Designing macrocycles with Schrödinger tools
Tools for conformational search, permeability prediction, enumeration, conformer stability calculation, docking,
and free energy calculations of macrocycles
Macrocycles in drug design
- Macrocycles are pervasive (63 on market, 35 in development)1
- Cyclization can expand Ro5 restrictions2
- Cyclization can infer conformational stability of known binder (linear compound, PPIs)
1. Giordanetto, F., & Kihlberg, J. (2014). Journal of Medicinal Chemistry, 57(2), 278–295. http://doi.org/10.1021/jm400887j
2. Villar, E. A. et al (2014). Nature Chemical Biology, 10(9), 1–10. http://doi.org/10.1038/nchembio.1584
Challenges in macrocycle design
- Difficult synthetically- Cyclization reactions can be difficult, time-consuming
- High MW can be burdensome
- Computation is difficult- Cyclization impedes sampling
- Force field parameterization requires “unphysical” twisting of ring torsions
- Chemical software not typically designed for large flexible rings
- Design workflows must be adapted for macrocycle caveats
Schrodinger’s macrocycle design solutions
1) Requires StructureBased ADME license2) Requires Glide & Prime Licenses
Schrodinger is expanding the limits of in silico macrocycle design
- Prime macrocycle conformational sampling
- Passive membrane permeability1
- Docking with Glide2
- Bioactive conformer stability- Free Energy Perturbation with FEP+
Macrocycle Sampling
Sampling: Attributes of a functional sampler
Sampling algorithm designed to efficiently explore macrocycle conformational space, especially major ring conformations
- Accuracy- Match experimental conformations
- Speed- Fast enough to enable workflows
- Quick turnaround during ideation
- Diversity- Sample not only crystal conformation, but
also permeable, solvent, exploitable
Speed
Accuracy
Diversity
PrimeMCS algorithm
Recognize macrocyclic topology,assign bond rotamers from library
Break* major ring, combinatorially sample “half-loops” using bond rotamers.
Find pairs of half-loops which form reasonable closed rings. If there’s enough, cluster them, otherwise, go back to 2 with finer resolution
Build out r-groups from clustered ring backbones using bond rotamers. Minimize to relax backbone and repeat once.
Algorithm details of note
Originally developed @ Prof Matt P. Jacobsons Lab based on loop sampling algorithm
In step 2, we cut the ring 10 different ways around the ring (spinroot 10), enabling 10x parallelism
Cross-links are accounted for via inverse clash restraints
Only one independent ring system is thoroughly sampled
Test sets and protocols for evaluation
Sindhikara et al, JCIM, 2017 (DOI: 10.1021/acs.jcim.7b00052)
• Test set– 208 diverse crystal structures were curated from
• The Cambridge Structural Database (CSD, 130 macrocycles)
• Biologically Interesting Molecule Reference Dictionary (BIRD, 18 macrocycles)
• The Protein Databank (PDB, 60 macrocycles)
• Protocol comparison– Prime Macrocycle Conformational Search (PrimeMCS)
– Macromodel Macrocycle Baseline Search (MMBase)
– Moe Low Mode MD (Moe-LMMD)
– Desmond Molecular Dynamics (MD, 24ns)
Two ways to measure backbone structure
Backbone RMSDAssuming that backbone sampling is the difficult part. Here we measure the best backbone RMSD of a structure in the ensemble to the crystal reference
Backbone Torsional FingerprintUsing torsional scanning profiles, we can bin torsions into states. The superposition of all backbone states (torsional fingerprint) can be used to compare discretized conformations.
a b c
Measuring accuracy
Proportion of compounds matching RMSD threshold
PrimeMCS finds < 1.0Å structure 90% of the time, worst outlier 2.1Å
Number of EXACT matches to crystal backbone fingerprint.
PrimeMCS most often finds exactly correct structure
Measuring Diversity
Left:The number of output conformations compared to the number of UNIQUE output conformations.PrimeMCS consistently yields vastly more unique output conformations
Right:Difference in radius of gyration of largest vs
smallest conformer in the output ensemble (Span of
rgyr). PrimeMCS most often yields a large range in
conformer sizes.
Speed
Calculation times on serial CPU architecture
PrimeMCS calculations are much faster than the methods compared to here, usually taking less than 10 minutes per compound in serial.
Further, the jobs are 10x parallelizable, enabling extremely fast turnaround times.
Ser
ial
Targeted sampling with PrimeMCS
Default PrimeMCS works well across diverse macrocycles
For focused sampling many options are available:
- Ring shape constraints- Restrain by SMARTS*- Fix bonds by atom number*- Return more conformers*- Manual rotamer specification** Command-Line Only Implementation in 17-2 Beta
Green: Input Conformation (ring only)
Blue: Unrestrained PrimeMCS SamplingRed: Ring Sampling Restrained Around Input Conformation
Macrocycle Solutions
Macrocycle Solutions utilizing Prime Macrocycle Sampling
Cyclization of linear ligands (Beta) - Linker enumeration with chem_enumerate* - Rapid Bioactive Conformer Stability
calculations** with macrocycle_stability- FEP+ to compare affinity of macrocycles to
linear molecules
Predicting permeability- Prime structure-based membrane permeability
predictor- Prime energy visualizer
Determining binding mode (Beta)- Macrocycle Docking using Glide
Optimizing macrocyclic leads (Beta)- Bioactive Conformer Stability calculations- FEP+ to compare affinity between macrocyclic
design ideas
* Command-line only** Command-line only, GUI Coming in 17-4
Predicting Stable Cyclizations of Linear Ligands
Linker Enumeration (Command-Line Only)
1. Prep known linear structure
2. Specify linker substitutions
3. Define substitution rules
4. Enumerate (chem_enumerate.py)
- ['', O]- [C,CC,CCC,CCCC,CCCCC,CCCCCC,CCCCCCC]- ['', O]
Enumerate linkers to optimize linker lengths and chemistries.
Bioactive conformation is one of many low
energy Conformation
Bioactive Conformer Stability (Command-line only, GUI will be available in 17-4)
When macrocyclizing a linear ligand with a known bioactive conformation, the macrocycle stability workflow can quickly screen compounds based on whether they can access the known bioactive conformation
Stablemacrocyclized
compound
Bioactive conformation is
inaccessible
Reference linear compound
Retrospective design by stabilization:Macrocyclic BACE inhibitors from linear inhibitors
Known active
Here we test Schrodinger macrocycle design tools reproducing a successful macrocyclization of a known linear binder.
Huang Y, Strobel ED, Ho CY et al. Macrocyclic BACE inhibitors: optimization of a micromolar hit to nanomolar leads. Bioorg. Med. Chem. Lett. 20(10), 3158–3160 (2010).
Idea 1 Idea 2 Idea 3
Idea 4 *linear intermediate
BACE inhibitor stabilization
Prime energy vs RMSD of Bioactive SMARTS
Macrocycle
MacrocycleMacrocycleMacrocycle
1) Each ligand is sampled using PrimeMCS to find it’s ensemble of low energy structures
2) A heavy atom RMSD is calculated of each member of the ensemble to the interacting portion of the known active liganda) Requires the bio-active conformation
of the reference ligand to be known3) RMSD and relative energy are plotted.
Conformations that are both stable and similar to the known active are found in the bottom left of the plot.
4) A Boltzmann weighting of the RMSD over the ensemble allows us to locate ligands that are stable in the bioactive conformation without being able to access to many other unproductive conformations.
BACE inhibitor stabilization
Prime energy vs RMSD of bioactive SMARTSScatter plot of energies in output ensemble suggest macrocycles maintain smaller RMSD, but difficult to distinguish.
Macrocycle
MacrocycleMacrocycleMacrocycle
Expected RMSD of bioactive region For macrocycles, the Boltzmann-weighted expectation value (macrocycle_stability.py) orders compounds identically as experimental affinity.
Macrocycle FEP+
Macrocycle FEP+
• Core-hopping technology (soft bond scaling) has enabled macrocyclization reactions in FEP+• Macrocycles are automatically detected
and run with optimal scaling parameters
Scaled soft bond potential (above) enables smooth macrocycle bond formation across FEP+ lambda schedule
Yu H. et al, in preparationBonds can be formed even across long distances
Macrocycle FEP+ retrospective study results
• Seven retrospective cases of macrocyclization• ∆∆G
– MUE: 0.69
– RMSE 0.98
Macrocycle FEP+ currently achieves accuracy on par with small molecule FEP+
Yu H. et al, in preparation
Predicting Macrocycle Binding Modes
Integrating Prime Macrocycle Sampling with Glide
PrimeMCSEfficiently samples macrocycle ring
conformationsDoesn’t directly account for environmental
effectsGlide dockingEfficiently accounts for receptor environmentRelies on templates for ring conformations
Glide “macrocycle mode” Generates ring templates using PrimeMCS “on-the-fly” for contextual sampling
Sampling and docking parameters optimized for macrocycle docking
Integrating PrimeMCS into Glide
- Glide uses a filtering-based workflow to go from conformers to poses
- Ring templates generated via PrimeMCS to be utilized in the initial core conformer ensemble generation (confgen)
- In “Macrocycle mode” we use an “expanded funnel” to reflect the additional conformational and pose complexity added by macrocycles
Design tools: PrimeMCS-integrated Glide docking
We’ve integrated PrimeMCS sampling into a Glide “macrocycle mode”
- Appropriate Prime sampling, ring template generation and sidechain sampling, etc, all done automatically when mode is activated
PrimeMCS-integrated Glide docking results (2017-2)
Dataset contains 67 cocrystallized macrocycles, median 16 backbone atoms
PrimeMCS-based Glide docking found top poses under 2.0Å 67% of the time, significantly better than with rigid rings, but not as good as docking the native conformation.
1d4k 1.1Å docked structure
Native conformation
Rigid ring PrimeMCS-based
% Top pose under 2.0Å
91% 43% 67%
Median Serial CPU time
~1m ~1m 28m
Predicting Macrocycle Permeability
Developing a “Global” Permeability Model
• Predictor methods can lose significant signal when applied to noncongeneric sets
• A highly consistent experimental assay is required– PAMPA and CaCo2 Permeability assays across
multiple labs can be inconsistent up to 3 log units
• RRCK cell assay relatively high consistency R2~0.8• Di et al Journal of Pharmaceutical Sciences,
Vol. 100, 4974–4985 (2011)
• Leung + Jacobson propose modified global partition model with volume term
• Model is parameterized to RRCK data, but has similar correlation to Caco2 and PAMPA permeabilities as well as long as all data points were acquired in the same lab using the same procedure
Modeling passive permeability
Charged in solvent Neutralized in solventMembrane
conformation in solvent
Membrane conformation in
membrane
Assumption: Neutralization, transfer, and volume are rate limiting
Workflow
LigPrep, Epik
Prime Sampling
(Low dielectric)
Prime Energy(High
dielectric)
RRCK Permeability Model Applied to RRCK Data
R2 0.49
RMSE (log10(cm/s)) 0.42
R2 (Small Mol) 0.41
RMSE (SM) 0.42
R2 (macrocycles) 0.72
RMSE (macrocycles) 0.28
Refs* 1 2 3 4 5 6 7 8*see accessory slides
Symbol size by volume
Ringed circles • are macrocycles
178 Small molecules + 23 macrocycles = 201 compounds
RRCK Predictor results
“OK” permeability “Good” permeability “Great” permeability
(cm/s) (cm/s) (cm/s)
Accuracy/Recovery: 0.84/0.98 Accuracy/Recovery: 0.76/0.87 Accuracy/Recovery: 0.75/0.66
Macrocycle Tools Summary
Summary of Schrödinger macrocycle design tools
- Schrodinger tools offer comprehensive capabilities for macrocycles- Though tools are state-of-the-art, scientific and technical development is
ongoing- Prime-MCSs sampling enables many workflows
- Macrocycle bioactive conformer stability calculations
- Docking within Glide
- Membrane permeability predictions
- Look forward to more GUIs, additional workflows, and enhanced FEP+ reliability for macrocycles
Acknowledgements
- Sampling- BMS
- Shana Posy
- Steve Spronk
- Dan Cheney
- UCSF
- Matthew P Jacobson
- Siegfried Leung
- Schrödinger
- Dan Sindhikara
- Ken Borrelli
- Tyler Day
- FEP+- Haoyu Yu- Lingle Wang- Robert Abel- Yuqing Deng
- Docking- Ivan Tubert-Brohman
- Prime- Ed Miller
- Infrastructure- Tor Colvin- Dan Nealschneider
Supplementary slides below
Cyclic Peptide Stability
Preliminary Results
Design concept: Cyclic peptide stabilization
dE
Conserved SMARTS RMSD
Known activeProject to mutate LARGE cyclic peptide while maintaining binding
- Binding residues identified by crystallography
- Stabilization using default Prime-MCS reveals sampling is insufficient! (see image right)
- No apparent low RMSD low E structures!
Focusing sampling for large cyclic peptide
dE
Conserved SMARTS RMSD Conserved SMARTS RMSD
dE
Active Inactive
- Adding ring shape constraints, increased output and extra spinroot. Median 20 cpu hours (45 min wallclock).
- Scatterplots show clear differentiation between active and inactive stability
Large cyclic peptide retrospective prediction results
Results of macrocycle stability calculation on large cyclic peptide with modified sampling options.
Stability metric, P(RMSD < 1.5A), clearly identifies some true positives with no false positives with high cutoff
However there is a large false negative rate and some false positives with a low cutoff.
False negative: Worst case appears that shape constraint still too loose.
False positive: Worst case appears that ”nonconserved” region clashing with protein (not taken into account in metric)
RRCK Dataset References
Index Ref Type N Congeneric?
1 White et al. Nat. Chem. Biol. 2011, 7 810-817 Cyclic peptides 8 Yes
2 Stepan et al. J. Med. Chem. 2012, 55, 3414-3424. Small molecules 21 Yes
3 Varma et al. Mol. Pharmaceutics 2012, 9, 1199-1212. Small mol + macrocycles 104 No
4 Rand et al. Med. Chem. Commun. 2012, 3, 1282-1289. Cyclic peptides 16 Yes
5 Guzman-Perez et al. Bioorg. Med. Chem. Lett. 2013, 23, 3051-3058. Small molecules 9 Yes
6 Filipski et al. Bioorg. Med. Chem. Lett. 2013, 23, 4571-4578. Small molecules 23 Yes
7 Dow et al. Bioorg. Med. Chem. 2013, 21, 5081-5097. Small molecules 34 Yes
8 Griffith et al. J. Med. Chem. 2013, 56, 7110-7119. Small molecules 22 Yes