Presented to the ZHAW Anisimova Group on 2018-05-24

1. ARMADILLO PROTEIN EVOLUTION &
BINDING PREDICTION
Spencer Bliven
Anisimova Journal Club
2018-05-24

2. ARMADILLO REPEAT PROTEINS

3. WHY ARMADILLOS? BIOLOGICAL INTEREST
 Roles in localization, protein transport, & more
 ß-catenin: cell adhesion, development
 α-importin: nuclear localization
 APC: tumor suppressor gene, linked to colorectal cancer
 Probably homologous to HEAT repeat family
 Slightly different structure
 huntingtin: disease, nerve signaling, protein transport
 ß-importin: nuclear localization
 phosphotases/kinases
 Ancient family (primarily eukaryotic, but predates metazoans)

4. WHY ARMADILLOS? PROTEIN-PROTEIN BINDING
 Bind peptides, so could be used like antibodies or
DARPins (therapeutics, biotech, assays, etc.)
 Bind extended chains
 Target disordered regions and termini
 Linear epitope, so much easier to design
 Modular binding
[5AEI]

5. ARMADILLO EVOLUTION
Armadillidium vulgare

6. TANDEM REPEAT EVOLUTION
 Duplications & fusions within a gene lead to tandem repeats
 Speciation and gene duplication lead to orthologs and paralogs
 Pattern of repeats tells us the sequence of evolutionary events

7. HEAT & ARM
Andrade MA, Petosa C, O'donoghue SI, Müller CW, Bork P. Comparison of ARM and HEAT protein repeats. J Mol Biol.
Academic Press; 2001 May 25;309(1):1–18.

8. ARM FAMILY
Gul, I. S., Hulpiau, P., Saeys, Y., & van Roy, F. (2017) Cellular and Molecular Life Sciences, 74(3), 525–541
∂-catenins & ARM formins
ß-catenin not with ∂-catenins
ß-importin
HEAT
(outgroup)
Catenin
beta-like
α-importin

9. LIMITATIONS OF PRIOR STUDIES
 Don’t model repeat evolution
 Either use full-length sequences (no support for copy variation) or single
repeats (inconsistent boundaries, repeats segregate differently between
species)
 No reconciliation between gene tree and repeat tree
 Older papers use limited species and sequences
 Inconsistent inclusion of HEAT repeats
MY APPROACH
 Detect repeats with TRAL (cpHMM)
 Alignment & tree inference with ProGraphML+TR
 Joint gene tree and repeat tree inference (future work)

10. TRAL
 Tandem Repeat Annotation Library
 Circularly permuted Hidden Markov Model (cpHMM) for tandem
repeat alignment
 Integrates repeat detection software
 Important for expanding analysis beyond ArmRP family
Schaper et al. (2015). TRAL: tandem repeat annotation library. Bioinformatics, 31(18), 3051–3053.
Schaper E, Gascuel O, Anisimova M. Deep conservation of human protein tandem repeats within the eukaryotes. Mol Biol Evol. 2014
May;31(5):1132–48.

11. DETECTED REPEATS BY SPECIES (GUL HMM)
Species ArmRP Proteins
Macrostomum lignano 170
Echinostoma caproni 163
Lingula anatina 125
human 107
zebrafish 107
scaled quail 100
tropical clawed frog 95
owl limpet 93
starlet sea anemone 93
Florida lancelet 90
Japanese sea cucumber 84
Schistocephalus solidus 84
Octopus bimaculoides 82
Biomphalaria glabrata 82
purple sea urchin 81
platypus 75
green sea turtle 75
Stylophora pistillata 75
Wild Bactrian camel 72
Amphimedon queenslandica 68
Number of Proteins
Numberofspecies
94 species

12. PROGRAPHML+TR
Szalkowski AM, Anisimova M. Graph-based modeling of tandem repeats improves global multiple sequence alignment. Nucleic Acids Res.
2013 Sep;41(17):e162–2.

13. OUTLOOK: EVOLUTION
 Improve Arm profiles based on structural searches
 MMTF-pySpark for rapid structural searches
 Finish phylogenetic reconstruction with ProGraphML+TR on diverse
species
 Joint gene-repeat reconstruction
 Analogous to joint species-gene tree inference (e.g. Szöllosi et al, 2015)

14. ARM BINDING

15. MOTIVATION
 Nature’s solution to binding
molecules
 Used in diagnostics,
therapy, labelling,
biochemistry research
 $105 billion industry (2016)
 3D epitope
 Produced in vivo in
animals (polyclonal) then
optimized biochemically
(monoclonal)
Antibodies

16. MOTIVATION
 Nature’s solution to binding
molecules
 Used in diagnostics,
therapy, labelling,
biochemistry research
 $105 billion industry (2016)
 3D epitope
 Produced in vivo in
animals (polyclonal) then
optimized biochemically
(monoclonal)
Antibodies DARPins
 Designed Ankyrin Repeat
Proteins
 Developed by Andreas
Plückthun, UZH
 Commercialized by
Molecular Partners AG
($571 million market cap)
 Similar uses to antibodies
 3D epitope
 Produced in vitro from a
randomized library

17. MOTIVATION
 Nature’s solution to binding
molecules
 Used in diagnostics,
therapy, labelling,
biochemistry research
 $105 billion industry (2016)
 3D epitope
 Produced in vivo in
animals (polyclonal) then
optimized biochemically
(monoclonal)
Antibodies DARPins dArmRP
 Designed Ankyrin Repeat
Proteins
 Developed by Andreas
Plückthun, UZH
 Commercialized by
Molecular Partners AG
($571 million market cap)
 Similar uses to antibodies
 3D epitope
 Produced in vitro from a
randomized library
 Designed Armadillo Repeat
Proteins
 Bind extended peptides
(tails, disordered regions,
denatured proteins)
 1D epitope
 Rationally designed in
silico?

18. ARM STRUCTURE & CONSERVATION
Gul 2017 Fig 1B
Structure: Repeat from designed ARM YIIIM5AII (Hansen…Plückthun, 2016) [5aei], colored and labeled as in the alignment
H1
H2
H3 H1 H2 H3
Hydrophobic core

19. BINDING HINTS FROM DARMRP ((KR)N BINDING)
Gul 2017 Fig 1B
Structure: Repeat from designed ARM YIIIM5AII (Hansen…Plückthun, 2016) [5aei], colored and labeled as in the alignment
H1
H2
H3
Nonspecific
binding
Mutants available for 7 residues in Arg pocket
Lys pocket has only one specific interaction
H1 H2 H3
Hydrophobic core

20. BINDING MODULARITY
 For dArmRP, binding is linear with the number of repeats and for
single-residue mutations
Predictable binding energies
Single-residue resolution
K->A
R->A
2K->2A
2R->2A

21. KERNEL MODEL
 Regression problem: predict binding affinity from sequence at 7
positions
 Extract 5 features based on amino acid properties (Atchley 2005)
 Use linear regression with various kernels
log10 𝑌 = 𝐾 𝐾 + 𝜆𝐼 log10 𝑌
 Linear kernel 𝑎, 𝑏 = 𝑎 𝑇 𝑏
 Gaussian kernel 𝑎, 𝑏 = 𝑒𝑥𝑝 −𝜎 𝑎 − 𝑏 2

22. RESULTS
 Train on 138 datapoints from Plückthun group
 Essentially all “positive” binding cases
 Leave-one-out cross validation for error estimation
 Linear: 0.42 standard error (log10 M units)
 Gaussian: 1.42, but numerically instable

23. LINEAR KERNEL
lambda=0.001
0.42 standard error (log10 M units)
R=.90
Measured Binding (log10)
PredictedBinding(log10)

24. GAUSSIAN KERNEL
lambda=10-4 sigma=108
1.42 standard error (log10 M units)
R=.13
Measured Binding (log10)
PredictedBinding(log10)

25. GAUSSIAN KERNAL
 Numerically unstable implementation (hat matrix is near-singular)
 No renormalization currently

26. OUTLOOK: BINDING
 Switch from regression to classification
 Additional training data from collaborators
 In particular, need non-binding examples
 More sophisticated classifiers
 Numerically stable implementation
 Better kernels?
 Proactively suggest informative instances for our collaborators to
measure

27. THANKS!
 ACGTeam: Maria Anisimova, Manuel Gil, Victor Garcia,
Lorenzo Gatti, Max Maiolo, Simone Ulzega, Erich
Zbinden
 Matteo Delucci & Lina Naef (ACLS masters) – TRAL
 Elke Schaper – TRAL
 Somayeh Danafar – Kernel methods
 Andreas Plückthun, Patrick Ernst, Yvonne Stark (UZH) –
Binding data
But wait, there’s more! MMTF format coming next…

28. MMTF DEMO?

29. MMTF
 Compression: data normalization, vectorization, run-length encoding,
delta encoding
 Optional lossy/course representation
 Now has widespread software support (BioPython, BioJava, most
molecular viewers, etc)

30.  MMTF Format: http://mmtf.rcsb.org/
 MMTF-Spark library:
 Java https://github.com/sbl-sdsc/mmtf-spark/
 Python https://github.com/sbl-sdsc/mmtf-pyspark/
 Fast, parallelized whole-PDB analysis

31. CE-SYMM OPEN REPEAT DETECTION
ß-catenin [1I7X]

32. This work is licensed under a Creative Commons Attribution-ShareAlike 3.0 Unported
License.