DMDGP Carlile Lavor
The MDGP
Introduction The DMDGP The BP algorithm MD-jeep
Solving NMR instances
First approach Making order An extension of BP:
iBP Preliminary experiments with NMR instances
Symmetries and symBP
Some theoretical results Computational experiments
The Discretizable Molecular Distance Geometry Problem
Carlile Lavor, Antonio Mucherino, Leo Liberti, Nelson Maculan
clavor@ime.unicamp.br
XVI ELAVIO, Vale dos Vinhedos, RS, Brazil
February 5 th - 10 th 2012
DMDGP Carlile Lavor
The MDGP
Introduction The DMDGP The BP algorithm MD-jeep
Solving NMR instances
First approach Making order An extension of BP:
iBP Preliminary experiments with NMR instances
Symmetries and symBP
Some theoretical results Computational experiments
Outline
1 The MDGP Introduction The DMDGP The BP algorithm MD-jeep
2 Solving NMR instances First approach Making order
An extension of BP: iBP
Preliminary experiments with NMR instances
3 Symmetries and symBP
Some theoretical results
Computational experiments
DMDGP Carlile Lavor
The MDGP
Introduction The DMDGP The BP algorithm MD-jeep
Solving NMR instances
First approach Making order An extension of BP:
iBP Preliminary experiments with NMR instances
Symmetries and symBP
Some theoretical results Computational experiments
Outline
1 The MDGP Introduction The DMDGP The BP algorithm MD-jeep
2 Solving NMR instances First approach Making order
An extension of BP: iBP
Preliminary experiments with NMR instances
3 Symmetries and symBP
Some theoretical results
Computational experiments
DMDGP Carlile Lavor
The MDGP
Introduction The DMDGP The BP algorithm MD-jeep
Solving NMR instances
First approach Making order An extension of BP:
iBP Preliminary experiments with NMR instances
Symmetries and symBP
Some theoretical results Computational experiments
Protein conformations
Proteins play many vital functions in the bodies of living beings.
They are chains of amino acids. They fold into unique
three-dimensional conformations, which define their functions.
DMDGP Carlile Lavor
The MDGP
Introduction The DMDGP The BP algorithm MD-jeep
Solving NMR instances
First approach Making order An extension of BP:
iBP Preliminary experiments with NMR instances
Symmetries and symBP
Some theoretical results Computational experiments
The Protein Data Bank (PDB)
It’s a web database of protein conformations ( ˜76000 ).
Two techniques are mainly employed:
1 X-ray diffraction
provides (in general) conformations with a good resolution;
requires the crystallization of the molecule;
2 Nuclear Magnetic Resonance (NMR)
analyzes the molecule in solution (crystallization not required);
only provides information from which the conformation can be obtained.
What does NMR provide?
Distances between some pairs of atoms of the molecule.
DMDGP Carlile Lavor
The MDGP
Introduction The DMDGP The BP algorithm MD-jeep
Solving NMR instances
First approach Making order An extension of BP:
iBP Preliminary experiments with NMR instances
Symmetries and symBP
Some theoretical results Computational experiments
The Protein Data Bank (PDB)
It’s a web database of protein conformations ( ˜76000 ).
Two techniques are mainly employed:
1 X-ray diffraction
provides (in general) conformations with a good resolution;
requires the crystallization of the molecule;
2 Nuclear Magnetic Resonance (NMR)
analyzes the molecule in solution (crystallization not required);
only provides information from which the conformation can be obtained.
What does NMR provide?
Distances between some pairs of atoms of the molecule.
DMDGP Carlile Lavor
The MDGP
Introduction The DMDGP The BP algorithm MD-jeep
Solving NMR instances
First approach Making order An extension of BP:
iBP Preliminary experiments with NMR instances
Symmetries and symBP
Some theoretical results Computational experiments
The Molecular Distance Geometry Problem
Let G = (V , E, d) be a weighted undirected graph, where V the set of vertices of G − corresponds to the set of atoms;
E the set of edges of G − corresponds to the set of known distances;
d the weights associated to the edges of G
the numerical value of each weight corresponds to the known distance.
Definition
Find a conformation x = {x 1 , x 2 , . . . , x n } such that all the following constraints are satisfied:
||x i − x j || = d ij ∀i , j : i 6= j,
where ||x i − x j || is the computed distance between x i and x j , and
d ij is the generic weight of the graph G.
DMDGP Carlile Lavor
The MDGP
Introduction The DMDGP The BP algorithm MD-jeep
Solving NMR instances
First approach Making order An extension of BP:
iBP Preliminary experiments with NMR instances
Symmetries and symBP
Some theoretical results Computational experiments
PDB and MDGP
The majority of protein conformations on the PDB related to NMR experiments are obtained by solving the corresponding MDGP by the meta-heuristic Simulated Annealing (SA).
Disadvantages of this approach:
SA is one of the less performing meta-heuristic searches;
there is no guarantee to converge to the optimal solution(s);
even though multi-start techniques are used, there is no
hope to identify all optimal solutions.
DMDGP Carlile Lavor
The MDGP
Introduction The DMDGP The BP algorithm MD-jeep
Solving NMR instances
First approach Making order An extension of BP:
iBP Preliminary experiments with NMR instances
Symmetries and symBP
Some theoretical results Computational experiments
Outline
1 The MDGP Introduction The DMDGP The BP algorithm MD-jeep
2 Solving NMR instances First approach Making order
An extension of BP: iBP
Preliminary experiments with NMR instances
3 Symmetries and symBP
Some theoretical results
Computational experiments
DMDGP Carlile Lavor
The MDGP
Introduction The DMDGP The BP algorithm MD-jeep
Solving NMR instances
First approach Making order An extension of BP:
iBP Preliminary experiments with NMR instances
Symmetries and symBP
Some theoretical results Computational experiments
The Discretizable MDGP (DMDGP)
Instances satisfying the following two assumptions can be discretized.
Given a graph G = (V, E, d ) and a total ordering on V , Ass.1 (1, 2, 3) ⊂ V must be a clique and
for each i ∈ V such that i > 3,
{(i − 3, i ), (i − 2, i), (i − 1, i)} ⊂ E ;
Ass.2 for each i ∈ V such that i > 2, the strict triangular inequality d i −2,i < d i−2,i−1 + d i−1,i
must hold.
C. Lavor, L. Liberti, N. Maculan, A. Mucherino, The Discretizable Molecular Distance Geometry Problem,
Computational Optimization and Applications, Online First Articles, March 2011.
DMDGP Carlile Lavor
The MDGP
Introduction The DMDGP The BP algorithm MD-jeep
Solving NMR instances
First approach Making order An extension of BP:
iBP Preliminary experiments with NMR instances
Symmetries and symBP
Some theoretical results Computational experiments
Discretizing with torsion angles
If the two assumptions hold, the cosine of the torsion angle among each quadruplet of consecutive atoms can be computed.
i − 3
i − 2
i − 1
i
i
0θ
i−3,i−1θ
i−2,id
i−3,id
i−3,iAn atomic position can be computed for each of the two values of
the corresponding torsion angle.
DMDGP Carlile Lavor
The MDGP
Introduction The DMDGP The BP algorithm MD-jeep
Solving NMR instances
First approach Making order An extension of BP:
iBP Preliminary experiments with NMR instances
Symmetries and symBP
Some theoretical results Computational experiments
Discretizing with three spheres
If the assumptions hold, the two possible positions for the atom x i can be computed by intersecting three spheres:
S −1 : center: x i−1 , radius: d i−1,i ; S −2 : center: x i−2 , radius: d i−2,i ; S −3 : center: x i−3 , radius: d i−3,i .
The intersection of the spheres S − 1 , S − 2 and S − 3 can be
one point, two points, a circle, empty.
The intersection can be computed by solving two linear systems.
DMDGP Carlile Lavor
The MDGP
Introduction The DMDGP The BP algorithm MD-jeep
Solving NMR instances
First approach Making order An extension of BP:
iBP Preliminary experiments with NMR instances
Symmetries and symBP
Some theoretical results Computational experiments
Discretizing with three spheres
If the assumptions hold, the two possible positions for the atom x i can be computed by intersecting three spheres:
S −1 : center: x i−1 , radius: d i−1,i ; S −2 : center: x i−2 , radius: d i−2,i ; S −3 : center: x i−3 , radius: d i−3,i .
The intersection of the spheres S − 1 , S − 2 and S − 3 can be
one point (probability 0), two points,
a circle (impossible, because of the strict triangular inequalities), empty (impossible, because of protein steric constraints).
The intersection can be computed by solving two linear systems.
DMDGP Carlile Lavor
The MDGP
Introduction The DMDGP The BP algorithm MD-jeep
Solving NMR instances
First approach Making order An extension of BP:
iBP Preliminary experiments with NMR instances
Symmetries and symBP
Some theoretical results Computational experiments
The Discretizable MDGP (DMDGP)
When these two assumptions are satisfied, the domain of the penalty function can be reduced to a discrete set.
In particular, a binary tree of atomic positions can be defined, and
solutions to the problem can be found by exploring this tree.
DMDGP Carlile Lavor
The MDGP
Introduction The DMDGP The BP algorithm MD-jeep
Solving NMR instances
First approach Making order An extension of BP:
iBP Preliminary experiments with NMR instances
Symmetries and symBP
Some theoretical results Computational experiments
Complexity
Definition
S UBSET -S UM . Given nonnegative integers a 1 , . . . , a n , is there a partition into two sets, encoded by s ∈ {−1, +1} n , such that each subset has the same sum, i.e. P n
i=1 s(i )a i = 0?
By reduction from the Subset-sum problem (which is known to be NP-hard), we can prove the following:
Theorem
The DMDGP is NP-hard.
C. Lavor, L. Liberti, N. Maculan, A. Mucherino, The Discretizable Molecular Distance Geometry Problem,
Computational Optimization and Applications, Online First Articles, March 2011.
DMDGP Carlile Lavor
The MDGP
Introduction The DMDGP The BP algorithm MD-jeep
Solving NMR instances
First approach Making order An extension of BP:
iBP Preliminary experiments with NMR instances
Symmetries and symBP
Some theoretical results Computational experiments
Outline
1 The MDGP Introduction The DMDGP The BP algorithm MD-jeep
2 Solving NMR instances First approach Making order
An extension of BP: iBP
Preliminary experiments with NMR instances
3 Symmetries and symBP
Some theoretical results
Computational experiments
DMDGP Carlile Lavor
The MDGP
Introduction The DMDGP The BP algorithm MD-jeep
Solving NMR instances
First approach Making order An extension of BP:
iBP Preliminary experiments with NMR instances
Symmetries and symBP
Some theoretical results Computational experiments
The Branch & Prune (BP) algorithm
Branches of the binary tree are pruned as soon as at least one of
the known distances is violated. In this way, an exhaustive search
on the remaining branches is not so expensive.
DMDGP Carlile Lavor
The MDGP
Introduction The DMDGP The BP algorithm MD-jeep
Solving NMR instances
First approach Making order An extension of BP:
iBP Preliminary experiments with NMR instances
Symmetries and symBP
Some theoretical results Computational experiments
The Branch & Prune (BP) algorithm
Branches of the binary tree are pruned as soon as at least one of
the known distances is violated. In this way, an exhaustive search
on the remaining branches is not so expensive.
DMDGP Carlile Lavor
The MDGP
Introduction The DMDGP The BP algorithm MD-jeep
Solving NMR instances
First approach Making order An extension of BP:
iBP Preliminary experiments with NMR instances
Symmetries and symBP
Some theoretical results Computational experiments
The Branch & Prune (BP) algorithm
Branches of the binary tree are pruned as soon as at least one of
the known distances is violated. In this way, an exhaustive search
on the remaining branches is not so expensive.
DMDGP Carlile Lavor
The MDGP
Introduction The DMDGP The BP algorithm MD-jeep
Solving NMR instances
First approach Making order An extension of BP:
iBP Preliminary experiments with NMR instances
Symmetries and symBP
Some theoretical results Computational experiments
The Branch & Prune (BP) algorithm
Branches of the binary tree are pruned as soon as at least one of
the known distances is violated. In this way, an exhaustive search
on the remaining branches is not so expensive.
DMDGP Carlile Lavor
The MDGP
Introduction The DMDGP The BP algorithm MD-jeep
Solving NMR instances
First approach Making order An extension of BP:
iBP Preliminary experiments with NMR instances
Symmetries and symBP
Some theoretical results Computational experiments
The Branch & Prune (BP) algorithm
Branches of the binary tree are pruned as soon as at least one of
the known distances is violated. In this way, an exhaustive search
on the remaining branches is not so expensive.
DMDGP Carlile Lavor
The MDGP
Introduction The DMDGP The BP algorithm MD-jeep
Solving NMR instances
First approach Making order An extension of BP:
iBP Preliminary experiments with NMR instances
Symmetries and symBP
Some theoretical results Computational experiments
The Branch & Prune (BP) algorithm
Branches of the binary tree are pruned as soon as at least one of
the known distances is violated. In this way, an exhaustive search
on the remaining branches is not so expensive.
DMDGP Carlile Lavor
The MDGP
Introduction The DMDGP The BP algorithm MD-jeep
Solving NMR instances
First approach Making order An extension of BP:
iBP Preliminary experiments with NMR instances
Symmetries and symBP
Some theoretical results Computational experiments
The Branch & Prune (BP) algorithm
Branches of the binary tree are pruned as soon as at least one of
the known distances is violated. In this way, an exhaustive search
on the remaining branches is not so expensive.
DMDGP Carlile Lavor
The MDGP
Introduction The DMDGP The BP algorithm MD-jeep
Solving NMR instances
First approach Making order An extension of BP:
iBP Preliminary experiments with NMR instances
Symmetries and symBP
Some theoretical results Computational experiments
Outline
1 The MDGP Introduction The DMDGP The BP algorithm MD-jeep
2 Solving NMR instances First approach Making order
An extension of BP: iBP
Preliminary experiments with NMR instances
3 Symmetries and symBP
Some theoretical results
Computational experiments
DMDGP Carlile Lavor
The MDGP
Introduction The DMDGP The BP algorithm MD-jeep
Solving NMR instances
First approach Making order An extension of BP:
iBP Preliminary experiments with NMR instances
Symmetries and symBP
Some theoretical results Computational experiments
MD-jeep
MD-jeep is an implemention of the BP algorithm in C programming language.
Legend of the following Tables:
Name - instance name;
n - instance dimension;
|E| - number of known distances;
#Sol - number of found solutions;
CPU - CPU time;
LDE - Largest Distance Error:
LDE = 1
|E|
X
(i,j)∈E
| ||x i − x j || − d ij |
d ij
DMDGP Carlile Lavor
The MDGP
Introduction The DMDGP The BP algorithm MD-jeep
Solving NMR instances
First approach Making order An extension of BP:
iBP Preliminary experiments with NMR instances
Symmetries and symBP
Some theoretical results Computational experiments
Generation of the instances
We downloaded a subset of proteins from the PDB, and we generated a subset of instances for the DMDGP.
Only backbone atoms N–C α –C are considered, and only distances shorter than 6Å are included in the instance.
Most papers written by the operation research community consider this kind of instances.
DMDGP Carlile Lavor
The MDGP
Introduction The DMDGP The BP algorithm MD-jeep
Solving NMR instances
First approach Making order An extension of BP:
iBP Preliminary experiments with NMR instances
Symmetries and symBP
Some theoretical results Computational experiments
Generation of the instances
We downloaded a subset of proteins from the PDB, and we generated a subset of instances for the DMDGP.
Only backbone atoms N–C α –C are considered, and only distances shorter than 6Å are included in the instance.
Most papers written by the operation research community consider this kind of instances.
DMDGP Carlile Lavor
The MDGP
Introduction The DMDGP The BP algorithm MD-jeep
Solving NMR instances
First approach Making order An extension of BP:
iBP Preliminary experiments with NMR instances
Symmetries and symBP
Some theoretical results Computational experiments
Other software for distance geometry
DGSOL
by Jorge Moré, Argonne National Laboratory
It is based on the idea of approximating the objective function with a sequence of smoother functions converging to the original objective function.
Class of problems: DGP with interval data.
SDP-based facial reduction method by Nathan Krislock, INRIA, Nice
It is based on the idea of reducing the search space to subsets of faces, where a subproblem is defined and solved independently from the others.
Class of problems: DGP with exact data and anchors.
DMDGP Carlile Lavor
The MDGP
Introduction The DMDGP The BP algorithm MD-jeep
Solving NMR instances
First approach Making order An extension of BP:
iBP Preliminary experiments with NMR instances
Symmetries and symBP
Some theoretical results Computational experiments
Some computational experiments
Instance BP-One BP-All SDP-based DGSOL
Name n |E| CPU LDE CPU #Sol CPU LDE CPU LDE
1brv 57 476 0.00 1.54e-14 0.00 1 0.03 1.24e-14 1.48 2.74e-01
1aqr 120 929 0.00 1.86e-09 0.00 2 0.06 2.54e-13 7.77 4.88e-01
2erl 120 1136 0.00 1.33e-14 0.00 1 0.06 2.52e-13 9.38 2.92e-01
1crn 138 1250 0.00 2.24e-13 0.00 1 0.06 2.24e-14 9.47 2.24e-01
1ahl 147 1205 0.00 1.50e-09 0.00 8 0.07 2.41e-14 6.95 1.46e-01
1ptq 150 1263 0.00 2.30e-13 0.00 1 0.08 2.54e-14 9.16 1.21e-01
1brz 159 1394 0.00 3.53e-13 0.00 2 0.07 2.01e-13 11.39 4.66e-01
1hoe 222 1995 0.00 3.18e-13 0.00 1 0.12 1.31e-13 16.83 2.06e-01
1lfb 232 2137 0.00 5.31e-14 0.00 1 0.11 1.86e-14 38.94 2.88e-01
1pht 249 2283 0.00 2.73e-12 0.00 1 0.10 9.54e-14 42.50 2.00e-01
1jk2 270 2574 0.00 2.09e-13 0.00 1 0.15 2.74e-14 86.98 4.05e-01
1f39a 303 2660 0.00 2.68e-12 0.00 1 0.12 3.91e-13 37.24 2.80e-01
1acz 324 3060 0.00 3.15e-12 0.02 4 0.13 3.04e-13 35.97 3.97e-01
1poa 354 3193 0.00 1.36e-13 0.00 1 0.20 2.53e-12 64.03 4.67e-01
1fs3 378 3443 0.00 8.08e-13 0.01 1 0.17 2.27e-13 54.68 2.69e-01
1mbn 459 4599 0.00 1.36e-09 0.00 1 0.22 9.67e-14 124.24 4.46e-01
1rgs 792 7626 0.00 4.22e-13 0.01 1 0.42 1.58e-13 237.93 4.69e-01
1m40 1224 20382 0.02 1.00e-12 5.26 1 0.71 1.08e-12 1142.49 4.89e-01
1bpm 1443 14292 0.02 2.85e-13 0.02 1 0.76 7.73e-13 398.29 5.06e-01
1n4w 1610 16940 0.02 1.19e-12 0.02 1 0.86 5.44e-13 994.51 5.26e-01
1mqq 2032 19564 0.02 4.90e-12 0.06 1 1.22 6.17e-13 451.58 5.40e-01
1rwh 2265 21666 0.02 2.08e-13 0.06 1 1.38 3.01e-12 934.29 5.38e-01
3b34 2790 29188 0.07 1.17e-11 0.07 1 1.68 3.00e-13 940.95 6.47e-01
2e7z 2907 42098 0.08 4.26e-12 0.09 1 1.88 2.88e-13 915.39 6.40e-01
1epw 3861 35028 0.16 3.19e-12 0.25 1 2.31 1.45e-12 2037.86 4.92e-01
DMDGP Carlile Lavor
The MDGP
Introduction The DMDGP The BP algorithm MD-jeep
Solving NMR instances
First approach Making order An extension of BP:
iBP Preliminary experiments with NMR instances
Symmetries and symBP
Some theoretical results Computational experiments
Visualizing the results
Solutions can be printed in text files in PDB format.
Visualization software (such RasMol) are able to read files in this format.
Two views of one of the found solutions for the instance 1mbn .
DMDGP Carlile Lavor
The MDGP
Introduction The DMDGP The BP algorithm MD-jeep
Solving NMR instances
First approach Making order An extension of BP:
iBP Preliminary experiments with NMR instances
Symmetries and symBP
Some theoretical results Computational experiments
Comparison to DGSOL
Comparison between two solutions, one found by BP, one found by DGSOL.
BP: LDE = 1.36e-09, RMSD = 2.87e-07 DGSOL: LDE = 4.46e-01, RMSD = 1.18e+01
RMSD: Root Mean Square Deviation.
One of the solutions found by the BP algorithm is actually the conformation
the protein 1mbn has.
DMDGP Carlile Lavor
The MDGP
Introduction The DMDGP The BP algorithm MD-jeep
Solving NMR instances
First approach Making order An extension of BP:
iBP Preliminary experiments with NMR instances
Symmetries and symBP
Some theoretical results Computational experiments
Outline
1 The MDGP Introduction The DMDGP The BP algorithm MD-jeep
2 Solving NMR instances First approach Making order
An extension of BP: iBP
Preliminary experiments with NMR instances
3 Symmetries and symBP
Some theoretical results
Computational experiments
DMDGP Carlile Lavor
The MDGP
Introduction The DMDGP The BP algorithm MD-jeep
Solving NMR instances
First approach Making order An extension of BP:
iBP Preliminary experiments with NMR instances
Symmetries and symBP
Some theoretical results Computational experiments
An artificial backbone of hydrogens
Given the set of hydrogens related to the protein backbone, how can we sort them so that the necessary assumptions are satisfied?
The order is specified through the labels on the edges.
DMDGP Carlile Lavor
The MDGP
Introduction The DMDGP The BP algorithm MD-jeep
Solving NMR instances
First approach Making order An extension of BP:
iBP Preliminary experiments with NMR instances
Symmetries and symBP
Some theoretical results Computational experiments
Reconstructing the real backbone
||a − b 1 || = d a,b
1||a − b 2 || = d a,b
2||a − b 3 || = d a,b
3||a − b 4 || = d a,b
4DMDGP Carlile Lavor
The MDGP
Introduction The DMDGP The BP algorithm MD-jeep
Solving NMR instances
First approach Making order An extension of BP:
iBP Preliminary experiments with NMR instances
Symmetries and symBP
Some theoretical results Computational experiments
Reconstructing the real backbone
||a − b 1 || = d a,b
1||a − b 2 || = d a,b
2||a − b 3 || = d a,b
3||a − b 4 || = d a,b
4DMDGP Carlile Lavor
The MDGP
Introduction The DMDGP The BP algorithm MD-jeep
Solving NMR instances
First approach Making order An extension of BP:
iBP Preliminary experiments with NMR instances
Symmetries and symBP
Some theoretical results Computational experiments
Reconstructing the real backbone
||a − b 1 || = d a,b
1||a − b 2 || = d a,b
2||a − b 3 || = d a,b
3||a − b 4 || = d a,b
4DMDGP Carlile Lavor
The MDGP
Introduction The DMDGP The BP algorithm MD-jeep
Solving NMR instances
First approach Making order An extension of BP:
iBP Preliminary experiments with NMR instances
Symmetries and symBP
Some theoretical results Computational experiments
Reconstructing the real backbone
||a − b 1 || = d a,b
1||a − b 2 || = d a,b
2||a − b 3 || = d a,b
3||a − b 4 || = d a,b
4DMDGP Carlile Lavor
The MDGP
Introduction The DMDGP The BP algorithm MD-jeep
Solving NMR instances
First approach Making order An extension of BP:
iBP Preliminary experiments with NMR instances
Symmetries and symBP
Some theoretical results Computational experiments
Outline
1 The MDGP Introduction The DMDGP The BP algorithm MD-jeep
2 Solving NMR instances First approach Making order
An extension of BP: iBP
Preliminary experiments with NMR instances
3 Symmetries and symBP
Some theoretical results
Computational experiments
DMDGP Carlile Lavor
The MDGP
Introduction The DMDGP The BP algorithm MD-jeep
Solving NMR instances
First approach Making order An extension of BP:
iBP Preliminary experiments with NMR instances
Symmetries and symBP
Some theoretical results Computational experiments
A particular ordering for protein backbones
The ordering is defined by the arrows in red.
If this special ordering is used, we are able to discretize real instances containing data from NMR.
C. Lavor, L. Liberti, A. Mucherino, The interval Branch-and-Prune Algorithm for the Discretizable Molecular
Distance Geometry Problem with Inexact Distances, to appear in Journal of Global Optimization, 2011.
DMDGP Carlile Lavor
The MDGP
Introduction The DMDGP The BP algorithm MD-jeep
Solving NMR instances
First approach Making order An extension of BP:
iBP Preliminary experiments with NMR instances
Symmetries and symBP
Some theoretical results Computational experiments
A particular ordering for protein backbones
How to discretize real instances?
All distances required for the discretization (Ass.1) are obtained from the chemical structure of the protein backbones (so they are independent from the instance);
Distances between pairs (i, i + 1) and (i , i + 2) are always exact;
Distances between pairs (i, i + 3) may be represented by intervals:
1 d i,i+3 is 0: it means that this is a duplicated atom, there is no
branching because it can only take the same position of its previous copy;
2 d i,i+3 is exact: the standard discretization process is applied, and
hence two possible positions for the current atom are computed;
3 d i,i+3 is represented by an interval: D sample distances are taken
from the interval and the discretization process is applied for any
chosen sample distance; 2 × D atomic positions are generated.
DMDGP Carlile Lavor
The MDGP
Introduction The DMDGP The BP algorithm MD-jeep
Solving NMR instances
First approach Making order An extension of BP:
iBP Preliminary experiments with NMR instances
Symmetries and symBP
Some theoretical results Computational experiments
One branch of the new tree
DMDGP Carlile Lavor
The MDGP
Introduction The DMDGP The BP algorithm MD-jeep
Solving NMR instances
First approach Making order An extension of BP:
iBP Preliminary experiments with NMR instances
Symmetries and symBP
Some theoretical results Computational experiments
Outline
1 The MDGP Introduction The DMDGP The BP algorithm MD-jeep
2 Solving NMR instances First approach Making order
An extension of BP: iBP
Preliminary experiments with NMR instances
3 Symmetries and symBP
Some theoretical results
Computational experiments
DMDGP Carlile Lavor
The MDGP
Introduction The DMDGP The BP algorithm MD-jeep
Solving NMR instances
First approach Making order An extension of BP:
iBP Preliminary experiments with NMR instances
Symmetries and symBP
Some theoretical results Computational experiments
interval Branch & Prune
0: IBP(i, n, d , nbranches) if (x
iis a duplicated atom) then
assign to x
ithe same coordinates of its previous copy;
IBP(i + 1,n,d ,nbranches);
else
if (d(i − 3, i) is exact) then b = 2;
else
b = nbranches;
end if for (k = 1, b) do
compute the k
thatomic position for the i
thatom: x
ik; check the feasibility of the atomic position x
ik: if (x
ikis feasible) then
if (i = n) then a solution is found;
else
IBP(i + 1,n,d ,nbranches);
end if else
the current branch is pruned;
end if
end for
end if
DMDGP Carlile Lavor
The MDGP
Introduction The DMDGP The BP algorithm MD-jeep
Solving NMR instances
First approach Making order An extension of BP:
iBP Preliminary experiments with NMR instances
Symmetries and symBP
Some theoretical results Computational experiments
Outline
1 The MDGP Introduction The DMDGP The BP algorithm MD-jeep
2 Solving NMR instances First approach Making order
An extension of BP: iBP
Preliminary experiments with NMR instances
3 Symmetries and symBP
Some theoretical results
Computational experiments
DMDGP Carlile Lavor
The MDGP
Introduction The DMDGP The BP algorithm MD-jeep
Solving NMR instances
First approach Making order An extension of BP:
iBP Preliminary experiments with NMR instances
Symmetries and symBP
Some theoretical results Computational experiments
Instance with 3 amino acids
We firstly consider this simple molecule.
Bond lengths and bond angles are constant, torsion angles can
vary into a predefined interval.
DMDGP Carlile Lavor
The MDGP
Introduction The DMDGP The BP algorithm MD-jeep
Solving NMR instances
First approach Making order An extension of BP:
iBP Preliminary experiments with NMR instances
Symmetries and symBP
Some theoretical results Computational experiments
Preliminary experiments
layer atom duplicated? branches w/out pruning branches with pruning atom
1 N 1 no 1 1 N 1
2 H 1 no 1 1 H 1
3 H 1 no 1 1 H 1
4 CA 1 no 2 2 CA 1
5 N 1 yes 2 2 N 1
6 HA 1 no 24 18 HA 1
7 CA 1 yes 24 18 CA 1
8 C 1 no 48 36 C 1
9 N 2 no 576 360 N 2
10 CA 2 no 1152 720 CA 2
11 H 2 no 2304 10 H 2
12 N 2 yes 2304 10 N 2
13 CA 2 yes 2304 10 CA 2
14 HA 2 no 27648 70 HA 2
15 C 2 no 55296 140 C 2
16 CA 2 yes 55296 140 CA 2
17 N 3 no 663552 1400 N 3
18 C 2 yes 663552 1400 C 2
19 CA 3 no 1327104 2800 CA 3
20 H 3 no 2654208 4 H 3
21 N 3 yes 2654208 4 N 3
22 CA 3 yes 2654208 4 CA 3
23 HA 3 no 31850496 9 HA 3
24 C 3 no 63700992 18 C 3
25 CA 3 yes 63700992 18 CA 3
26 O 3 no 764411904 52 O 3
27 C 3 yes 764411904 52 C 3
28 O 3 no 9172942848 10 O 3
DMDGP Carlile Lavor
The MDGP
Introduction The DMDGP The BP algorithm MD-jeep
Solving NMR instances
First approach Making order An extension of BP:
iBP Preliminary experiments with NMR instances
Symmetries and symBP
Some theoretical results Computational experiments
Outline
1 The MDGP Introduction The DMDGP The BP algorithm MD-jeep
2 Solving NMR instances First approach Making order
An extension of BP: iBP
Preliminary experiments with NMR instances
3 Symmetries and symBP
Some theoretical results
Computational experiments
DMDGP Carlile Lavor
The MDGP
Introduction The DMDGP The BP algorithm MD-jeep
Solving NMR instances
First approach Making order An extension of BP:
iBP Preliminary experiments with NMR instances
Symmetries and symBP
Some theoretical results Computational experiments
The first symmetry
It is possible to prove that:
If the DMDGP problem has a solution, then it has at least another symmetric solution.
Pairs of solutions of the DMDGP are symmetric.
Theorem
Let x : G → R 3 be a solution for the DMDGP, defined by the
torsion angles ω 1,4 , . . . , ω n − 3,n . If we invert the sign of sin ω i − 3,i ,
for i = 4, ..., n, then we obtain a new solution x 0 : G → R 3 for the
DMDGP.
DMDGP Carlile Lavor
The MDGP
Introduction The DMDGP The BP algorithm MD-jeep
Solving NMR instances
First approach Making order An extension of BP:
iBP Preliminary experiments with NMR instances
Symmetries and symBP
Some theoretical results Computational experiments
The other symmetries
We recently proved that, for each vertex i belonging to this set B = {v ∈ V :6 ∃(u, w) s.t. u + 3 < v ≤ w}
there is a symmetry on the binary tree.
In other words, for each i ∈ B, there are two symmetric branches on the tree having as root a common position for x i − 1 .
Theorem
With probability 1, the number of solutions of the DMDGP is a
power of 2.
DMDGP Carlile Lavor
The MDGP
Introduction The DMDGP The BP algorithm MD-jeep
Solving NMR instances
First approach Making order An extension of BP:
iBP Preliminary experiments with NMR instances
Symmetries and symBP
Some theoretical results Computational experiments
An instance with 2 symmetries (and 4 solutions)
The set B related to this small instance is {4, 6}. Feasible branches are marked in light yellow.
If one solution is already available, BP can exploit B in order to
guide the search towards the feasible branches of the tree. This is
the idea behind symBP.
DMDGP Carlile Lavor
The MDGP
Introduction The DMDGP The BP algorithm MD-jeep
Solving NMR instances
First approach Making order An extension of BP:
iBP Preliminary experiments with NMR instances
Symmetries and symBP
Some theoretical results Computational experiments
Outline
1 The MDGP Introduction The DMDGP The BP algorithm MD-jeep
2 Solving NMR instances First approach Making order
An extension of BP: iBP
Preliminary experiments with NMR instances
3 Symmetries and symBP
Some theoretical results
Computational experiments
DMDGP Carlile Lavor
The MDGP
Introduction The DMDGP The BP algorithm MD-jeep
Solving NMR instances
First approach Making order An extension of BP:
iBP Preliminary experiments with NMR instances
Symmetries and symBP
Some theoretical results Computational experiments
Comparisons between BP and symBP
instance BP symBP
name n |E| #Sol #DDF time #Sol #DDF time
lavor_10 10 29 4 44 0.00 4 2 0.00
lavor_20 20 77 64 86 0.00 64 7 0.00
lavor_30 30 161 64 592 0.01 64 18 0.00
lavor_40 40 194 2048 30720 0.05 2048 19 0.01 lavor_50 50 203 1024 46728 0.07 1024 49 0.01
lavor_60 60 357 256 71352 0.12 256 89 0.01
lavor_70 70 591 16 232 0.00 16 67 0.00
lavor_100 100 605 2 4460924 7.55 2 815010 1.37
lavor_200 200 1844 32 35298 0.09 32 394 0.01
lavor_300 300 2505 4 38364 0.07 4 9378 0.03
Experiments with Lavor instances.
DMDGP Carlile Lavor
The MDGP
Introduction The DMDGP The BP algorithm MD-jeep
Solving NMR instances
First approach Making order An extension of BP:
iBP Preliminary experiments with NMR instances
Symmetries and symBP
Some theoretical results Computational experiments
Comparisons between BP and symBP
instance BP symBP
name n |E| |B| #Sol time #Sol time
lavor_400 400 2600 35 10
58 10
52
lavor_500 500 4577 44 10
5310 10
54 lavor_600 600 5933 25 10
530 10
57 lavor_700 700 5679 71 10
57050 10
5134
All these experiments are stopped when 100,000 solutions are found. The
cardinality of B is so large that neither BP nor symBP are able to enumerate all
solutions.
DMDGP Carlile Lavor
The MDGP
Introduction The DMDGP The BP algorithm MD-jeep
Solving NMR instances
First approach Making order An extension of BP:
iBP Preliminary experiments with NMR instances
Symmetries and symBP
Some theoretical results Computational experiments
Comparisons between BP and symBP
instance BP symBP
name n |E| #Sol #DDF time #Sol #DDF time
1m40 1224 20382 2 2488 0.03 2 115 0.02
1bpm 1443 14292 2 2890 0.04 2 519 0.02
1mqq 2032 19564 2 4066 0.08 2 737 0.04
3b34 2790 29188 2 5664 0.15 2 1798 0.10
1epw 3861 35028 2 11974 0.45 2 3269 0.21
Experiments with PDB instances.
DMDGP Carlile Lavor
The MDGP
Introduction The DMDGP The BP algorithm MD-jeep
Solving NMR instances
First approach Making order An extension of BP:
iBP Preliminary experiments with NMR instances
Symmetries and symBP
Some theoretical results Computational experiments