacement; additional changes of ~5 Å occur in the αC-helix/β4/β5 region, whereas the G-loop and β2/β3 region remain below 3 Å (Figure 22). The global RMSD is ~4.55 Å. The OUT structure (1FIN) is a cyclin A–CDK2 complex, and cyclin binding is known to reposition the αC/PSTAIRE helix and activation loop. Therefore, part of the observed difference probably reflects cyclin-dependent activation rather than an intrinsic IN-to-OUT transition of isolated CDK2 (Jeffrey et al., 1995).
Figure 22. (A) Superposition of the two CDK2 kinase conformations (PDB IDs: 1FIN and 6Q4K, shown in brown and blue, respectively) and (B) per-residue Cα displacement comparison between chain A of PDB 1FIN (OUT) and chain A of PDB 6Q4K (IN) for the CDK2 catalytic domain.
The three-category preliminary classification (IN, OUT, and SWAPPED) is consistent with the literature (Adams, 2003; Modi & Dunbrack, 2019) and shows that OUT conformations predominate (55%) among the resolved kinase structures surveyed, whereas IN conformations account for 38%. SWAPPED conformations accounted for 6% in the broad survey, but only two examples remained in the fully resolved 68-structure subset; this category is therefore retained as a descriptive observation pending validation in a larger dataset (Beenstock et al., 2016; Gilburt et al., 2017). Subclassification by secondary-structure content within the activation segment, particularly the β6–β9 and β10–β11 sheets, provides a finer structural vocabulary that has not been systematically used in previous classifications (Möbitz, 2015; Ung et al., 2018).
The association diagrams show that the active OUT conformation is associated with DFG-IN in both serine/threonine and tyrosine kinases and tends to co-occur with αC-helix-IN. The αC-helix association was strong in serine/threonine kinases but did not reach conventional statistical significance in tyrosine kinases (p = 0.058). These patterns are consistent with canonical kinase-activation models (Nolen et al., 2004; Taylor & Kornev, 2011) and the KLIFS framework (Kooistra et al., 2016). By contrast, A-loop-IN structures tolerate both DFG-IN and DFG-OUT states. In serine/threonine kinases, DFG-IN was modestly more frequent, consistent with partial uncoupling between A-loop geometry and regulatory-spine/DFG organization (Hu et al., 2015; Modi & Dunbrack, 2019). The more varied αC-helix states observed among inactive tyro-sine kinases may reflect greater structural plasticity and warrant further study.
These associations were quantified with Fisher’s exact test (Supplementary Table S5): across the dataset, the A-loop state is significantly associated with both the DFG state (p = 4.1×10–5) and the αC-helix state (p = 1.2×10–8). We interpret them as descriptive statistical associations, not causal or mechanistic linkages. As an independent functional check, the regulatory spine (RS3–RS4 proxy) was found assembled in all active-cluster representatives, and a catalytic-motif screen (HRD-Asp, DFG-Asp) indicated that pseudokinase inclusion does not drive the clustering (Supplementary Tables S3–S4).
The K-means clustering defines seven descriptive energy/conformation clusters, indicating that interaction energy and conformational state together define a richer classification landscape than geometry alone. The most stabilized cluster (Cluster 4, mean −110.2 kJ/mol) is dominated by active DFG-in/A-loop-out kinases (AGC-rich), whereas the SWAPPED cluster (Cluster 2, −63.7 kJ/mol) and the DFG-out, predominantly inactive cluster (Cluster 0, −74.6 kJ/mol), are among the least stabilized; the substantial within-cluster spread (SD up to ~22 kJ/mol; Table A1) confirms that interaction energy carries information beyond the geometric IN/OUT label. The SWAPPED configurations, both assigned to Cluster 2, are retained as observations only; because they are based on n = 2 and include one high-B-factor structure, they should not be generalized as a validated energetic class.
The spatial heat maps make explicit what global energy values obscure: that the activation segment–catalytic loop interaction (yielding the β6–β9 sheet) and the activation segment–αEF/ αF loop interaction (yielding the β10–β11 sheet in active forms such as HCK) are the dominant inter-element interactions. Internal activation segment interactions are consistently the strongest in absolute terms, underscoring the structural self-organization of the activation segment independent of its external contacts (Galgonek et al., 2017). The progressive evolution of these interaction maps between inactive and active forms provides a residue-level map of how energy redistribution accompanies conformational switching.
The per-residue Cα displacement analysis identifies the A-loop as the region with the largest difference between inactive and active end states. These magnitudes are not measures of thermal motion, which would require a molecular-dynamics trajectory or conformational ensemble. A-loop displacements reached 14–19 Å in the IN/OUT comparisons and 23 Å only in the AURA IN/SWAPPED comparison; these changes substantially exceeded those of other structural regions and were consistent with the A-loop’s role as the principal gating element for substrate access (Chiu et al., 2013; Steichen et al., 2012). Secondary displacements in the αC-helix (notably ~11 Å in AKT1) and G-loop reflect coordinated structural reorganization accompanying A-loop repositioning. Their kinase-specific patterns in HCK, AKT1, AURA, and CDK2 argue against a universal conformational template and suggest that inhibitor design should account for kinase-specific structural signatures.
The AKT1 profile contains a localized feature near phosphorylated Thr308. By contrast, the AURA IN/SWP comparison shows a displacement centered near Thr288 even though 4C3P is dephosphorylated; this signal therefore reflects local conformational rearrangement rather than a phosphorylation event. Per-residue displacement profiles should be interpreted as structural differences between crystallographic end states, not as direct detectors of post- translational modification.
Several limitations should be noted. (i) The analysis relies on static crystal structures; the per-residue Cα displacement reported here quantifies conformational differences between end states, not thermal motion, and future work should integrate molecular dynamics to map transition pathways. (ii) The INTAA/AMBER parm03 interaction energies are computed in vacuo and should be read as relative structural signatures; for a representative subset we verified, under a Poisson–Boltzmann implicit solvent with PROPKA-assigned protonation at pH 7.4 and physiological ionic strength, the activation-segment interactions remain favorable (Supplementary Table S7), while a full force-field re-ranking additionally including van der Waals terms (MM-PBSA/MM-GBSA) remains a priority for future work. (iii) A few activation than the full four-residue spine. Extending the approach to broader kinome datasets, including disease-associated variants and explicit pseudokinase filtering, is an important next step.
Conclusion
This study analyzed a broad dataset of protein kinases from an integrated structural and energetic perspective and established the following. (i) Kinases were classified into three activation-segment configurations: IN, in which the segment occludes the substrate-binding cleft; OUT, in which it is displaced from that cleft and is compatible with an active arrangement; and a minor SWAPPED configuration associated with dimeric trans-autophosphorylation assemblies. (ii) Association diagrams showed that the OUT activation segment is associated with DFG-IN and tends to co-occur with αC-helix-IN, whereas the IN activation segment is compatible with multiple regulatory-element states. (iii) Incorporating interaction energy alongside the A-loop, DFG, and αC-helix states yielded seven descriptive clusters, and spatial heat maps localized the strongest activation segment–catalytic-domain contacts. (iv) Static inter-conformational comparisons showed that A-loop repositioning dominates the structural transition, with kinase-specific secondary changes in the αC-helix and adjacent elements. This classification complements existing geometric schemes and provides a quantitative framework for studying kinase activation and for designing inhibitors that target particular conformational states, including type II (DFG-out) and allosteric inhibitors.