def validate_input_arrays(predictions, measurements, uncertainties, prior_pops=None):
"""Check input data for correct shape and dtype
Parameters
----------
predictions : ndarray, shape = (num_frames, num_measurements)
predictions[j, i] gives the ith observabled predicted at frame j
measurements : ndarray, shape = (num_measurements)
measurements[i] gives the ith experimental measurement
uncertainties : ndarray, shape = (num_measurements)
uncertainties[i] gives the uncertainty of the ith experiment
prior_pops : ndarray, shape = (num_frames), optional
Prior populations of each conformation. If None, skip.
Notes
-----
All inputs must have float64 type and compatible shapes.
"""
num_frames, num_measurements = predictions.shape
ensure_type(predictions, np.float64, 2, "predictions")
ensure_type(measurements, np.float64, 1, "measurements", shape=(num_measurements,))
ensure_type(uncertainties, np.float64, 1, "uncertainties", shape=(num_measurements,))
if prior_pops is not None:
ensure_type(prior_pops, np.float64, 1, "prior_pops", shape=(num_frames,))
def test_ensure_type_25():
with warnings.catch_warnings(record=True) as w:
warnings.simplefilter("always")
val = ensure_type(a, np.float64, 1, '', length=10, warn_on_cast=False)
assert val.dtype == np.float64
assert a.dtype == np.float32 # a should not be changed
assert len(w) == 0 # no warning since we set warn_on_cast to False
def test_ensure_type_2():
with warnings.catch_warnings(record=True) as w:
warnings.simplefilter("always")
val = ensure_type(a, np.float64, 1, '', length=10)
assert val.dtype == np.float64
assert a.dtype == np.float32 # a should not be changed
assert len(w) == 1
assert issubclass(w[-1].category, TypeCastPerformanceWarning)
def compute_translation_and_rotation(mobile, target):
"""Returns the translation and rotation mapping mobile onto target.
Parameters
----------
mobile : ndarray, shape = (n_atoms, 3)
xyz coordinates of a `single` frame, to be aligned onto target.
target : ndarray, shape = (n_atoms, 3)
xyz coordinates of a `single` frame
Returns
-------
translation : ndarray, shape=(3,)
Difference between the centroids of the two conformations
rotation : ndarray, shape=(3,3)
Rotation matrix to apply to mobile to carry out the transformation.
"""
ensure_type(mobile, 'float', 2, 'mobile', shape=(None, 3))
ensure_type(target, 'float', 2, 'target', shape=(target.shape[0], 3))
mu1 = mobile.mean(0)
mu2 = target.mean(0)
translation = mu2
mobile = mobile - mu1
target = target - mu2
correlation_matrix = np.dot(np.transpose(mobile), target)
V, S, W_tr = np.linalg.svd(correlation_matrix)
is_reflection = (np.linalg.det(V) * np.linalg.det(W_tr)) < 0.0
if is_reflection:
V[:, -1] = -V[:, -1]
rotation = np.dot(V, W_tr)
return translation, rotation
def compute_translation_and_rotation(mobile, target):
"""Returns the translation and rotation mapping mobile onto target.
Parameters
----------
mobile : ndarray, shape = (n_atoms, 3)
xyz coordinates of a `single` frame, to be aligned onto target.
target : ndarray, shape = (n_atoms, 3)
xyz coordinates of a `single` frame
Returns
-------
translation : ndarray, shape=(3,)
Difference between the centroids of the two conformations
rotation : ndarray, shape=(3,3)
Rotation matrix to apply to mobile to carry out the transformation.
"""
ensure_type(mobile, 'float', 2, 'mobile', warn_on_cast=False, shape=(None, 3))
ensure_type(target, 'float', 2, 'target', warn_on_cast=False, shape=(target.shape[0], 3))
mu1 = mobile.mean(0)
mu2 = target.mean(0)
translation = mu2
mobile = mobile - mu1
target = target - mu2
correlation_matrix = np.dot(np.transpose(mobile), target)
V, S, W_tr = np.linalg.svd(correlation_matrix)
is_reflection = (np.linalg.det(V) * np.linalg.det(W_tr)) < 0.0
if is_reflection:
V[:, -1] = -V[:, -1]
rotation = np.dot(V, W_tr)
return translation, rotation
def kmeans_mds(xyzlist, k=10, max_iters=100, max_time=10, threshold=1e-8, nearest_medoid=False):
"""k-means clustering with the RMSD distance metric.
this is an iterative algorithm. during each iteration we first move each cluster center to
the empirical average of the conformations currently assigned to it, and then we re-assign
all of the conformations given the new locations of the centers.
to compute the average conformations, we use a form of classical multidimensional
scaling / principle coordinate analysis.
"""
xyzlist = ensure_type(xyzlist, np.float32, 3, name='xyzlist', shape=(None, None, 3), warn_on_cast=False)
# center
for x in xyzlist:
centroid = x.astype('float64').mean(0)
assert centroid.shape == (3,)
x -= centroid
# setup for the rmsd calculation
n_frames, n_atoms = xyzlist.shape[0:2]
xyzlist_irmsd, n_atoms_padded = rmsd.reshape_irmsd(xyzlist)
xyzlist_G = rmsd.calculate_G(xyzlist)
# setup for the clustering stuff
# assignments[i] = j means that the i-th conformation is assigned to the j-th cluster
assignments = -1*np.ones(n_frames, dtype=np.int64)
assignments[0:k] = np.arange(k)
np.random.shuffle(assignments)
# the j-th cluster has cartesian coorinates centers[j]
centers = np.zeros((k, xyzlist.shape[1], 3))
# assignment_dist[i] gives the RMSD between the ith conformation and its
# cluster center
assignment_dist = np.inf * np.ones(len(xyzlist))
# previous value of the clustering score
# all of the clustering scores
scores = [np.inf]
times = [time.time()]
for n in itertools.count():
# recenter each cluster based on its current members
for i in range(k):
structures = xyzlist[assignments == i, :, :]
if len(structures) == 0:
# if the current state has zero assignments, just randomly
# select a structure for it
print 'warning: cluster %5d contains zero structures, reseeding...' % i
print '(if this error appears once or twice at the beginning and then goes away'
print 'don\'t worry. but if it keeps up repeatedly, something is wrong)'
new_center = xyzlist[np.random.randint(len(xyzlist))]
else:
medoid = average_structure(structures)
medoid -= medoid.mean(0)
if nearest_medoid:
# instead of actually using the raw MDS average structure, we choose
# the data point in xyzlist[assignments == i, :, :] that is closest,
# by RMSD, to this MDS structure.
# reshape the medoid for RMSD
medoid = medoid[np.newaxis, :, :]
medoid_g = rmsd.calculate_G(medoid)
medoid_irmsd, _ = rmsd.reshape_irmsd(medoid)
# actually compute the RMSDs
d = IRMSD.rmsd_one_to_all(medoid_irmsd, xyzlist_irmsd[assignments == i, :, :],
medoid_g, xyzlist_G[assignments == i, :, :], n_atoms, 0)
# choose the structure that was closest to be the medoid
medoid = xyzlist[assignments == i, :, :][np.argmin(d)]
centers[i] = medoid
# prepare the new centers for RMSD
centers_G = rmsd.calculate_G(centers)
centers_irmsd, _ = rmsd.reshape_irmsd(centers)
# reassign all of the data
assignments = -1 * np.ones(len(xyzlist))
assignment_dist = np.inf * np.ones(len(xyzlist))
for i in range(k):
d = IRMSD.rmsd_one_to_all(centers_irmsd, xyzlist_irmsd, centers_G, xyzlist_G, n_atoms, i)
where = d < assignment_dist
assignments[where] = i
assignment_dist[where] = d[where]
# check how far each cluster center moved during the last iteration
# and break if necessary
scores.append(np.sqrt(np.mean(np.square(assignment_dist))))
times.append(time.time())
print 'round %3d, RMS radius %8f, change %.3e' % (n, scores[-1], scores[-1] - scores[-2])
if threshold is not None and scores[-2] - scores[-1] < threshold:
print 'score decreased less than threshold (%s). done\n' % threshold
break
if max_iters is not None and n >= max_iters:
print 'reached maximum number of iterations. done\n'
break
if max_time is not None and times[-1] >= times[0] + max_time:
print 'reached maximum amount of time. done\n'
break
print 'RMSD KMeans Performance Summary (py)'
print '------------------------------------'
print 'n frames: %d' % n_frames
print 'n states: %d' % k
print 'mean time per round (s) %.4f' % np.mean(np.diff(times))
print 'stddev time per round (s) %.4f' % np.std(np.diff(times))
print 'total time (s) %.4f' % (times[-1] - times[0])
return centers, assignments, assignment_dist, np.array(scores), np.array(times)
def test_ensure_type_7():
c = ensure_type(a, np.float32, ndim=2, name='', add_newaxis_on_deficient_ndim=True)
assert c.shape == (1, len(a))
def test_ensure_type_6():
val = ensure_type(b, np.float64, 2, '', shape=(10,10))
assert val.flags.c_contiguous is True
def test_ensure_type_5():
ensure_type(a, np.float32, 1, '', length=11, can_be_none=True)
def test_ensure_type_4():
ensure_type(None, np.float64, 1, '', length=11, can_be_none=True)
def test_ensure_type_8():
c = ensure_type(np.zeros((5,10)), np.float32, ndim=2, name='', shape=(None, 10))
assert c.shape == (5, 10)
def _center(conformation):
"""Center the conformation"""
ensure_type(conformation, 'float', 2, 'conformation', warn_on_cast=False, shape=(None, 3))
centroid = np.mean(conformation, axis=0)
centered = conformation - centroid
return centered
def test_ensure_type_12():
ensure_type(np.zeros((2,2)), np.float32, ndim=3)
def test_ensure_type_13():
ensure_type(np.zeros((2,2)), np.float32, ndim=2, name='', shape=(None, None, None))
def test_ensure_type_11():
c = ensure_type(0, np.float32, ndim=1, name='', add_newaxis_on_deficient_ndim=True)
assert c.shape == (1,)
def test_ensure_type_10():
c = ensure_type([0,1], np.float32, ndim=2, name='')
def test_ensure_type_9():
c = ensure_type(np.zeros((5,11)), np.float32, ndim=2, name='', shape=(None, 10))
def _center(conformation):
"""Center the conformation"""
ensure_type(conformation, 'float', 2, 'conformation', shape=(None, 3))
centroid = np.mean(conformation, axis=0)
centered = conformation - centroid
return centered
def test_ensure_type_3():
ensure_type(a, np.float32, 1, '', length=11)
def rmsd_qcp(conformation1, conformation2):
"""Compute the RMSD with Theobald's quaterion-based characteristic
polynomial
Rapid calculation of RMSDs using a quaternion-based characteristic polynomial.
Acta Crystallogr A 61(4):478-480.
Parameters
----------
conformation1 : np.ndarray, shape=(n_atoms, 3)
The cartesian coordinates of the first conformation
conformation2 : np.ndarray, shape=(n_atoms, 3)
The cartesian coordinates of the second conformation
Returns
-------
rmsd : float
The root-mean square deviation after alignment between the two pointsets
"""
ensure_type(conformation1, np.float32, 2, 'conformation1', shape=(None, 3))
ensure_type(conformation2, np.float32, 2, 'conformation2', shape=(conformation1.shape[0], 3))
A = _center(conformation1)
B = _center(conformation2)
if not A.shape[0] == B.shape[0]:
raise ValueError('conformation1 and conformation2 must have same number of atoms')
n_atoms = len(A)
# the inner product of the structures A and B
G_A = np.einsum('ij,ij', A, A)
G_B = np.einsum('ij,ij', B, B)
# print 'GA', G_A, np.trace(np.dot(A.T, A))
# print 'GB', G_B, np.trace(np.dot(B.T, B))
# M is the inner product of the matrices A and B
M = np.dot(B.T, A)
# unpack the elements
Sxx, Sxy, Sxz = M[0, :]
Syx, Syy, Syz = M[1, :]
Szx, Szy, Szz = M[2, :]
# do some intermediate computations to assemble the characteristic
# polynomial
Sxx2 = Sxx * Sxx
Syy2 = Syy * Syy
Szz2 = Szz * Szz
Sxy2 = Sxy * Sxy
Syz2 = Syz * Syz
Sxz2 = Sxz * Sxz
Syx2 = Syx * Syx
Szy2 = Szy * Szy
Szx2 = Szx * Szx
SyzSzymSyySzz2 = 2.0 * (Syz * Szy - Syy * Szz)
Sxx2Syy2Szz2Syz2Szy2 = Syy2 + Szz2 - Sxx2 + Syz2 + Szy2
# two of the coefficients
C2 = -2.0 * (Sxx2 + Syy2 + Szz2 + Sxy2 + Syx2 + Sxz2 + Szx2 + Syz2 + Szy2)
C1 = 8.0 * (Sxx * Syz * Szy + Syy * Szx * Sxz + Szz * Sxy * Syx - Sxx * Syy * Szz - Syz * Szx * Sxy - Szy * Syx * Sxz)
SxzpSzx = Sxz + Szx
SyzpSzy = Syz + Szy
SxypSyx = Sxy + Syx
SyzmSzy = Syz - Szy
SxzmSzx = Sxz - Szx
SxymSyx = Sxy - Syx
SxxpSyy = Sxx + Syy
SxxmSyy = Sxx - Syy
Sxy2Sxz2Syx2Szx2 = Sxy2 + Sxz2 - Syx2 - Szx2
# the other coefficient
C0 = Sxy2Sxz2Syx2Szx2 * Sxy2Sxz2Syx2Szx2 \
+ (Sxx2Syy2Szz2Syz2Szy2 + SyzSzymSyySzz2) * (Sxx2Syy2Szz2Syz2Szy2 - SyzSzymSyySzz2) \
+ (-(SxzpSzx) * (SyzmSzy) + (SxymSyx) * (SxxmSyy - Szz)) * (-(SxzmSzx) * (SyzpSzy) + (SxymSyx) * (SxxmSyy + Szz)) \
+ (-(SxzpSzx) * (SyzpSzy) - (SxypSyx) * (SxxpSyy - Szz)) * (-(SxzmSzx) * (SyzmSzy) - (SxypSyx) * (SxxpSyy + Szz)) \
+ (+(SxypSyx) * (SyzpSzy) + (SxzpSzx) * (SxxmSyy + Szz)) * (-(SxymSyx) * (SyzmSzy) + (SxzpSzx) * (SxxpSyy + Szz)) \
+ (+(SxypSyx) * (SyzmSzy) + (SxzmSzx) * (SxxmSyy - Szz)) * (-(SxymSyx) * (SyzpSzy) + (SxzmSzx) * (SxxpSyy - Szz))
E0 = (G_A + G_B) / 2.0
f = lambda x: x ** 4.0 + C2 * x ** 2. + C1 * x + C0
df = lambda x: 4 * x ** 3.0 + 2 * C2 * x + C1
max_eigenvalue = scipy.optimize.newton(f, E0, df)
rmsd = np.sqrt(np.abs(2.0 * (E0 - max_eigenvalue) / n_atoms))
return rmsd
def rmsd_qcp(conformation1, conformation2):
"""Compute the RMSD with Theobald's quaterion-based characteristic
polynomial
Rapid calculation of RMSDs using a quaternion-based characteristic polynomial.
Acta Crystallogr A 61(4):478-480.
Parameters
----------
conformation1 : np.ndarray, shape=(n_atoms, 3)
The cartesian coordinates of the first conformation
conformation2 : np.ndarray, shape=(n_atoms, 3)
The cartesian coordinates of the second conformation
Returns
-------
rmsd : float
The root-mean square deviation after alignment between the two pointsets
"""
ensure_type(conformation1, np.float32, 2, 'conformation1', warn_on_cast=False, shape=(None, 3))
ensure_type(conformation2, np.float32, 2, 'conformation2', warn_on_cast=False, shape=(conformation1.shape[0], 3))
A = _center(conformation1)
B = _center(conformation2)
if not A.shape[0] == B.shape[0]:
raise ValueError('conformation1 and conformation2 must have same number of atoms')
n_atoms = len(A)
# the inner product of the structures A and B
G_A = np.einsum('ij,ij', A, A)
G_B = np.einsum('ij,ij', B, B)
# print 'GA', G_A, np.trace(np.dot(A.T, A))
# print 'GB', G_B, np.trace(np.dot(B.T, B))
# M is the inner product of the matrices A and B
M = np.dot(B.T, A)
# unpack the elements
Sxx, Sxy, Sxz = M[0, :]
Syx, Syy, Syz = M[1, :]
Szx, Szy, Szz = M[2, :]
# do some intermediate computations to assemble the characteristic
# polynomial
Sxx2 = Sxx * Sxx
Syy2 = Syy * Syy
Szz2 = Szz * Szz
Sxy2 = Sxy * Sxy
Syz2 = Syz * Syz
Sxz2 = Sxz * Sxz
Syx2 = Syx * Syx
Szy2 = Szy * Szy
Szx2 = Szx * Szx
SyzSzymSyySzz2 = 2.0 * (Syz * Szy - Syy * Szz)
Sxx2Syy2Szz2Syz2Szy2 = Syy2 + Szz2 - Sxx2 + Syz2 + Szy2
# two of the coefficients
C2 = -2.0 * (Sxx2 + Syy2 + Szz2 + Sxy2 + Syx2 + Sxz2 + Szx2 + Syz2 + Szy2)
C1 = 8.0 * (Sxx * Syz * Szy + Syy * Szx * Sxz + Szz * Sxy * Syx - Sxx * Syy * Szz - Syz * Szx * Sxy - Szy * Syx * Sxz)
SxzpSzx = Sxz + Szx
SyzpSzy = Syz + Szy
SxypSyx = Sxy + Syx
SyzmSzy = Syz - Szy
SxzmSzx = Sxz - Szx
SxymSyx = Sxy - Syx
SxxpSyy = Sxx + Syy
SxxmSyy = Sxx - Syy
Sxy2Sxz2Syx2Szx2 = Sxy2 + Sxz2 - Syx2 - Szx2
# the other coefficient
C0 = Sxy2Sxz2Syx2Szx2 * Sxy2Sxz2Syx2Szx2 \
+ (Sxx2Syy2Szz2Syz2Szy2 + SyzSzymSyySzz2) * (Sxx2Syy2Szz2Syz2Szy2 - SyzSzymSyySzz2) \
+ (-(SxzpSzx) * (SyzmSzy) + (SxymSyx) * (SxxmSyy - Szz)) * (-(SxzmSzx) * (SyzpSzy) + (SxymSyx) * (SxxmSyy + Szz)) \
+ (-(SxzpSzx) * (SyzpSzy) - (SxypSyx) * (SxxpSyy - Szz)) * (-(SxzmSzx) * (SyzmSzy) - (SxypSyx) * (SxxpSyy + Szz)) \
+ (+(SxypSyx) * (SyzpSzy) + (SxzpSzx) * (SxxmSyy + Szz)) * (-(SxymSyx) * (SyzmSzy) + (SxzpSzx) * (SxxpSyy + Szz)) \
+ (+(SxypSyx) * (SyzmSzy) + (SxzmSzx) * (SxxmSyy - Szz)) * (-(SxymSyx) * (SyzpSzy) + (SxzmSzx) * (SxxpSyy - Szz))
E0 = (G_A + G_B) / 2.0
f = lambda x: x ** 4.0 + C2 * x ** 2. + C1 * x + C0
df = lambda x: 4 * x ** 3.0 + 2 * C2 * x + C1
max_eigenvalue = scipy.optimize.newton(f, E0, df)
rmsd = np.sqrt(np.abs(2.0 * (E0 - max_eigenvalue) / n_atoms))
return rmsd