henhaobio
Sign in

bioRxivBioinformaticsPreprint

AlphaFold2 models of the active form of all 437 catalytically-competent typical human kinase domains

Bulat FaezoviD, Roland L. DunbrackiD

PDFDOICodeData
In-depth full-text digest available · Jump to digest ↓

Abstract

AbstractHumans have 437 catalytically competent protein kinase domains with the typical kinase fold, similar to the structure of Protein Kinase A (PKA). Only 155 of these kinases are in the Protein Data Bank in their active form. The active form of a kinase must satisfy requirements for binding ATP, magnesium, and substrate. From structural bioinformatics analysis of 40 unique substrate-bound kinases, we derived several criteria for the active form of protein kinases. We include requirements on the DFG motif of the activation loop but also on the positions of the N-terminal and C-terminal segments of the activation loop that must be placed appropriately to bind substrate. Because the active form of catalytic kinases is needed for understanding substrate specificity and the effects of mutations on catalytic activity in cancer and other diseases, we used AlphaFold2 to produce models of all 437 human protein kinases in the active form. This was accomplished with templates in the active form from the PDB and shallow multiple sequence alignments of orthologs and close homologs of the query protein. We selected models for each kinase based on the pLDDT scores of the activation loop residues, demonstrating that the highest scoring models have the lowest or close to the lowest RMSD to 22 non-redundant substrate-bound structures in the PDB. A larger benchmark of all 130 active kinase structures with complete activation loops in the PDB shows that 80% of the highest-scoring AlphaFold2 models have RMSD < 1.0 [A] and 90% have RMSD < 2.0 [A] over the activation loop backbone atoms. Models for all 437 catalytic kinases are available at http://dunbrack.fccc.edu/kincore/activemodels. We believe they may be useful for interpreting mutations leading to constitutive catalytic activity in cancer as well as for templates for modeling substrate and inhibitor binding for molecules which bind to the active state.

In-depth summary

Claude Opus 5.5 · xhigh effort · 7 Nov 2026 · from the full text of version 2 (3 Sept 2023) · CC BY 4.0

With active-state templates and shallow ortholog alignments, AlphaFold2 models all 437 human catalytic kinase domains as active; 80% of 130 benchmarked kinases are under 1 Å activation-loop RMSD.

Only 155 of 437 human catalytic kinase domains have an active structure in the PDB, and 48% have an active model in the EBI AlphaFold2 set, yet active forms are needed to study substrate specificity, activating cancer mutations and drug development. Beyond the usual DFG and salt-bridge checks, the authors add substrate-binding criteria for both ends of the activation loop, derived from 40 substrate-bound complexes, and use them to pick templates and models.

Key results

  1. ShownFive criteria define an active, substrate-capable kinase: DFGin, BLAminus XDFG dihedrals, a formed Lys/Glu salt bridge, a backbone hydrogen bond between activation-loop residue 6 and the residue before HRD, and an APE9 Cα contact with the HRD-Arg carbonyl.

    DFG6/XHRD hydrogen bond (<3.6 Å) present in 386 of 391 ATP-bound BLAminus structures, 97% of BLAminus structures without ATP, and 24% of non-BLAminus ones. With the salt bridge formed, 99% of ATP-bound structures are BLAminus vs 19% when broken. APE9-hRd distance is 3.4 to 4.2 Å in Ser/Thr substrate complexes and 6.5 to 7.4 Å in TYR ones. · BLAminus conformation and Salt bridge formation; ActLoopNT; ActLoopCT; Active structures of catalytic kinases in the…

  2. ShownNo single combination of templates, alignment source and depth gave active models for every kinase, but pooling roughly 200 models per kinase did so for 435 of 437. OBSCN-2 needed an introduced D7929G mutation, and LMTK2 needed extra distillation templates from AATK and LMTK3.

    Active models per source: EBI AlphaFold2 database 208 of 437; default AlphaFold2 with UniRef90 281 (PDB70 templates) or 298 (no templates); each setup using ActivePDB or distillation templates with ortholog or family alignments 371 to 421; ActivePDB templates with family and ortholog alignments combined 435. About 90,000 models were generated overall. · Generation of active models of catalytic protein…; DISCUSSION

  3. ShownRanking models by their minimum pLDDT over the activation loop recovers active PDB conformations: the top-ranked model is usually the best or close to the best match, and for most of the 130 benchmarked kinases it is within 1 Å.

    For most of 22 substrate-bound kinases, the top min-pLDDT model had the best or close to best RMSD. Across 130 active PDB kinases, top models were under 1 Å backbone RMSD (N, Cα, C, O; after C-terminal domain superposition) for 104 (80%) and under 2.0 Å for 117 (90%); outliers include MAPK7 (5.74 Å), MAPK3 (5.1 Å) and MAP3K14 (3.97 Å). · Picking the best model with pLDDTs scores of the…; Benchmarking

  4. SuggestedFeeding AlphaFold2's own active models back as 'distillation' templates (ActiveAF2) raised confidence in the activation loop and sometimes improved accuracy, but these templates can also carry over loop conformations that do not resemble substrate-bound ones.

    In the 22-kinase benchmark, ActiveAF2 gave higher pLDDT for most kinases and markedly better models than ActivePDB for CDK2 and PIM1 (no test statistic shown). LMTK2 and LMTK3 became active only with distillation templates and 5 or fewer ortholog sequences. For MAPK1, ActiveAF2 templates reproduced the bulged alternate loop and ActivePDB the substrate-like one (both about 0.3 Å). · Generation of active models of catalytic protein…; Picking the best model with pLDDTs scores of the…; Data Input and Preparation

  5. SuggestedSome PDB structures that pass the criteria are likely not substrate-capable, and the models match substrate-bound relatives instead. Human SRC 1Y57, often used in molecular dynamics of active SRC, does not resemble substrate-bound TYR kinases, whereas chicken SRC 3DQW matches ABL1 2G2I.

    CDK2: best model 0.48 Å RMSD (min pLDDT 84.2) to substrate-bound 1QMZ, whereas the closest models to three other passing conformations were 1.09 to 2.78 Å with min pLDDT under 50. HIPK3 models resemble HIPK2 7NCF (activation loops 96% identical) rather than HIPK3 7O7I. Chicken and human SRC kinase domains differ at two residues, neither in the activation loop. · Picking the best model with pLDDTs scores of the…; Figure legends

Methods, data and reproducibility

Five active-state criteria were derived from 40 substrate-bound complexes and 391 ATP-bound BLAminus kinase structures in the PDB. AlphaFold2 was run on 437 domains with active PDB or AlphaFold2-derived templates and shallow ortholog or family alignments (5 to 90 sequences), keeping models that pass the criteria and ranking them by minimum activation-loop pLDDT (AlphaFold2's per-residue confidence).

  • Kinase-substrate and pseudosubstrate complexes in the PDB, including 11 crystal autophosphorylation complexes · multiple (includes human AKT1 and a plant kinase, Arabidopsis BAK1) · n = 40 unique complexes · PDB IDs in Table 1 (e.g. 4ekk, 2qvs, 3tl8)
  • Catalytic kinase structures bound to ATP, ADP or an ATP analog in the DFGin BLAminus state · n = 391 structures of 74 kinases
  • ActivePDB template set of active catalytic kinase structures · multiple (includes non-human kinases) · n = 165 kinase domains from 278 PDB entries
  • Human catalytic typical kinase domains modeled · human · n = 437 domains in 429 genes
  • Ortholog/close-paralog sequence sets from UniRef100 kinase-domain sequences (Pfam PF00069, PF07714), filtered at >50% identity and >90% coverage, then CD-HIT at 90% · n = 1.68 million UniRef100 sequences searched
  • Benchmark 1: substrate-bound PDB structures with complete activation loops · n = 22 kinases
  • Benchmark 2: active PDB structures with complete activation loops, with chicken SRC substituted for human SRC · human (one chicken SRC structure) · n = 170 structures of 130 kinases · e.g. PDB 3DQW, 1QMZ, 4MNE
  • AlphaFold2 models from the EBI database, used for comparison · human · n = 437 kinases, one model per protein
Methods & tools
AlphaFold2, AMBER relaxation, PSI-BLAST, CD-HIT, hh-suite, PyMOL (CE alignment), UniRef90, UniRef100, PDB70, Pfam (PF00069, PF07714), Kincore / Kincore-standalone2, AlphaFold-Multimer (exploratory)
Software & environment
Ubuntu 20.04, CUDA11, Python 3.8, TensorFlow 2.3.1; AlphaFold2 from the DeepMind GitHub repository with recommended settings, models 3 to 5 commented out in config.py, an added MSAlimit variable in pipeline.py, and mgnify, BFD, small BFD and uniref30 disabled.
Assay conditions & compute
Workstations with NVIDIA GeForce GPUs (8, 12 or 24 GB each); about 90,000 models generated in total; run times not reported.
Reproducibility gaps
The AlphaFold2 modifications (MSAlimit, swapped sequence and template databases) are described in prose only; no repository or scripts for the modeling pipeline are given, and run times are not reported.
Limitations (6)
  • Some conformations that pass the criteria may still block substrate binding, and no general rule was found to flag them; MAPK1, MAPK3 and MAPK7 share such an alternate loop in PDB structures. · stated by the authors
  • For some kinases the activation loop still takes several conformations after filtering, and more than one may support catalysis; the released set keeps the one most like substrate-bound structures. · stated by the authors
  • Kinases distantly related to the rest of the kinome, or with long activation loops, get low loop pLDDT; all three RAF kinases have minimum loop pLDDT below 40, and for BRAF it is unclear whether the PDB structure or the models are substrate-capable. · stated by the authors
  • AlphaFold-Multimer does not yet reliably model kinase-substrate complexes: settings that produce an active loop can disrupt substrate placement, and a robust protocol is still in development. · stated by the authors
  • The criteria, template sets and benchmarks were all defined by the authors, and disagreements with PDB structures (MAPK3, HIPK3, ACVR1) are settled by resemblance to substrate-bound relatives, not by binding or activity data. · noted in this summary
  • pLDDT ranking is not always right: for BUB1 and WNK3, lower-scoring models sit closer to the likely active structure than the top-pLDDT model, which is the one added to Kincore. · noted in this summary

Read this if you need active-state models of human kinases to interpret mutations or to model substrates and inhibitors, or want a recipe for steering AlphaFold2 to a chosen conformation with templates and shallow alignments.

Citations
45

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISGizzio '26Wang '26Al-Masri '22Modi '19Trozzi '24Herrington '23
97% semantic matchBioinformatics
Joan Gizzio, Bulat Faezov, Qifang Xu+1 more·bioRxiv·Feb 21, 2026· 1 citations