Small Molecule-Protein Interactions

1
Small Molecule-Protein Interactions

• Howard Feldman
• The Blueprint Initiative
• Toronto, Ontario
• hfeldman_at_blueprint.org

2
Drug Discovery Pipeline

• Most drugs are small molecules, and the
  interactions they make with proteins determine
  their effects, and toxicity, to the human body
• Clinical trials are most expensive part of the
  pipeline if failure can be predicted before
  this point, it saves time and money

3
Drug Discovery Pipeline

• It is of utmost importance to identify lead
  compounds in the early stages of drug discovery
  that will be most likely to succeed
• Recent study by Tufts Center for the Study of
  Drug Development showed that bringing one drug to
  market costs an average of 800M!
• 5/5,000 potential new drugs tested on animals
  reach clinical trials, and only one ultimately
  wins FDA approval

4
How small is a small molecule?

• Small molecule generally considered anything
  which may interact with protein-DNA
• Must be biologically relevant
• Examples include ions, polysaccharides, peptides,
  drugs

5
Small Molecules

• No absolute maximum size, though drug-like
  molecules often have molecular weight of 500 Da
  or less
• However can get complex branched poly
  saccharides, cyclic antibiotics, etc.
• Normally not interested in detergents, buffers,
  solvents, denaturants, non-biological ions

6
How Many?

• Recently a number of public small molecule
  databases have become available
• CAS Registry 26,000,000 substances
• Cambridge Structural Database 300,000 3D
  structures
• PDBSum 6700 3D ligands from PDB
• NCI databse 250,000 molecules
• NCBI PubChem 700,000 compounds
• ChemBank 1,100,000 molecules
• Problem data can be very messy, sparse

7
Popular small molecules and domains

• Not surprisingly, divalent cations and ATP are
  the most common small molecules found interacting
  with proteins
• AAA is an ATPase domain, the next three are all
  helicases, which bind various nucleotides as well

8
Toxicity

• Caused when drug interferes with biological
  pathway(s) in the host
• Less side-effects, the better
• Must be determined in early stages of discovery,
  or very costly
• Hence predicting toxicity is very important and
  desirable
• Boils down to predicting interactions, or rather,
  non-interactions

9
Predicting Toxicity

• Inverse docking Chen and Zhi developed a
  database of cavities in PDB structures
• INVDOCK searches cavities for potential
  interactions to ligand of interest, using scoring
  function
• Compare energy to absolute threshold, as well as
  energy of observed PDB ligand(s) at that site

10
Example 4H-Tamoxifen

• Used to treat breast cancer
• INVDOCK finds 22 putative protein targets at
  least 10 of which have some experimental backing
  (including the ones shown here)
• Estrogen receptor (the drug target)
• Alcohol dehydrogenase (enhances sedative effect
  of alcohol)
• IgG light chain (modulates immune response)
• 17b-hydroxysteroid dehydrogenase (tumor
  regression)
• GST (suppressed activity, genotoxicity,
  carcinogenicity)

11
Drug Docking
12
Drug Docking

• Shares much in common with structure prediction
• Two components
• Exploration of conformational space
• Scoring function
• Plus one additional component
• Locating the binding site

13
Drug Docking Level of Detail

• Rigid body docking protein remains fixed, small
  molecule has 6 degrees of freedom (DOF) 3
  translational and 3 rotational

14
Drug Docking Level of Detail

• Flexible-ligand docking protein remains fixed,
  small molecule has standard 6 DOF plus internal
  DOF can rotate about bonds
• More time consuming, but necessary for complex
  ligands if binding conformation is unknown
• Flexible docking as above, and in addition
  protein atoms in neighbourhood of binding site
  can move
• Largest conformational space to search
• Often done by using multiple static protein
  conformers, and treating each by flexible ligand
  docking
• Often important when docking to apo-protein e.g.
  allosteric effects

15
Drug Docking Level of Detail

• Some methods such as FlexX perform incremental
  construction within the binding pocket rather
  than docking per se

16
Drug Docking Techniques

• Drug docking algorithms share much with protein
  structure prediction, and include
• Monte Carlo search
• Molecular Dynamics
• Genetic Algorithms
• Fragment Assembly
• Tabu Search
• Many more

17
Drug Docking

• When ligand and target are known, can allow
  complete flexible docking
• For HTS, can usually only afford rigid body for
  initial pass
• Location to dock to on protein target may be
  known ahead of time, or may be computed through
  binding pocket detection
• Often binding site can be predicted if 3D
  structure is available using cavity-detection
  algorithms
• Search must be efficient, as with protein
  folding, since exhaustive search is not possible
• Scoring function must be selective and efficient

18
Drug Docking Example

• Study by Thorntons group (Nature Biotech. 22(8)
  (2004) p 1039-1045
• Took 120 enzymes and 125 metabolites from EcoCyc
  subset of 29 complexes have crystal structures
• Docked all-vs-all with AUTODOCK

19

• Energy plots for docking (a) and reverse docking
  (b) for subset of 29 with crystal structures
  triangles represent crystal complex
• Note from (a), enzymes are not that selective
  about substrate, nor are substrates that specific
  for enzyme in (b)

20
Drug Docking Example

• Computed P value ability of substrate or enzyme
  to recognize its partner based on energy
  distribution
• Now with 4 exceptions, the docked pairs show
  either enzyme OR substrate OR both are specific

21
Transition state
22
Transition state

• Most potent inhibitors are not substrate
  analogues but rather transition-state analogues
• Important to remember when screening compounds

23
Interaction Databases

• BIND (Protein-ligand interactions from PDB and
  literature, SLRI)
• Het-PDB Navi (Protein-ligand interactions from
  PDB, Nagahama Inst. Bio-Science)
• EcoCyc (metabolic pathways, SRI)
• KEGG (pathway database, Kyoto)

24
Blueprints Small Molecule Resources

• BIND-3DSM Division
• 23,584 Filtered Small Molecule Biopolymer
  interactions, automatically derived from crystal
  structures
• Biologically insignificant records removed (i.e.
  crystal packing, non-biological ions)
• Published Biopolymers. 2001-2002 61(2)111-20
• SMID
• 48886 records matching 4283 small molecules (from
  PDB structures) to 2807 protein families (CDD,
  SMART, PFAM)
• SMID-BLAST
• BLAST calibre tool to attach small molecule
  binding annotation (residue-level) to genomic
  sequence
• SMID-Genomes
• SMID-BLAST vs all completely sequenced genomes
• 9.6 Million high-quality small molecule
  interaction annotations mapped to sequences
• Database interface to browse/compare/investigate
  small molecule specificity across organisms

25
A 3DSM Record
www.bind.ca
26
BIND record binding site
27
Interaction Example

• Taxol is derived from natural products, and was
  discovered to be effective against certain types
  of cancer
• Interacts with tubulin and
• stabilizes tubules forming
• cell cytoskeleton, preventing
• mitosis and leading to cell death

28
Visualizing Binding Sites
29
SMID

• http//smid.blueprint.org/
• Small Molecule Interaction Database
• Matches small molecule binding sites in
  structures to protein domains in NCBI's Conserved
  Domain Database
• 4283 small molecules from PDB

30
Creating SMID Records
Start with an MMDB record (PDB record) containing
more than one molecule.
Small Molecule A (smA)
Protein A (ProA)
Small Molecule B (smB)
Find atoms from one molecule in proximity (0.5 Å)
of atoms from another molecule.
321
336
345
357
371
401
62
74
83
ProA

• Interactions Found
• Residues 62, 74 83 interacting with smA.
• Residues 321, 336, 345, 357, 371 401
  interacting with smB.

smA
smB
RPS-BLAST
BIND Records
31
Creating SMID Records
RPS-BLAST
123
31
44
62
73
86
105
98
RPS-BLAST all sequences found to interact with a
small molecule in order to obtain alignments with
conserved domains (DomA DomB).
DomA
DomB
321
336
345
357
371
401
62
74
83
ProA
Overlay small molecule protein interaction on
aligned conserved domains.
smA
SMID records made

• Interactions Found
• DomA (residues 98, 105 123) interacting with
  smA.
• DomB (residues 31, 44, 62, 73, 86 interacting
  with smB.

smB
123
31
44
62
73
86
105
98
DomA
DomB
smA
smB
32
Use Cases for SMID

• Domain Studies
• Binding site analysis
• Domain family binding site conservation
• Small molecule to the domain families that bind
• Structural Genomics
• Domain/ligand/binding site identification
• Some ligands go over domain boundaries
• Easier pattern recognition for interactions
• Quickly identify candidate co-crystalization
  ligands

33
Taxol ligand conservation in Tubulin/FtsZ domain
family
34
SMID-BLAST

• Uses RPS-BLAST (unmodified) with a new scoring
  scheme to improve domain family hits using
  specific ligand conservation information
• Validation - 1652 new unique interactions
  deposited into PDB
• 1027 (62) of these interactions are predicted
  within our selected ligand score cutoff
• Of these 262 (25) were top predictions
• This is very good, as the test set is not
  comprehensive
• we do not have a set of all possible ligands to
  each protein crystal structure
• we can only use exact small molecule matches (not
  similar molecules, e.g. ATP vs ATP-gamma-S)
• Specificity able to distinguish closely related
  Trp- and Tyr- aminoacyl-tRNA synthetases that hit
  the same protein domain families

35
Use Cases for SMID-BLAST

• Annotation of Newly Sequenced Genomes
• New enzyme discovery
• Rhodococcus genome
• William Mohn (UBC)
• Metabolic diversity
• PCB degradation
• Drug Docking
• Can help prioritize experiments
• Homology Modelling
• May help in template selection phase

36
SMID-BLAST Results Summary
37
Summary

• Understanding and cataloguing biopolymer-small
  molecule interactions is critical to the drug
  discovery process
• Drug docking can help explain toxicity and side
  effects, and can be useful in understanding the
  forces behind interactions
• Transition state analogues make the best
  inhibitors
• Tools like SMID-BLAST provide a simple, powerful
  way to predict what ligands may interact with a
  protein, and vice-versa