kmayerb
commented
3 years ago Andreas,
Great question, sorry for the slower than average response:
Assuming you have full matrices and not sparse ones, combined Alpha and Beta TCRdists into one distance is as simple as
tr.pw_alpha + tr.pw_betaHowever, if you have more than 10K clones, you might be using the sparse representation that will drop all values less than some pre specified value, you will need to be careful combining those: Try this, method shown below, but this is not a fully supported feature yet. Let me know if this resolves your question?
from scipy import sparse
def add_sparse_pwd(a, b, neg1_fmt=False):
"""
Add sparse pairwise distance matrices a and b.
Can work with two "formats": true zero distances as -1
or true zero distances as 0. Returns same format as received.
Does not attempt to add sparse zeros (ie NA),
so returns an intersection of data in a and b (summed).
Returns a scipy.sparse.csr matrix (can take csr or any
format as inputs).
Parameters
----------
a, b : scipy.sparse.csr or coo
Sparse pairwise distance matrices. D=0 can be
represented as -1 or 0 but should be specified
using neg1_fmt
neg1_fmt : bool
Indicator whether to assume D=0 is represented as -1
Returns
-------
tot : scipy.sparse.csr
Sparse matrix representing sum of a and b
where a and b both have values.
"""
a_coo = a.tocoo()
b_coo = b.tocoo()
nr, nc = a_coo.shape
ind_a = a_coo.row.astype(np.int64) * nc + a_coo.col
ind_b = b_coo.row.astype(np.int64) * nc + b_coo.col
# ind_ab = np.array(tuple(set(ind_a).intersection(set(ind_b))))
ind_ab = np.intersect1d(ind_a, ind_b, assume_unique=True)
keep_a = np.in1d(ind_a, ind_ab, assume_unique=True)
keep_b = np.in1d(ind_b, ind_ab, assume_unique=True)
if neg1_fmt:
a_data = a_coo.data[keep_a]
a_data[a_data == -1] = 0
b_data = b_coo.data[keep_b]
b_data[b_data == -1] = 0
tot = a_data + b_data
else:
tot = a_coo.data[keep_a] + b_coo.data[keep_b]
"""Need to add -1 as placeholder upon creation of new sparse matrix"""
tot[tot == 0] = -1
aplusb = sparse.coo_matrix((tot,
(a_coo.row[keep_a], a_coo.col[keep_a])), shape=(nr, nc))
aplusb = aplusb.tocsr()
if not neg1_fmt:
"""Convert it back to 0 = true zero distance"""
aplusb.data[aplusb.data == -1] = 0
return aplusb
## TEST
import numpy as np
from scipy.sparse import csr_matrix
A = np.matrix([ [1, 2, 3],
[3, 4, 0]])
B = np.matrix([ [9, 8, 0],
[0, 1, 0]])
As = csr_matrix(A)
Bs = csr_matrix(B)
Cs = As + Bs
# Regular Addition
Cs.todense()
np.all(Cs.todense() == A + B)
print(Cs.todense())
# TCRDIST ADDITION OF SPARES (ANY ZERO STAY ZEROS?)
Ns = add_sparse_pwd(As, Bs)
print(Ns.todense())
andreas-wilm
KJMaroney
KevinMaroney
CaryBond
Dear Koshlan and TCRdist3 devs,
I would like to know whether it is possible to get a distance matrix from TCRdist3 that represents the combined distance of the alpha and beta chains. It's clear from the documentation how to get pairwise beta and alpha distances (
tr.pw_alphaandtr.pw_betaresp), but I was thinking of one score (in case of paired input data) as shown in this TCRdist2 figure. Or can I simply add/average the two and that's it?Apologies in advance if this clear from the documentation or if I misunderstood this.
Many thanks, Andreas