Neighbor joining is similar to UPGMA/WPGMA, but infers unrooted trees. As a consequence, and unlike UPGMA/WPGMA, it does not require that the multiple sequence alignment (MSA) has been generated according to a molecular clock along an ultrametric tree.

There are a few differences from UPGMA/WPGMA.

Choosing which nodes to merge

As with UPGMA, two nodes at a time are merged. However instead of choosing them based on the off-diagonal element of the distance matrix with the lowest value, a second matrix called the Q-matrix is derived from the distance matrix. Then the nodes which correspond to the off-diagonal element of the Q-matrix with the lowest value are merged.

Each element i, j of the Q-matrix is calculated as (n - 2) × di,j - Σk di,k - Σk di,k where d is the distance matrix, n is the length of the distance matrix along one dimension, and k is the kth index of the distance matrix.

Calculating branch lengths

As with UPGMA, we merge two nodes at a time, reducing the pool of available nodes by one for each iteration. The calculation of branch lengths is different because we no longer assume that the tree is ultrametric, so separate branch lengths must be estimated for each child.

Given a node whose children A and B correspond to the lowest value off-diagonal element with the indices f, g, we can calculate the branch length of A (LA), and then derive the branch length of B (LB) as dA, B - LA.

LA = df,g / 2 + (Σk df,k - Σk dg,k) / 2(n - 2)

Calculating new genetic distances

After each step shrinking the matrix, as with UPGMA we have to calculate new distances in the matrix to the new node. For the kth element of the matrix this is simply (df,k + dg,k - d(f,g)) / 2.

Besides the above differences, there is one less node in an unrooted tree so there is one less shrinking of the matrix. Instead at the last step, in addition to the two branch lengths we normally calculate, we also calculate a third branch length for the last taxon which we can call child C: (LC) = dA, C - LA.

I have implemented neighbor joining in Python as an example. This code reads in a PHYLIP formatted MSA with the filename “alignment.phy”, uses neighbor joining to estimate a tree under the Jukes–Cantor model of nucleotide evolution, and writes the result as a Newick format string with the filename “nj.tree”.

import numpy

# A function to print out a matrix (e.g. a distance matrix) for human viewing
def print_matrix(title, matrix_map, matrix):
	print(title + ":")
	print("  " + "".join(["%7d" % (column) for column in matrix_map]))
	for row_i, row in enumerate(matrix):
		print("%2d" % (matrix_map[row_i]) + "".join(["%7.2f" % (element) for element in row]))
	print("")

# A function to write a tree to a file as a Newick-format string 
def write_tree(newick_path, tree, taxon_labels):
	newick_string = make_newick_string(len(tree) - 1, tree, taxon_labels) + ";"

	newick_file = open(newick_path, "w")
	newick_file.write(newick_string)
	newick_file.close()

# Recursively build a Newick-format string from an adjacency list
def make_newick_string(node_i, tree, taxon_labels):
	if len(tree[node_i]) == 0: # no outgoing edges, so must be a leaf node
		return taxon_labels[node_i]
	else: # an internal node
		newick_substrings = []
		for child_i in sorted(tree[node_i]):
			branch_length = tree[node_i][child_i]
			substring = make_newick_string(child_i, tree, taxon_labels)
			newick_substrings.append("%s:%f" % (substring, branch_length))

		return "(" + ",".join(newick_substrings) + ")"

# Read in a PHYLIP-format multiple sequence alignment
def read_phylip(phylip_path):
	phylip_file = open(phylip_path)
	phylip_header = phylip_file.readline().strip().split()

	n_taxa = int(phylip_header[0])
	sequence_length = int(phylip_header[1])

	sequence_labels = []
	msa = numpy.zeros((n_taxa, sequence_length), dtype = "uint8")

	for i in range(n_taxa):
		line = phylip_file.readline()
		label, sequence = line.strip().split()

		sequence_labels.append(label)
		msa[i] = numpy.fromstring(sequence, dtype = "uint8")

	return (sequence_length, sequence_labels, msa)

# Find the coordinates of whatever off-diagonal element of a matrix has the lowest value
def get_lowest_off_diagonal_value_coordinate(matrix):
	lowest_value = None
	lowest_value_coordinate = None

	for i, row in enumerate(matrix):
		for j, value in enumerate(row):
			if i != j:
				if lowest_value == None or value < lowest_value:
					lowest_value = value
					lowest_value_coordinate = (i, j)

	return lowest_value_coordinate

# Compute the neighbor joining Q-matrix from a distance matrix
def compute_q_matrix(distance_matrix):
	q_matrix = (matrix_length - 2.0) * distance_matrix

	for i in range(matrix_length):
		for j in range(matrix_length):
			if i != j:
				q_matrix[i][j] -= numpy.sum(distance_matrix[i]) + numpy.sum(distance_matrix[j])

	return q_matrix

# Read in the multiple sequence alignment
sequence_length, taxon_labels, msa = read_phylip("alignment.phy")
n_taxa = len(taxon_labels)
n_nodes = n_taxa + n_taxa - 2

matrix_length = n_taxa
p_distance_matrix = numpy.zeros((matrix_length, matrix_length))

for i in range(n_taxa):
	msa_i = msa[i]
	for j in range(n_taxa):
		msa_j = msa[j]
		identity = float(numpy.sum(msa_i == msa_j))
		p_distance_matrix[i][j] = 1.0 - identity / sequence_length

matrix_map = [n for n in range(n_taxa)] # mapping matrix rows and columns to node indices
distance_matrix = -0.75 * numpy.log(1.0 - 1.3333333333 * p_distance_matrix) # using the Jukes-Cantor 1969 (JC69) model

print_matrix("P-distance matrix", matrix_map, p_distance_matrix)
print_matrix("Distance matrix", matrix_map, distance_matrix)

tree = []
for i in range(n_nodes):
	tree.append({})

for u in range(n_taxa, n_nodes): # we call internal nodes "u"
	if u == n_nodes - 1:
		f, g = 0, 1 # when this is the seed node, don't have to find the next nodes to branch off
	else:
		q_matrix = compute_q_matrix(distance_matrix)
		f, g = get_lowest_off_diagonal_value_coordinate(q_matrix) # these are the next nodes to branch off

	fg_distance = distance_matrix[f][g]
	f_length = 0.5 * fg_distance + (numpy.sum(distance_matrix[f]) - numpy.sum(distance_matrix[g])) / (2.0 * (matrix_length - 2))
	g_length = fg_distance - f_length

	# add the edges and branch lengths
	tree[u][matrix_map[f]] = f_length
	tree[u][matrix_map[g]] = g_length

	# if this is the seed node, fill in the last root branch length and stop calculating
	if u == n_nodes - 1:
		tree[u][matrix_map[2]] = distance_matrix[0][2] - f_length
		break

	new_distance_matrix = numpy.zeros((matrix_length - 1, matrix_length - 1))

	# a and b are the old indices, i and j are the new indices
	i = 0
	new_matrix_map = [u]
	for a in range(matrix_length):
		if (a != f) and (a != g): # skip the rows to be merged
			i += 1
			j = 0

			new_matrix_map.append(matrix_map[a])

			ua_distance = 0.5 * (distance_matrix[f][a] + distance_matrix[g][a] - fg_distance)
			new_distance_matrix[0][i] = ua_distance
			new_distance_matrix[i][0] = ua_distance

			for b in range(matrix_length): # skip the columns to be merged
				if (b != f) and (b != g):
					j += 1
					new_distance_matrix[i][j] = distance_matrix[a][b]

	print_matrix("Distance matrix", new_matrix_map, new_distance_matrix)

	distance_matrix = new_distance_matrix
	matrix_map = new_matrix_map
	matrix_length = matrix_length - 1

# save the result
output_path = "nj.tree"
write_tree(output_path, tree, taxon_labels)