Hydrogen bonds in Rosetta: a phenomonological study Jack Snoeyink Dept. of Computer Science UNC Chapel Hill
Key points My biases Hydrogen bonds in Rosetta Structure-derived potential of KMB03 Existing definition/scoring Comparing natives & decoys Proposed recategorization Bad smells in code Open questions
Phenomenology defined: Movement originated by E. Husserl in 1905 A philosophy based on the premise that reality consists of objects and events as they are perceived or understood in human consciousness and not of anything independent of human consciousness.
Phenomenology defined: Movement originated by E. Husserl in 1905 A philosophy based on the premise that reality consists of objects and events as they are perceived or understood in human consciousness and not of anything independent of human consciousness.
Structure-derived potential KMB03 Energy from observed structures: distance dependence for helix
Structure-derived potential KMB03 Energy from observed structures: statistically derived energies…
Structure-derived potential KMB03 Energy from observed structures: as implemented in Rosetta…
Three tasks in Hbond scoring Identify pairs of atoms that Hbond Classify Hbond types Evaluate energies for Hbonds Rosetta++ mixes these tasks together…
Three tasks in Hbond scoring As described in KMB03 Identify pairs of atoms that Hbond Params: AH distance, , Classify Hbond types BB: helix, strand, other; AH distance SS,BS,SB: acceptor hybridization; AH dist Evaluate energies for Hbonds Sum three potentials on AH distance, , , Amino acid weights Residue neighbors for donor/acceptor
Three tasks in Hbond scoring As implemented in Rosetta++ Identify pairs of atoms that Hbond Params: AH distance, , Classify Hbond types BB: separation short |sep|≤4; long range SS,BS,SB: acceptor hybridization; AH dist Evaluate energies for Hbonds Sum three potentials on AH distance, , Amino acid weights OR Residue neighbors for donor/acceptor
SS bonds: native & decoy sp 2 ED QN bb sp 3 TS Y ring H dist
SS bonds: native & decoy sp 2 ED QN bb sp 3 TS Y ring H dist Natives: Dunbrack set of 3157 structures some pdb errors Decoys: Best 20 for each of Rhiju’s ab initio runs on 62 structures small proteins few parallel beta strands Rosetta places Hs & determines Hbonds Filter energies < -0.1 Visualization: Tufte’s small multiples Normalization Express counts as fraction of all Hbonds to support comparison of colors in each plot Plot with common x axis; scale y to max height
Energy distribution of bonds involving a sidechain atom before/after filtering
Number (and percentage) of bonds under the existing classification CountsPercentage NativeDecoysNativeDecoys BB Helix (+/-4) 185,204 38, Turn (+/-3) 79,110 8, Other 150,945 19, S sp2 ED QB bb 132,522 6, sp3 TS Y 23,641 2, ring H 2, TOTALS 573,747 75,
Observations Rosetta does well at optimizing what it is told. Decoy distributions are more sharply peaked than natives. Relax preserves more non-helix bonds than ab initio, but produces same shapes for param distrib’ns. To test changes, it suffices to run relax.
SS bonds: native & decoy sp 2 ED QN bb sp 3 TS Y ring H dist
SS bonds: native & decoy sp 2 ED QN bb sp 3 TS Y ring H dist
SS,BS,SB bonds: native & decoy sp 2 ED QN bb sp 3 TS Y ring H dist
AH Distance NB: donor effects small # omit C bimodal H R & QNacc
Theta A-H-D angle NB: small #s width R N E &N H
Psi AHD angle NB: R E & EDacc
Chi A 2 -A torsion NB: Polar & charged prefs
__ H-D torsion NB: Polar & charged prefs
Three tasks in Hbond scoring Proposed changes Identify pairs of atoms that Hbond Params: AH distance, , Classify Hbond types BB: finer separation (Beta?) SS,BS,SB: finer don/acc chemical types Evaluate energies for Hbonds: options 1. Sum three potentials on AH distance, , 2. Potential on three variables AH distance, , 3. Add neighbors 4. Add a torsion as 4 th or 5 th variable Weights for tuning different terms
Backbone bonds AH distance
Backbone bonds theta
Parallel vs Anti-parallel beta The standard figures are misleading; parallel and anti-parallel form similar distance, , distributions.
Backbone bonds psi
Backbone bonds chi torsion
Backbone bonds AH-DD 2 torsion
Refactoring Hbonds Recategorizing should eliminate long-range & short-range Hbonds; which are used outside of hbonds.cc – they shouldn’t need to be. Duplicated code in minimizers needs to be brought back into hbonds.cc
Refactoring In code, a function should do one thing well. When a function you work with is doing too many things, split it. Duplicating code indicates that something is designed wrong. Avoid magic numbers.