Reading cross-reactivity off the kinase pocket

Imatinib was designed against one kinase, BCR-ABL1, and also inhibits half a dozen others (KIT, PDGFRA, CSF1R, LCK) well enough to matter clinically. Every human kinase folds the same ATP-binding pocket around the same three landmarks, so a drug built to fit one kinase’s pocket often fits several. Across the human kinase family, we:

list the kinases and fetch their sequences
align the kinase domain to Pfam’s HMM and build a tree
find the landmark columns from ABL1 and band them over a sequence logo
compare the gatekeeper residue between imatinib’s targets and the CDKs
check the gatekeeper tally against KLIFS

Prerequisites

curl
python3
docker, to run HMMER and FastTree as biocontainers (quay.io/biocontainers/hmmer, quay.io/biocontainers/fasttree); both are also on apt and brew

The data

UniProt supplies the kinase list and the sequences, and InterPro serves Pfam’s kinase domain model. Nothing to download: the build script fetches these.

What it is	File
UniProt’s own classification of the human and mouse kinomes, release 2026_03	pkinfam.txt
One UniProt sequence per accession through the `/stream` endpoint, which the script batches 50 accessions to a request, for example	uniprotkb-stream
The Pfam protein kinase domain HMM, PF00069, as InterPro serves it	PF00069

The files this page writes, hosted for the figure links

What it is	File
The alignment the figures below load	kinase-pocket.afa
Its tree	kinase-pocket.nwk
And its row metadata (kinase group, UniProt accession)	kinase-pocket-metadata.json

1. List the human kinases

UniProt’s pkinfam.txt groups every human and mouse protein kinase by family. A few lines of Python keep the human entries and their groups:

import re

lines = open('pkinfam.txt').read().splitlines()
group = None
rows = []
i, n = 0, len(lines)
while i < n:
    line = lines[i]
    if re.fullmatch(r'=+', line.strip()) and i + 2 < n and re.fullmatch(r'=+', lines[i + 2].strip()):
        group = lines[i + 1].strip()
        i += 3
        continue
    m = re.match(r'^(\S+)\s+(\S+_HUMAN)\s+\((\w+)\s*\)', line)
    if m:
        gene, entry, acc = m.groups()
        rows.append((gene, acc, group))
    i += 1
print(len(rows))

The 512 entries fall in ten named groups (AGC, CAMK, CK1, CMGC, NEK, RGC, STE, TKL, Tyr, Other) plus six small “Atypical” families with a different fold. This page writes Tyr, the tyrosine kinases with ABL1 among them, as TK.

2. Fetch the sequences

The script fetches one UniProt sequence per accession, 50 accessions per request:

import urllib.parse
import urllib.request

query = ' OR '.join(f'accession:{a}' for a in accessions[:50])
url = 'https://rest.uniprot.org/uniprotkb/stream?query=' + urllib.parse.quote(query) + '&format=fasta'
with urllib.request.urlopen(url) as resp:
    fasta_text = resp.read().decode()

The requests return 512 full-length records for the 512 accessions.

3. Align the kinase domain

hmmalign against Pfam’s Pkinase model (PF00069) extracts the catalytic domain, leaving the regulatory regions, SH2/SH3 modules and transmembrane spans behind, and puts every copy in the same 262 columns:

curl -sf "https://www.ebi.ac.uk/interpro/api/entry/pfam/PF00069/?annotation=hmm" |
  gzip -dc > pf00069.hmm

# score every candidate first, and keep only what clears Pfam's own gathering
# threshold. That drops the atypical, non-ePK-fold kinases (PI3/PI4-kinase,
# ADCK, alpha-type...) that would otherwise align to noise
docker run --rm -v "$PWD:/work" -w /work quay.io/biocontainers/hmmer:3.4--h7d74f8d_5 \
  hmmsearch --cut_ga --tblout hmmsearch.tbl -o hmmsearch.out pf00069.hmm all.fasta

docker run --rm -v "$PWD:/work" -w /work quay.io/biocontainers/hmmer:3.4--h7d74f8d_5 \
  hmmalign --trim --outformat afa -o kinase.afa pf00069.hmm kept.fasta

474 of the 512 entries clear the gathering threshold, and 30 of the 38 dropped belong to the six “Atypical” families. --trim removes the unaligned tails, and a following step strips the lowercase insert-state characters hmmalign leaves, so all 474 rows come out at the HMM’s 262 columns.

All 474 kinase domains beside their tree, with vertical bands at the columns the kinases share and ragged stretches at loops that differ from kinase to kinase.

4. Build a tree

FastTree builds the tree:

docker run --rm -v "$PWD:/work" -w /work quay.io/biocontainers/fasttree:2.2.0--h7b50bb2_1 \
  FastTree kinase-pocket.afa > kinase-pocket.nwk

Total time: 13.70 seconds  Unique: 474/474  Bad splits: 0/471

Clicking a tip opens a dialog with that row’s UniProt accession and kinase group from kinase-pocket-metadata.json.

5. Find the pocket landmarks

The ATP pocket has three landmarks: a catalytic lysine that anchors ATP’s phosphates, a gatekeeper that sets the room in the back pocket, and a DFG motif that switches the kinase between active and inactive states. In ABL1 numbering they are K271, T315 and D381-F382-G383, so walking the ABL1 row gives each landmark’s column:

full_abl1 = uniprot_sequence[metadata['ABL1']['uniprot']]
aligned_abl1 = alignment['ABL1']
residues = [(col, c) for col, c in enumerate(aligned_abl1, 1) if c != '-']
pos = full_abl1.find(''.join(c for _, c in residues[:8]))
abl1_number = {}
for col, c in residues:
    pos = full_abl1.index(c, pos)
    abl1_number[col] = pos + 1
    pos += 1

alignment and uniprot_sequence map a kinase’s name and accession to its sequence, and metadata is kinase-pocket-metadata.json. The script prints:

ABL1: 1130 aa, 262 alignment columns
catalytic K: column 30, ABL1 K271
gatekeeper: column 77, ABL1 T315
DFG: column 141-143, ABL1 D381 F382 G383

Column 102, outside the pocket, is the control:

The sequence logo over the alignment, with the catalytic K (blue), gatekeeper (red), DFG (green) and control (grey) columns banded. The catalytic lysine column is 93.5% K, and the control column, between the gatekeeper and the DFG motif, has no majority residue: F is the most common at 13.3%, then S, K, R, L and H.

ggmsaview reads the same hosted alignment and tree, and geom_msa_highlight() takes the four columns as a data frame. w opens the viewer in RStudio or a notebook, and msaview_plot() draws the logo above the rows as a ggplot2 figure:

library(ggmsaview)
pocket <- data.frame(
  start = c(30, 77, 141, 102),
  end = c(30, 77, 143, 102),
  label = c("catalytic K", "gatekeeper", "DFG", "control"),
  color = c("#1565c040", "#e3242b40", "#2e7d3240", "#75757540")
)
w <- msaview(
  msa = "https://gmod.org/JBrowseMSA/demo/data/kinase-pocket/kinase-pocket.afa",
  tree = "https://gmod.org/JBrowseMSA/demo/data/kinase-pocket/kinase-pocket.nwk",
  color_scheme = "clustalx_protein_dynamic", row_height = 2
) + geom_msa_highlight(pocket)
w
ggplot2::ggsave("kinase-pocket-logo-r.png",
  msaview_plot(w, tracks = c("conservation", "sequence-logo")),
  width = 12, height = 8)

All 474 rows over the 262 Pkinase columns, the tree beside them, under the conservation bars and the sequence logo. The logo stands tall at the catalytic K, at the DFG motif and in the HRD loop before it, and stays low at the control column. row_height = 2 leaves room in the figure for the tracks.

def tally(col, top):
    counts = Counter(row[col - 1] for row in alignment.values())
    return ', '.join(f'{r} {n} ({100 * n / total:.1f}%)' for r, n in counts.most_common(top))

print(f'catalytic K, column 30 over {total} rows: {tally(30, 3)}')
print(f'control, column 102: {tally(102, 8)}')

catalytic K, column 30 over 474 rows: K 443 (93.5%), - 15 (3.2%), C 4 (0.8%)
control, column 102: F 63 (13.3%), S 50 (10.5%), K 49 (10.3%), R 46 (9.7%), L 30 (6.3%), H 29 (6.1%), Y 27 (5.7%), D 26 (5.5%)

6. Compare the gatekeeper across two families

A small gatekeeper (threonine) leaves the back pocket open, and a bulkier one (phenylalanine, methionine, leucine) fills it. Two bands of the tree show column 77: the ABL band, and the CDK band, a family imatinib does not inhibit:

The gatekeeper column in the ABL band, where ABL1, ABL2 and CSK read threonine and MATK methionine.

The gatekeeper column in the CDK band, where CDK1, CDK2 and CDK3 read phenylalanine, as do the flanking CDK16 and CDK17.

The script prints the residue each named kinase has in this column:

ABL1=T ABL2=T KIT=T PDGFRA=T CSF1R=T LCK=T
MATK=M CSK=T
CDK1=F CDK2=F CDK3=F CDK16=F CDK17=F

7. Check the gatekeeper against KLIFS

The script tallies the gatekeeper column across all 474 rows:

counts = Counter(row[76] for row in alignment.values())  # column 77, 0-indexed
for r, n in counts.most_common(5):
    print(f'{r}  {n:3d}  {100 * n / total:.1f}%')

M  185  39.0%
T   90  19.0%
L   76  16.0%
F   69  14.6%
V   15  3.2%

Threonine is the gatekeeper in 19.0% of these human kinase domains. KLIFS, which curates an 85-residue pocket from solved structures, puts the same position at 18.4% threonine over its 521 human kinase entries. KLIFS’ terms do not allow redistribution, so the script queries its API live (https://klifs.net/api/kinase_information?species=Human) and prints KLIFS: 96 of 521 human pockets read T at residue 45 (18.4%).

Reproduce it end to end

build_kinase_pocket.sh runs the steps above in order:

It reads UniProt’s kinase classification, keeps the human entries with their groups, and fetches their sequences.
HMMER keeps the sequences that clear Pfam’s Pkinase gathering threshold and aligns their kinase domains to the HMM’s 262 columns.
FastTree infers a tree, and the script writes each kinase’s group and UniProt accession as row metadata.
It prints the landmark columns, the logo shares and the gatekeeper tally, then queries KLIFS for the closing comparison.

It needs the tools under Prerequisites.

curl -O https://raw.githubusercontent.com/GMOD/JBrowseMSA/main/docs/tutorials/scripts/build_kinase_pocket.sh
bash build_kinase_pocket.sh

With no arguments the script writes kinase-pocket.afa, kinase-pocket.nwk and kinase-pocket-metadata.json in the working directory.

References

Manning G, Whyte DB, Martinez R, Hunter T, Sudarsanam S. The protein kinase complement of the human genome. Science 298:1912-1934 (2002).
Blum M, et al. InterPro: the protein sequence classification resource in 2025. Nucleic Acids Research 53:D444-D456.
Eddy SR. Accelerated profile HMM searches. PLoS Computational Biology 7:e1002195 (2011).
Price MN, Dehal PS, Arkin AP. FastTree 2: approximately maximum-likelihood trees for large alignments. PLoS ONE 5:e9490 (2010).
Kooistra AJ, Kanev GK, van Linden OP, Leurs R, de Esch IJ, de Graaf C. KLIFS: a structural kinase-ligand interaction database. Nucleic Acids Research 44:D365-D371 (2016).