KinaseConf DFG-state classifier

⬢

KinaseConf — DFG-state classifier

Classifies a kinase activation loop as DFG-in, DFG-out, or DFG-inter from a single structure — with the Modi/Dunbrack cluster, a confidence score, and a Type-II druggability estimate.

Drop a .pdb/.cif file, fetch by PDB ID or UniProt accession, or pick an example below. Runs entirely in your browser.

Cross-validated accuracy 99.57% on 7,459 KinCoRe chains, with independent external checks. See the Validation tab.

1ATP · PKA active 1IEP · Abl + imatinib 1FVR · Hck inactive 3GVU · c-Met type-II

Consensus call

— awaiting

—

Drop a structure to see its DFG-state classification.

—

awaiting

—

confidence

Geometric metrics

distances and torsions vs. published thresholds

Method ensemble

three independent published criteria

Method	Call	Detail

Diagnostics

structure quality and edge-case warnings

ready · no structure loaded

Batch classify

Drop multiple PDB / mmCIF files (or paste a list of PDB IDs, one per line). Each structure is parsed and classified locally; results download as a single CSV. Maximum 200 inputs per batch.

↥ Drop or browse multiple .pdb / .cif files

Or paste PDB IDs

New here? Load example classifies a curated set — PKA (active, DFG-in), and Abl·imatinib & c-Met (DFG-out, type-II inhibitor complexes) — so you can see the full table, dashboard, distribution, confidence/druggability bars, and Notes flags at a glance.

Source	Chain	Conformation	Cluster	Confidence	Type-II druggability	Active	Notes

How to read the columns

The dashboard at the top summarises the whole batch: how many chains were classified (and across how many structures), how many are catalytically-primed (active), how many have a high type-II druggability score, and the mean classifier confidence. The coloured bar beneath it shows the DFG-state distribution — the fraction of chains in each conformation.

Source / Chain — The input file or PDB ID, and the protein chain within it. Every chain ≥100 residues is classified separately, so one multi-chain structure yields several rows.

Conformation (coloured pill) — Consensus DFG state from three published criteria (Vijayan distances, Modi/Dunbrack spatial group, Möbitz pseudo-torsion). DFG-in = active-ready geometry; DFG-out = Phe flipped into the ATP pocket (type-II inhibitor territory); DFG-inter = an in-between state.

Cluster — Modi/Dunbrack 8-state spatial-plus-Ramachandran label refined by the ONNX MLP. BLAminus is the canonical active state; BBAminus is the canonical DFG-out cluster occupied by classical type-II inhibitor complexes (imatinib, sorafenib, ponatinib).

Confidence (0–100, coloured bar) — Composite of method agreement (40), structure completeness (25), B-factor quality (20), and threshold proximity (15). The bar is green ≥75, yellow 50–74, red <50.

Type-II druggability (0–100, coloured bar) — Accessibility of the allosteric back pocket to a type-II inhibitor. It asks: "Could a type-II inhibitor extending from the hinge reach the cryptic pocket opened by a DFG flip?". Inputs are gatekeeper bulk (smaller → more accessible), αC-helix dilation, and the geometric viability of a linker from the gatekeeper Cβ to the αC-Glu Cβ. ≥75 implies a chemically reasonable type-II growth vector exists; ≤25 implies the pocket is sterically gated. This is not a generic druggability metric.

Active — ✓ active means the chain passes the Kincore-style "catalytically primed" test (Modi & Dunbrack 2019): DFG-in and cluster BLAminus and the DFG-Phe/DFG-Asp atoms present and the αC-Glu/β3-Lys Cβ–Cβ distance ≤10 Å (catalytic salt bridge feasible). All four must hold; — means one or more fail (the default for DFG-out, DFG-inter, broken salt bridges, or non-BLAminus DFG-in).

Notes — Per-chain warnings worth a second look (e.g. an AlphaFold model, a low-resolution or disordered DFG region, an out-of-distribution flag). Hover to read the full text. A blank cell means no warnings were raised.

The full numeric feature vector for every chain (all Vijayan/Modi/Möbitz distances and angles, B-factors, flags) is included in the downloadable CSV.

Benchmarks

The classifier is exercised on a hand-picked reference set of canonical PDB structures drawn from the foundational literature on kinase-conformation classification. Each structure has a published expected DFG state; disagreement indicates an algorithmic gap.

Structures are fetched a few at a time from RCSB and classified locally; the run takes a few seconds and results are cached, so re-runs are instant.

PDB	Reference	Expected	KinaseConf call	Cluster	Confidence	Match

Note on scope. The gallery above is a sanity-check set of well-characterised structures provided for at-a-glance verification. The complete quantitative validation is reported in the Validation tab and comprises stratified five-fold cross-validation on the 7,459-chain KinCoRe training set (accuracy 99.57%, macro-F₁ 0.989), independent external agreement with the Simonson laboratory KinActive annotations (98.7% on 923 chains absent from training), agreement with FDA-approved inhibitor binding modes (95.5% on 44 unique drug–target pairs), and probability calibration (ECE = 0.0071).

Methods

KinaseConf assigns a conformational-state label to a protein-kinase domain from a single crystal, cryo-EM, or AlphaFold structure. The output identifies the activation-loop state (DFG-in, DFG-out, or an intermediate) and the corresponding Modi & Dunbrack cluster, both of which determine which classes of small-molecule inhibitor are likely to bind. All computation is performed client-side; no structural data leaves the browser. The sections below describe, in order: the classification pipeline, the anchor residues and geometric features used, how the consensus call is constructed, how reliability is assessed, and scope and limitations.

Methods · 01

1. The KinaseConf pipeline

A structure goes through five stages: parse → locate the four conserved kinase motifs → measure 9 geometric features → classify with a small neural network → compose a verdict with a confidence score and an out-of-distribution check.

Figure 1. The five steps of a single classification. The network sees only the nine geometric features measured in step 3 — not sequence — the same kind of measurements a structural biologist records by hand. The rule check in step 4 applies published DFG-state criteria alongside the network.

Technical detail (struct-bio reader)

The browser parses the input via a streamed PDB / mmCIF reader (with support for multi-model NMR ensembles and DCD/XTC trajectory frames, see Trajectory tab). Stage 2 runs the field-standard HMMER profile-HMM anchoring (Pfam PF00069 Ser/Thr kinase + PF07714 Tyr kinase, union-classified by bit score — matches KinActive 2024 and Kincore3). Stage 3 measures: ϕ/ψ of the X-DFG residue, ϕ/ψ of DFG-Asp, ϕ/ψ of DFG-Phe, χ₁ of DFG-Phe (the rotamer flag), and two Cα–Cζ distances localising the Phe ring (D₁ from αC-Glu+4 Cα, D₂ from β3-Lys Cα). Stage 4 evaluates three independent published criteria (section 3) AND an MLP trained on 7,459 KinCoRe chains; the consensus call defaults to the Modi/Dunbrack spatial assignment when criteria disagree. Stage 5 composes a 0–100 confidence score from four orthogonal components and assigns an out-of-distribution tier from a per-cluster Mahalanobis distance.

1 · Parse & locate the kinase domain

local structure parsing · Pfam PF00069 / PF07714

The coordinate file is read in the browser. The catalytic domain is identified by hidden-Markov-model search against the Pfam protein-kinase families, so the active-site anchors can be placed even when the deposited residue numbering is non-canonical.

2 · Locate the conserved motifs

KLIFS pocket numbering · 4 motifs

The β3 lysine, the αC glutamate, the HRD aspartate and the DFG aspartate–phenylalanine–glycine triad are mapped onto the structure through the KLIFS 85-residue pocket alignment.

3 · Measure the geometry

7 dihedrals + 2 distances → 16-d vector

Side-chain dihedrals and inter-anchor distances are computed directly from the coordinates. These nine numbers capture whether the DFG motif points into or out of the ATP cleft — no sequence information is used.

4 · Classify the state

16→32→32→8 network + published criteria

The feature vector is passed through a small neural network and, in parallel, three published rule-based criteria. The geometry is mapped to a DFG-in, DFG-out or intermediate state under the Modi & Dunbrack reference scheme; each criterion is evaluated separately so the call can be inspected.

5 · Compose the verdict

state · cluster · confidence · out-of-distribution check

The assigned state is reported with its supporting criteria, a calibrated confidence score, the Modi & Dunbrack cluster, and an out-of-distribution flag indicating whether the structure resembles those the classifier was fitted on.

Methods · 02

2. Where in the kinase the model looks

KinaseConf reads only four anchor residues — the same four that every published kinase-conformation scheme uses. From those four anchors it measures 9 numbers that summarise the geometry of the activation loop, and feeds those 9 numbers (and only those 9) to the classifier.

PKA · cAPK DFG-in · αC-in · Active PDB 1ATP · chain E

ABL1 · ABL kinase DFG-out · αC-out · Type-II PDB 1IEP · chain A · + imatinib

β3-Lys

αC-Glu

HRD-Asp

DFG triad

Bound ligand

How the feature vector reads these sidechains — one structural view per feature

Feature

Structural view

Active (PKA / 1ATP)

Inactive (ABL1 / 1IEP)

Units

d(K·NZ, E·OE1)

β3-Lys ↔ αC-Glu salt bridge

PKA active

3.0 Å salt bridge formed

~12 Å E rotated out

d(αC·E·OE1, F·CZ)

αC-Glu ↔ DFG-Phe ring

PKA active

6.8 Å αC-in, R-spine intact

11.5 Å αC swung out

χ1(F · CA-CB-CG-CD1)

DFG-Phe sidechain rotamer

PKA F185 · g−

ABL1 F382 · t

−70° g−, F away from αC

−175° t, F into cleft

d(HRD·D·CG, DFG·D·CG)

R-spine proxy (D166 ↔ D184)

PKA active

8.2 Å R-spine aligned

10.4 Å R-spine broken

d(K·NZ, F·CZ)

β3-Lys ↔ DFG-Phe ring

PKA · F away

ABL1 · F blocks K

9.1 Å Phe out of cleft

5.6 Å Phe blocks K

A-loop COM offset

resi 184-208 centroid vs C-lobe ref

PKA · A-loop highlighted

+6.5 Å extended outward

−2.1 Å collapsed in

Å (rel.)

Values are illustrative reference points from 1ATP / 1IEP after KLIFS-aligned anchor selection. The full feature vector is 17-D (10 inter-anchor distances + 4 χ angles + 3 motif-COM offsets); the six rows above are the most discriminative subset based on classifier feature importance. All insets are rendered from the same PDB structures used in the main panels.

Label

DFG-in

Phe-CZ rotated away from αC; K–E salt bridge intact; competent for ATP binding.

Label

DFG-out

Phe-CZ flipped ~180° into the back pocket; salt bridge broken; type-II inhibitor binding mode.

Label

Inter / other

Intermediate Phe χ1 or salt bridge partially formed; KinCore "inter" class.

Figure 2. Anchor residue side chains in the two reference structures used to train and benchmark the classifier. Left: PKA (PDB 1ATP, chain E) — canonical active DFG-in / αC-in conformation. K72–E91 salt bridge (amber dashed line) intact at 3 Å; F185-CZ points away from αC; HRD-Asp (D166) and DFG-Asp (D184) form an aligned R-spine. Right: ABL1 (PDB 1IEP, chain A) bound to imatinib — canonical DFG-out / αC-out conformation. F382-CZ rotated ~180° into the ATP pocket (occupied by imatinib, cyan), K271–E286 distance expanded to ~12 Å (faint grey dash, "broken"), αC swung outward. Inset table: the six most-discriminative features of the 17-D vector, each illustrated with a view of the exact atoms/sidechains measured, shown against the translucent kinase fold (β-sheet, αC helix and A-loop lightly tinted) so the motif's location in the domain is visible while the measured atoms remain readable. Sticks are coloured by motif (β3-Lys=rose, αC-Glu=amber, HRD-Asp=violet, DFG triad=green); imatinib carbons cyan.

Why these four residues, and how they are localised in practice

Why four anchors. Every published kinase-conformation scheme — Vijayan 2015, Möbitz 2015, Modi & Dunbrack 2019, KinCoRe — uses the same four conserved residues because their positions define the catalytic geometry: the αC-Glu / β3-Lys salt bridge is the catalytic ion pair, HRD-Asp is the catalytic base, and the DFG triad is the regulatory switch.

How they are located. The default anchor source is HMMER profile-HMM alignment against the union of Pfam PF00069 (Ser/Thr protein kinase) and PF07714 (tyrosine kinase / Pkinase_Tyr). On the full RCSB protein-kinase set (7,669 PDB entries) this yields anchors for 11,457 chains. PF07714 consensus anchor positions were empirically calibrated against ABL1 (PDB 1OPL), SRC (1Y57), and EGFR (2ITX); all three structures place β3-K at HMM position 34, αC-E at 51, and DFG-D at 145, with HRD-D at 127 (two of three). A sequence-regex fallback derived from the Eswaran & Knapp consensus spacer is used if HMMER fails. A motif-override panel in the Classify tab right rail accepts user corrections for atypical kinases.

Pseudokinases. Pseudokinases often have one or more of the four anchors replaced (e.g. αC-Glu → Ser/Thr/Asn in HER3, JAK pseudokinase domains, STRADα). HMMER still anchors them if the surrounding sequence is canonical, but the geometric interpretation degrades — the OOD detector (section 4) is designed to surface this.

Methods · 03

3. How the model makes the call

The model combines a small neural network (trained on the geometry of 7,459 kinase chains where the answer was already known) with three independent published rules. The four answers are reconciled into a single consensus call. When they disagree, the model says so and lowers the confidence.

Source	What it uses	What it outputs	Reference
MLP classifier	16-d encoded feature vector	8-cluster soft probability (BLAminus, BLAplus, BBAminus, BABtrans, ABAminus, BLBminus, BLBplus, BLBtrans)	this work, trained on KinCoRe
Vijayan rule	2 Cα–Cα distances	"classical DFG-out" boolean (D₁ ≤ 7.2 Å AND D₂ ≥ 9.0 Å)	Vijayan et al., J. Med. Chem. 2015
Modi/Dunbrack rule	2 Cα→Cζ distances + Ramachandran region + χ₁ rotamer	3 spatial groups × 8 clusters	Modi & Dunbrack, Proc. Natl. Acad. Sci. U. S. A. 2019
Möbitz rule	F→G pseudo-torsion + αC-E to F distance	3-state coarse DFG (active / FG-down / G-down)	Möbitz, Biochim. Biophys. Acta 2015

Table 1. The four independent sources of evidence that feed into the consensus call.

Architecture, training, and consensus details

MLP architecture. 16 → 32 → 32 → 8, ReLU activations, softmax output over the eight Modi/Dunbrack clusters. ~1.4 K parameters. Exported to ONNX, runs in WebAssembly. Inference < 50 ms / chain.

Training. 7,459 chains from the Kincore annotated database (Modi & Dunbrack, Nucleic Acids Res. 2022). Stratified 5-fold cross-validation. Adam optimizer, lr=1e-3, batch=256, 100 epochs with early stopping at validation loss plateau.

Calibration. The raw softmax outputs are already well-calibrated — expected calibration error (ECE) = 0.0071 over the 7,459-chain KinCoRe set — so no post-hoc temperature scaling is applied and the reported soft probabilities can be read as approximate true frequencies.

Consensus. The reported DFG state is the modal call across the three rule-based criteria; ties default to the Modi/Dunbrack spatial assignment. The cluster label is the MLP softmax argmax. Disagreement between the MLP and the rules contributes a confidence penalty (see section 4).

Catalytic-activity flag. A separate boolean appears on the verdict card: "catalytically primed" iff DFG state is DFGin AND cluster is BLAminus AND αC-Glu Cβ ↔ β3-Lys Cβ distance ≤ 10 Å (salt-bridge feasible) AND no missing DFG-Asp/Phe side-chain atoms. Matches the Kincore-style criterion (Modi & Dunbrack 2019).

Vijayan, R. S. K.; He, P.; Modi, V.; Duong-Ly, K. C.; Ma, H.; Peterson, J. R.; Dunbrack, R. L., Jr.; Levy, R. M. Conformational Analysis of the DFG-Out Kinase Motif and Biochemical Profiling of Structurally Validated Type II Inhibitors. J. Med. Chem. 2015, 58, 466–479.

Modi, V.; Dunbrack, R. L., Jr. Defining a New Nomenclature for the Structures of Active and Inactive Kinases. Proc. Natl. Acad. Sci. U. S. A. 2019, 116, 6818–6827.

Möbitz, H. The ABC of Protein Kinase Conformations. Biochim. Biophys. Acta, Proteins Proteomics 2015, 1854, 1555–1566.

Methods · 04

4. How the model knows when to be uncertain

A confident-looking soft probability ("99% BLAminus") is not enough — neural networks happily extrapolate into nonsense when given out-of-distribution input. KinaseConf surfaces two independent reliability signals: a 0–100 confidence score compounded from four orthogonal sources, and an out-of-distribution tier (green / yellow / red) read from the Mahalanobis distance to the assigned cluster's training centroid.

Confidence component	Points	What it penalises
Method agreement	0 – 40	Disagreement between MLP, Vijayan, Modi, and Möbitz calls
Structure completeness	0 – 25	Missing side-chain atoms on the four anchor residues
B-factor quality	0 – 20	Mean B-factor of DFG-triad Cα atoms (penalties at > 50 Å² and > 80 Å²)
Vijayan boundary proximity	0 – 15	Distances within 0.5 Å of a Vijayan threshold (boundary cases)

Table 2. The 0–100 confidence score components. The four are orthogonal — they can flag different kinds of unreliability — so a perfect score requires the structure to be complete, well-resolved, and unambiguous AND to have all four methods agreeing on the call.

Out-of-distribution detection (Mahalanobis tiers)

For every prediction, a Mahalanobis distance is computed between the input feature vector and the empirical centroid of the assigned cluster in the training set, using the cluster-specific covariance matrix. Inputs below the 95th percentile of the in-cluster distance distribution are tagged in-distribution (green); inputs between the 95th and 99th percentiles are atypical (yellow); inputs above the 99th percentile (and above the global p99 threshold) are out of distribution (red, "verify manually").

Thresholds were calibrated on the held-out portion of the 5-fold cross-validation set. Across the full RCSB protein-kinase set (11,457 chains classified), 9,786 (85.4%) fall in the in-distribution tier under this rule.

Methods · 05

5. Scope and limitations

KinaseConf is a canonical-kinase classifier and assumes the four conserved kinase-domain anchor residues are present at their structurally canonical positions. The list below catalogues the principal cases in which it should not be applied without manual review.

What this tool does: assigns the DFG conformational state of a protein-kinase catalytic domain from a single structure. What it does not do: predict ligand binding, potency, or selectivity, and it is not a substitute for manual inspection of borderline cases. Reported performance and its evaluation conditions are given in the Validation tab.

Atypical kinases and pseudokinases. When one or more anchor residues are non-canonical (most commonly αC-Glu replaced by Ser, Thr, or Asn in pseudokinases), the out-of-distribution banner flags the reduced reliability. Users may specify motif residues manually via the override panel in the Classify-tab rail.
Type-I½ inhibitor classification. Type-I½ inhibitors bind a DFG-in conformation with αC-out geometry. The classifier scores these as a match only when the predicted cluster lies among the αC-out clusters (BLBplus, BLBminus, BLBtrans). Because the D₁ distance alone does not separate αC state across clusters (for example, BLBplus has D₁ ≈ 8.4 Å whereas BLAminus has D₁ ≈ 16.0 Å, opposite of the naïve expectation), accuracy in this class is structurally limited.
NMR ensembles. When multiple MODEL records are present, a model selector appears in the rail and classification re-runs per selection. Aggregate statistics across models are not summarised.
Modi & Dunbrack eight-cluster classifier. The full probabilistic XGBoost classifier reported in the original publication is replaced here by a deterministic Ramachandran-region and χ₁-rotamer rule augmented by the MLP softmax. The two approaches concur on 5,820 of 5,831 KLIFS-shared chains.
Möbitz twelve-state scheme. Collapsed here to the dominant pseudo-torsion call. The full twelve-state assignment is reserved for future work.
Single-chain analysis. Each kinase chain is classified independently. Inter-subunit contacts in heteromeric complexes (for example CDK–cyclin assemblies, RAF dimers, or JAK pseudokinase–kinase pairs) are not modelled.
Trajectory file formats. Multi-model PDB and CHARMM, NAMD, or X-PLOR DCD are parsed natively in the browser. GROMACS XTC and TRR formats currently require conversion to PDB using mdconvert or gmx trjconv.
Sequence-only prediction. KinaseConf requires atomic coordinates; sequence-only DFG-state prediction is not provided. Preliminary experiments with protein language-model embeddings (ESM2) did not reach the agreement threshold adopted for inclusion.

References

Vijayan, R. S. K.; He, P.; Modi, V.; Duong-Ly, K. C.; Ma, H.; Peterson, J. R.; Dunbrack, R. L., Jr.; Levy, R. M. Conformational Analysis of the DFG-Out Kinase Motif and Biochemical Profiling of Structurally Validated Type II Inhibitors. J. Med. Chem. 2015, 58, 466–479.
Modi, V.; Dunbrack, R. L., Jr. Defining a New Nomenclature for the Structures of Active and Inactive Kinases. Proc. Natl. Acad. Sci. U. S. A. 2019, 116, 6818–6827.
Modi, V.; Dunbrack, R. L., Jr. Kincore: A Web Resource for Structural Classification of Protein Kinases and Their Inhibitors. Nucleic Acids Res. 2022, 50, D654–D664.
Möbitz, H. The ABC of Protein Kinase Conformations. Biochim. Biophys. Acta, Proteins Proteomics 2015, 1854, 1555–1566.
Reveguk, I.; Simonson, T. Classifying Protein Kinase Conformations with Machine Learning. Protein Sci. 2024, 33, e4918.
Rahman, R.; Ung, P. M.-U.; Schlessinger, A. KinaMetrix: A Web Resource to Investigate Kinase Conformations and Inhibitor Space. Nucleic Acids Res. 2019, 47, D361–D366.
Carles, F.; Bourg, S.; Meyer, C.; Bonnet, P. PKIDB: A Curated, Annotated and Updated Database of Protein Kinase Inhibitors in Clinical Trials. Molecules 2018, 23, 908.
Eddy, S. R. Accelerated Profile HMM Searches. PLoS Comput. Biol. 2011, 7, e1002195.
van Linden, O. P. J.; Kooistra, A. J.; Leurs, R.; de Esch, I. J. P.; de Graaf, C. KLIFS: A Knowledge-Based Structural Database To Navigate Kinase–Ligand Interaction Space. J. Med. Chem. 2014, 57, 249–277.
Kanev, G. K.; de Graaf, C.; Westerman, B. A.; de Esch, I. J. P.; Kooistra, A. J. KLIFS: An Overhaul after the First 5 Years of Supporting Kinase Research. Nucleic Acids Res. 2021, 49, D562–D569.
Lin, Z.; Akin, H.; Rao, R.; Hie, B.; Zhu, Z.; Lu, W.; Smetanin, N.; Verkuil, R.; Kabeli, O.; Shmueli, Y.; dos Santos Costa, A.; Fazel-Zarandi, M.; Sercu, T.; Candido, S.; Rives, A. Evolutionary-Scale Prediction of Atomic-Level Protein Structure with a Language Model. Science 2023, 379, 1123–1130.
Jumper, J.; Evans, R.; Pritzel, A.; et al. Highly Accurate Protein Structure Prediction with AlphaFold. Nature 2021, 596, 583–589.

KinaseConf DFG-state classifier