CONCERTS: Dynamic Connection of Fragments as an Approach to de Novo Ligand Design

1
CONCERTS Dynamic Connection of Fragments as an
Approach to de Novo Ligand Design
Creation Of Novel Compounds by Evaluation of
Residues at Target Sites
David A. Pearlman and Mark A. Murkco Vertex
Pharmaceuticals Incorperated Cambridge, MA
2
Outline

• Background
• Implementation
• HIV-1 aspartyl protease
• FK506 binding protein
• Conclusions

3
Previous Work CONCEPTS

• Active site is filled with atoms
• Run MD simulations, and form/break bonds
• Generates useful de Novo leads
• Limitations
• Difficult to incorporate charge models
• Slow convergence, especially for spacer regions
• Only 1 suggestion per cpu-intensive run

4
CONCERTS Implementation
Modified AMBER/SANDER 4.0 minimization/MD program

1. Active site is filled with user-defined fragments
2. Connection vectors are chosen for each fragment
3. Define a volume for a known protein of interest
4. Randomly orient fragments in defined volume
5. Fragment minimization and MD (two steps)
6. Start CONCERTS

5
CONCERTS Implementation
6
CONCERTS Improvements
CONCERTS has several improvements over CONCEPTS

• Fragments can inherently have charge
• Fragments span larger region of space don't have
  to worry about spacer regions
• Many suggested molecules can be built during a
  run
• Greater control over types of molecules generated

7
CONCERTS Testing
Begin testing CONCERTS on two targets using 3
types of basis sets

1. 1000 copies of peptide fragment
2. 700 copies of benzene, 1000 copies each of
   methane, ammonia, formaldehyde, and water
3. 300 copies each of ammonia, benzene, cyclohexane,
   formic acid, ethane, ethylene, formaldehyde,
   formamide, methane, methanol, sulfinic acid,
   thiophene, and water

8
HIV-1 AP, Results A

• 82 macrofragments were found
• 35 tetra-, 27 penta-, 17 hexa-, and 3
  hepta-peptides
• Reproduces backbone of JG-365, a sub-nM
  peptide-based inhibitor
• Good fit suggested start with this structure, and
  add amino-acid side chains

9
HIV-1 AP, Results A2

• Start with 10 copies of previous fragment and 150
  copies of each standard amino acid side-chain
• A side-chain was added to each of the six a
  carbons in every peptide seed
• Lowest energy result mimics known inhibitor quite
  well

10
HIV-1 AP, Results B

• 138 macrofragments were generated
• Combination of 4 fragments
• Reproduces backbone of JG-365, despite not being
  made from amino acids
• Bonus only one chiral center!

11
HIV-1 AP, Results C

• 151 macrofragments were generated
• Combinations of 4 fragments
• Not good agreement with backbone of JG-365
• However, places atoms in regions of space for all
  but one of the side chains of the drug!

12
FKBP-12, Results A

• A number of macrofragments were identified
• Mimics the binding core of nM inhibitor FK506
• Interesting that peptide fragments modeled a
  non-peptide inhibitor reasonably well

13
FKBP-12, Results B

• 122 macrofragments were generated
• Places atoms in regions occupied by FK506
• Unfortunately, a significant number of fragments
  falls at the edge or outside of the active site
• Contains zero chiral centers

14
FKBP-12, Results C

• 130 macrofragments were generated
• A majority were outside or on the edge of the
  active site
• Less concise than B set
• Contains several chiral centers

15
Sampling Issues Thoroughness

• How well does CONCERTS sample the conformational
  space available?

20 hexamer or larger macrofragments during
peptide run (A set) against HIV1-AP
16
Sampling Issues Energy Function

• Does the energy function used in CONCERTS have
  predictive qualities?

• HIV-1 AP
• Hydrogen bonds with protein residues
• Enb for Set A inhibitors

17
Conclusion

• CONCERTS works it generates inhibitors
• For two targets HIV-1 protease and FKBP-12
• Peptide fragments produce more structures that
  are similar to known inhibitors
• More fragment types lead to increased diversity,
  but often have less similarity to inhibitors
• However, could produce new lead structures
• Less diverse fragment sets results in greater
  convergence
• For targets with unknown inhibitors, multiple
  structures can be generated
• Identify trends or new leads for better modeling