Kabsch

Intro

The Kabsch algorithm uses linear and vector algebra to find the optimal rotation and translation of two sets of points in N-dimensional space as to minimize the RMSD between them. The following program is a Python implementation of the Kabsch algorithm.

This program when called will align the two selections, optimally, convert the proteins in the selection to ribbons and change the color of the selections to show the matched alignments.

WHAT THIS DOESN'T DO: This program does NOT provide a pairwise alignment of two structures from scratch. You have to tell it what the equivlanced items are. This is NOT DALI. (Yet.)

NOTE: This will undergo more edits as I continue this project.

NOTE: This has NOT been tested on any other machine than mine, by me. It works on all PyMols 0.97 and newer (haven't tested earlier versions) and I use Python 2.3's Numeric Version 23.3.

To use

1. Save this script to Kabsch.py
2. Open PyMol
3. Load the alignment script: run Kabsch.py (The command optAlign is now defined in PyMol.)
4. Load your proteins
5. Align the proper segments (see below examples)

To align two equivalent sets of residues do:

optAlign SEL1 and n. CA and i. a-b, SEL2 and n. CA and c-d

where

• SEL1 is the first protein
• a-b is the range of residues to align in the first protein
• SEL2 is the second protein
• c-d is the range of residues to align in the second protein

Caveats

• Ensure that you're equivalencing N atoms to N atoms (run Count_Atoms over your two selections to ensure they are the same length).
• Sometimes PyMol doesn't seem to superimpose them right the first time. Hit the up-arrow and rerun the program if this happens. It always superimposes the correctly the second time. I think it has something to do with the orientation. I'll fix this when I find the error.
• The RMSD is only between the equivlanced atoms. Use PyMol's Rms_Cur if you want a full RMSD.
• Make sure your atom selections are numbered correctly. Many times PDB files start residue numbers at something other than 0 or 1. To ensure you're aligning the right things, do

  set seq_view,1
  

  to turn on the sequence viewer and double check your residue numbers. If a protein has residue one numbered as something other than one, say 2064, simply run

  alter (SEL), resi=str(int(resi)-2064)
  

  and then

  sort
  

  where SEL is the name of the protein and 2064 is the offset to adjust by. Your protein will now work as needed. See Alter.
• Windows users are having problems running the script. Python tells them first off "TypeError: Can't convert rank-0 arrays to Python scalars." The fix to that breaks some code in Numeric -- which I don't maintain.

Examples

optAlign 1cll and n. CA and i. 4-20+30-60, 1ggz and n. CA and i. 4-20+30-60

optAlign 1kao and n. CA and i. 20-50, 1ctq and n. CA and i. 20-50

1cll and 1ggz loaded

1cll and 1ggz aligned to residues 5-50+55-80 shown in red

You already need a sequence alignment to effectively run this program. (This is VERY strongly recommended!!)

The Code

#!python
 
##############################################################################
#
# @SUMMARY: -- Kabsch.py.  A python implementation of the optimal superposition
#     of two sets of vectors as proposed by Kabsch 1976 & 1978.
#
# @AUTHOR: Jason Vertrees
# @COPYRIGHT: Jason Vertrees (c), 2005.
# DATE  : 2005-04-07
# REV   : 1
#
#############################################################################
#
#  Modifications for Win32 compatibility
# 
#  Shivender Shandilya 
#  09-28-2005
#
#  NOTE: In the "swapaxes" funtion in Numeric.py (line 250):
#            if axis2 < 0: axis2 = axis1 + n  should be replaced with
#            if axis2 < 0: axis2 = axis2 + n  for "dot" to work correctly.
# 
#  This seems to be an error within the Numeric.py source code.
#
#  NOTE: Always run the same "optAlign" command TWICE in a row to get 
#        the proper Kabsch fit!
#        [Why this needs to be done is still under investigation.]
#
#############################################################################
# math
import math
import Numeric
import LinearAlgebra
import Matrix
 
from array import *
 
import random
random.seed()
 
# system stuff
import os
import copy
 
# pretty printing
import pprint
import string
 
# for importing as a plugin into PyMol
#import tkSimpleDialog
#import tkMessageBox
from pymol import cmd
from pymol import stored
from pymol import selector
 
class kabsch:
	"""
	Kabsch alignment of two set of vectors to produce and optimal alignemnt.
 
	Steps
	=====	
		1.  Calculate the center of mass for each protein.  Then, move the protein
		to its center of mass.  We choose as a convention, to use the origin as 
		the center of mass of the first set of coordinates.  This will allow us to
		return one translation vector, instead of two.
		
		Update: Since any rotation around a point that's not the origin, is in fact
		an affine rotation.  So, to beat that, we translate both to the center.
		
		NAME: superpose(c1, c2)
		POST: T is now defined.
		
		2.  Calculate the matrix, R, by (eq7, 1976).  r_{i,j} = sum_n w_n * y_{ni} * x_{nj},
		where y_{ni} is the ith component of the vector y_n.
		NAME: calcR
		POST: R is now defined
		
		3.  Calculate RtR (R-transpose * R).
		NAME: calcRtR
		POST: RtR is now defined
		
		4.  Calculate the corresponding eigenpairs for RtR, and sort them accordingly:
		m1 >= m2 >= m3; set v3 = v1 x v2 to ensure a RHS
		NAME: calcEigenPairs
		POST: The eigen pairs are calculated, sorted such that m1 >= m2 >= m3 and
		v3 = v1 x v2.
		
		5.  Calculate R*v_k and normalize the first two vectors to obtain b_1, b_2 and set
		b_3 = b_1 x b_2.  This also takes care of m1 > m2 = 0. 
		NAME: calcBVectors
		POST: b-Vectors are defined and returned
		
		6.  Calculate U=(u_{ij})=(sum n b_{ki} * a_{kj}) to obtain the best rotation.  Set
		sigma_3 = -1 if b3(Ra3) < 0 else sigma_3 = +1.
		NAME: calcU
		POST: U is defined
		
		7.  Calculate the RMSD.  The residual error is then
		The E = E0 - sqrt(m1) - sqrt(m2) - sigma_3(sqrt(m3)).
		NAME: calcRMSD
		POST: RMSD is computed.
	
	 @note: This should be a static method that takes three parameters
		
		1. The first protein's coordinates, f.  This program will 
		accept coordinates in the form an array of 3D vectors/lists
		
		2. The second protein's coordinates, g.
		
		3. The array of integer pairs representing the pairs to align.
		Coordinates should be formatted as as array of 2D vectors/lists.
	
 
	
	"""
	
	def __init__(self):
		"""
		Constructor.  Doesn't do shit.  @see kabsch.align.
		
		"""
 
		#
		# Instance Variables:  All three of these will be updated
		# every time the user calls ~.align.  Just to warn ya'.
		#		
		# U, the rotation matrix
		self.U = []
		# T, the translation vector
		self.T = []
		# R, the RMSD
		self.R = -1.0
 
		#self.menuBar.addmenuitem('Plugin', 'command', 'Kabsch Align', label = "Align Selections to Optial RMSD", command = lamda s=self: fetchPDBDialog(s))
		
				
	def align(self, c1, c2, pairs):
		"""
		Finds the best alignment of c1 and c2's pairs resulting in
		the smallest possible RMSD.
 
				
		@note:
			- All weights in this first version are set to 1.  Kabsch allows,
			differential weighting.  In the future, I may extend to this option,
			and then pairs may become 3D (r1, r2, weight) or I may add another
			parameter.
		
			- Helper functions will soon be provided such that the user may
			just use this package alone to compute the rotation.
		
		@param c1: coordinats of the first vectors, as an array of 3D vectors.
		@type  c1: Python list
		   
		@param c2: coordinates of the second set of vectors, as an array of 3D vectors.
		@type  c2: Python list
		   
		@param pairs: the list of pairs as an array of 2D pairs.
		@type  pairs: Python list of 2D lists.
		
		@return: U, the rotation matrix that gives the optimal rotation between the two proteins.
			
			T, the translation vector given to align the two objects centers of mass.
			
			R, the RMSD of the final rotation.
		"""
		
		#
		# First we move the center of mass of one protein, to the
		# center of mass of the other.  This removes any translation
		# between the two.
		#
		T1, T2, c1, c2 = self.superpose(c1, c2)
		# Calculate the initial RMSD
		E0 = self.calcE0(c1, c2)
		# Calculate R via eq. 7.
		R = self.calcR(c1, c2)
		# Calculate R(transpose)*R
		RtR = self.calcRtR(R)
		# Determined the eigenpairs for the matrix RtR.
		eValues, eVectors = self.calcEigenPairs(RtR)
		# Determine the bVectors as required
		bVectors = self.calcBVectors(R, eVectors)
		# Calculate the roation matrix
		U = self.calcU(eVectors, bVectors)
		# Calculate the final RMSD using U.
		RMSD = self.calcRMSD(E0, eValues, eVectors, bVectors, R)
		
		return U, T1, T2, RMSD, c1, c2
		
		
	def superpose(self, c1, c2 ):
		"""
		Calculate the center of mass for each protein.  Then, move the protein
		to its center of mass.  We choose as a convention, to use the origin as 
		the center of mass of the first set of coordinates.  This will allow us to
		return one translation vector, instead of two.
		(CORRECT)
		
		@precondition: c1 and c2 are well defined lists of N-dimensional points with length > 0.
		@postcondition: T is now defined.
		
		@param c1: coordinats of the first vectors, as an array of 3D vectors.
		@type  c1: Python list
		   
		@param c2: coordinates of the second set of vectors, as an array of 3D vectors.
		@type  c2: Python list
		
		@return: T the translation vector.
		
			c2 one list of coordinates that of c2 translated to the COM of c1.
		
		"""
		
		# make sure we don't get bad data
		assert(len(c1) == len(c2) != 0)
 
		L = len(c1)
		
		#
		# Centers of Mass
		#
		c1COM = Numeric.zeros((3,1), Numeric.Float64)
		c2COM = Numeric.zeros((3,1), Numeric.Float64)
 
		# calculate the CsOM		
		for i in range(0, L):
			for j in range(0,3):
				c1COM[j] = c1COM[j] + c1[i][j]
				c2COM[j] = c2COM[j] + c2[i][j]
		
		T1 = - c1COM / L
		T2 = - c2COM / L
 
		# move everything back to the origin.
		for i in range(0, L):
			for j in range(0,3):
				c1[i][j] = c1[i][j] + T1[j]
				c2[i][j] = c2[i][j] + T2[j]
				
		return T1, T2, c1, c2
 
					
	def calcR( self, c1, c2 ):
		"""
		Calculate the matrix, R, by (eq7, 1976).  M{r_{i,j} = sum_n w_n * y_{ni} * x_{nj}},
		where M{y_{ni}} is the ith component of the vector M{y_n}.
		(CORRECT)
		
		@param c1: coordinats of the first vectors, as an array of 3D vectors.
		@type  c1: Python list
		   
		@param c2: coordinates of the second set of vectors, as an array of 3D vectors.
		@type  c2: Python list
 
		@postcondition: R is now defined.
		
		@return: R
		@rtype : 3x3 matrix
				
		"""
		
		# Create the 3x3 matrix
		R = Numeric.zeros((3,3), Numeric.Float64)
		L = len(c1)
		
		for k in range(0, L):
			for i in range(0, 3):
				for j in range(0, 3):
					# Get the first element of the multiplied array(s) -- S.S.
					R[i][j] = R[i][j] + ((c2[k][i] * c1[k][j])[0]) 
		
		# return R the 3x3 PSD Matrix.
		return R
		
	
	def calcRtR( self, R ):
		"""
		Calculate RtR (R-transpose * R).
		(CORRECT)
		
		@param R: Matrix
		@type  R: 3x3 Matrix
		
		@precondition: R is a the well formed matrix as per Kabsch.
		@postcondition: RtR is now defined
		
		@return: M{R^tR}
		@rtype : 3x3 matrix
		
		"""
		# Note: In the old (linux) code, "matrixmultiply" should be replaced with "dot".
		# Additionally, in the "swapaxes" funtion in Numeric.py (line 250):
		# if axis2 < 0: axis2 = axis1 + n  should be replaced with
		# if axis2 < 0: axis2 = axis2 + n  for "dot" to work correctly
		# -- S.S. 09-28-2005.
		#
		RtR = Numeric.dot(Numeric.transpose(R), R)
		
		return RtR
	
	
	def calcEigenPairs( self, RtR ):
		"""
		Calculate the corresponding eigenpairs for RtR, and sort them accordingly:
		M{m1 >= m2 >= m3}; set M{v3 = v1 x v2} to ensure a RHS
		(CORRECT?)
		
		@postcondition: The eigen pairs are calculated, sorted such that M{m1 >= m2 >= m3} and
		M{v3 = v1 x v2}.
		
		@param RtR: 3x3 Matrix of M{R^t * R}.
		@type  RtR: 3x3 Matrix
		@return : Eigenpairs for the RtR matrix.
		@rtype  : List of stuff
		
		"""
		
		eVal, eVec = LinearAlgebra.eigenvectors(RtR)
 
		# This is cool.  We sort it using Numeric.sort(eVal)
		# then we reverse it using nifty-crazy ass notation [::-1].
		eVal2 = Numeric.sort(eVal)[::-1]
		eVec2 = [[],[],[]] #Numeric.zeros((3,3), Numeric.Float64)
				
		# Map the vectors to their appropriate owners		
		if ( eVal2[0] == eVal[0]):
			eVec2[0] = eVec[0]
			if ( eVal2[1] == eVal[1] ):
				eVec2[1] = eVec[1]
				eVec2[2] = eVec[2]
			else:
				eVec2[1] = eVec[2]
				eVec2[2] = eVec[1]
		elif( eVal2[0] == eVal[1]):
			eVec2[0] = eVec[1]
			if ( eVal2[1] == eVal[0] ):
				eVec2[1] = eVec[0]
				eVec2[2] = eVec[2]
			else:
				eVec2[1] = eVec[2]
				eVec2[2] = eVec[0]
		elif( eVal2[0] == eVal[2]):
			eVec2[0] = eVec[2]
			if ( eVal2[1] == eVal[1] ):
				eVec2[1] = eVec[1]
				eVec2[2] = eVec[0]
			else:
				eVec2[1] = eVec[0]
				eVec2[2] = eVec[1]
 
#              print "v3 = v1 x v2"
#              print eVec2[2]
#              eVec2[2] = a1 CROSS a2
		eVec2[2][0] = eVec2[0][1]*eVec2[1][2] - eVec2[0][2]*eVec2[1][1]
		eVec2[2][1] = eVec2[0][2]*eVec2[1][0] - eVec2[0][0]*eVec2[1][2]
		eVec2[2][2] = eVec2[0][0]*eVec2[1][1] - eVec2[0][1]*eVec2[1][0]
# 	       print eVec2[2]
		
		return [eVal2, eVec2]
		
	
	def calcBVectors( self, R, eVectors ):
		"""
		Calculate M{R*a_k} and normalize the first two vectors to obtain M{b_1, b_2} and set
		M{b_3 = b_1 x b_2}.  This also takes care of {m2 > m3 = 0}. 
		(CORRECT)
		
		@postcondition: b-Vectors are defined and returned
		
		@return: The three B-vectors
		@rtype: List of 3D vectors (Python LOL).
		"""
		
		bVectors = Numeric.zeros((3,3), Numeric.Float64)
 
		for i in range(0,3):
			bVectors[i] = Numeric.matrixmultiply(R, eVectors[i])
 
#              what if the  bVector[0] == 0?
		bVectors[0] = bVectors[0] / Numeric.sqrt(Numeric.add.reduce(bVectors[0]**2))
		bVectors[1] = bVectors[1] / Numeric.sqrt(Numeric.add.reduce(bVectors[1]**2))
		bVectors[2] = bVectors[2] / Numeric.sqrt(Numeric.add.reduce(bVectors[2]**2))
		
#              b2 = b0 cross b1
# 	       print "b3 = b1 x b2"
# 	       print bVectors[2]
		bVectors[2][0] = bVectors[0][1]*bVectors[1][2] - bVectors[0][2]*bVectors[1][1]
		bVectors[2][1] = bVectors[0][2]*bVectors[1][0] - bVectors[0][0]*bVectors[1][2]
		bVectors[2][2] = bVectors[0][0]*bVectors[1][1] - bVectors[0][1]*bVectors[1][0]
#              bVectors[2] = bVectors[0] * bVectors[1]
# 	       print bVectors[2]
		
		return bVectors
		
		
		
	def calcU( self, eVectors, bVectors ):
		"""
		Calculate M{U=(u_{ij})=(sum n b_{ki} * a_{kj})} to obtain the best rotation.  Set
		M{sigma_3 = -1 if b3(Ra3) < 0 else sigma_3 = +1}.
		(CORRECT)
		
		@postcondition: U is defined
		
		@param eVectors: Eigenvectors for the system.
		@type  eVectors: Eigenvectors
		
		@param bVectors: BVectors as described by Kabsch.
		@type  bVectors: BVectors
		
		@return: U the rotation matrix.
		@rtype  :3x3 matrix.
		
		"""
		
		U = Numeric.zeros((3,3), Numeric.Float64)
		
		for k in range(0,3):
			for i in range(0,3):
				for j in range(0,3):
					U[i][j] = U[i][j] + Numeric.matrixmultiply(bVectors[k][i], eVectors[k][j])
		
		return U
	
		
	def calcE0( self, c1, c2 ):
		"""
		Calculates the initial RMSD, which Kabsch called E0.
		(CORRECT)
		
		@param c1: coordinats of the first vectors, as an array of 3D vectors.
		@type  c1: Python list
		   
		@param c2: coordinates of the second set of vectors, as an array of 3D vectors.
		@type  c2: Python list
		
		@return: E0 the initial RMSD.
		@rtype : float.
				
		"""
		
		E0 = 0.0
		
		L = len(c1)
		for i in range( 0, L ):
			for j in range(0, 3):
				E0 = E0 + 0.5*( (c1[i][j]*c1[i][j])+(c2[i][j]*c2[i][j]))
		
		return E0
	
	def calcRMSD( self, E0, eValues, eVectors, bVectors, R):
		"""
		Calculate the RMSD.  The residual error is then
		The M{E = E0 - sqrt(m1) - sqrt(m2) - sigma_3(sqrt(m3))}.
		
		@param E0: Initial RMSD as calculated in L{calcE0}.
		@type  E0: float.
		
		@param eVectors: Eigenvectors as calculated from L{calcEigenPairs}
		@type eVectors: vectors, dammit!
		
		@param bVectors: B-vectors as calc. from L{calcBVectors}
		@type  bVectors: More vectors.
		
		@param R: The matrix R, from L{calcR}.
		@type  R: 3x3 matrix.		
		
		@postcondition: RMSD is computed.
		@return: The RMSD.
		
		"""
		sigma3 = 0
		if ( Numeric.matrixmultiply(bVectors[2], Numeric.matrixmultiply( R, eVectors[2])) < 0):
			sigma3 = -1
		else:
			sigma3 = 1
			
		E = E0 - math.sqrt(eValues[0]) - math.sqrt(eValues[1]) - sigma3*(math.sqrt(eValues[2]))
		
		return E
		
		
	def calcSimpleRMSD( self, c1, c2 ):
		"""
		Calculates the usual concept of RMSD between two set of points.  The CalcRMSD above
		sticks to Kabsch's alignment method protocol and calculates something much different.
		@see kabsch.calcRMSD
		
		@param c1: List of points #1
		@type  c1: LOL
		
		@param c2: List of points #2
		@type  c2: LOL
		
		@return: RMSD between the two
		
		"""
		
		RMSD = 0.0
		for i in range(0, len(c1)):
			for j in range(0,3):
				RMSD = RMSD + (c2[i][j]-c1[i][j])**2
				
		RMSD = RMSD / len(c1)
		RMSD = Numeric.sqrt(RMSD)
		return RMSD
		
		
	#
	#
	# Thus, a typical user session with Kabsch will be:
	# 1. Read in the PDBs
	# c1 = parsePDB(fileName)
	# c2 = parsePDB(fileName2)
	# pairs = getPairs(c1,c2)
	# kasbch = new kabsch()
	# U, T, R = kabsch.align( c1, c2, pairs )
	#
	# The user will be handed back a rotation matrix U,
	# and translation vector T.
	#
	#
	
	#####################################################################
	#
	# UTIL Functions
	def rotatePoints(self, U, c2):
		"""
		Rotate all points in c2 based on the rotation matrix U.
		
		@param U: 3x3 Rotation matrix
		@type  U: 3x3 matrix
		
		@param c2: List of points to rotate
		@type  c2: List of 3D vectors
		
		@return: List of rotated points
		
		"""
	
		L = len(c2)
 
		for n in range(0,L):
			c2[n][0] = c2[n][0] * U[0][0] + c2[n][1] * U[1][0] + c2[n][2] * U[2][0]
			c2[n][1] = c2[n][0] * U[0][1] + c2[n][1] * U[1][1] + c2[n][2] * U[2][1]
			c2[n][2] = c2[n][0] * U[0][2] + c2[n][1] * U[1][2] + c2[n][2] * U[2][2]
		
		return c2
	
	def writeU( self, U, fileName ):
		"""
		Convenience function.  Writes U to disk.
		
		"""
		
		if ( len(fileName) == 0 ):
			fileName = "./U"
			
		outFile = open( fileName, "wb")
		for i in range(0,3):
			for j in range(0,3):
				outFile.write( string.ljust(str(U[i][j]),20) )
			outFile.write("\n")
		outFile.close()		
 
	
def optAlign( sel1, sel2 ):
	"""
	optAlign accurately aligns two selections to the lowest possible RMSD.
	@note: The selections, sel1 and sel2 must be the same length (for now).
	An example selection is: optAlign 1GGZ and i. 4-146 and n. CA, 1CLL and i. 4-146 and n. CA
	Coming soon: Monte Carlo simulation to choose the best pairs to align across the molecules.
	So, I'll soon be able to say, "optAlign 1GGZ, 1CLL".  :)
	
	@param sel1: First PyMol selection with N-atoms
	@param sel2: Second PyMol selection with N-atoms
	"""
	cmd.reset()
 
	# make the lists for holding coordinates
	# partial lists
	stored.sel1 = []
	stored.sel2 = []
	# full lists
	stored.mol1 = []
	stored.mol2 = []
 
	# now put the coordinates into a list
	# partials
	cmd.iterate_state(1, selector.process(sel1), "stored.sel1.append([x,y,z])")
	cmd.iterate_state(1, selector.process(sel2), "stored.sel2.append([x,y,z])")
	# full molecule
	mol1 = cmd.identify(sel1,1)[0][0]
	mol2 = cmd.identify(sel2,1)[0][0]
	cmd.iterate_state(1, mol1, "stored.mol1.append([x,y,z])")
	cmd.iterate_state(1, mol2, "stored.mol2.append([x,y,z])")
 
	K = kabsch()
	U, T1, T2, RMSD, c1, c2 = K.align(stored.sel1, stored.sel2, [])
	
	# Remove all backslashes in case you were using the earlier (linux) code -- S.S.
        # Oops! Sorry for the long line...
	stored.mol2 = map(lambda v:[ T2[0]+((v[0]*U[0][0])+(v[1]*U[1][0])+(v[2]*U[2][0])), T2[1]+((v[0]*U[0][1])+(v[1]*U[1][1])+(v[2]*U[2][1])), T2[2]+((v[0]*U[0][2])+(v[1]*U[1][2])+(v[2]*U[2][2])) ], stored.mol2)
 
	stored.mol1 = map(lambda v:[ v[0]+T1[0], v[1]+T1[1], v[2]+T1[2] ], stored.mol1)
	
	# Concatenate the array of arrays to get a useable array -- S.S.
	cmd.alter_state(1,mol1,"(x,y,z)=Numeric.concatenate(stored.mol1.pop(0))")
	cmd.alter_state(1,mol2,"(x,y,z)=Numeric.concatenate(stored.mol2.pop(0))")
	
	# Get the first element of the RMSD array -- S.S.
	print "RMSD=%f" % RMSD[0]
	
	cmd.hide('everything')
	cmd.show('ribbon', sel1 + ' or ' + sel2)
	cmd.color('magenta', mol1 )
	cmd.color('yellow', mol2 )
	cmd.color('red', 'visible')
	cmd.show('ribbon', 'not visible')
	cmd.center('visible')
	cmd.orient()
	cmd.zoom('visible')
 
cmd.extend("optAlign", optAlign)

References

[Kabsch, 1976] Kabsch, W. (1976). A solution for the best rotation to relate two sets of vectors. Acta. Crystal, 32A:922-923.

[Kabsch, 1978] Kabsch, W. (1978). A discussion of the solution for the best rotation to related two sets of vectors. Acta. Crystal, 34A:827-828.