Nested containment list

The Nested Containment List is a datastructure for interval overlap queries, like the interval tree. It is usually an order of magnitude faster than the interval tree both for building and query lookups.

The implementation here is a revived version of the one used in the now defunct PyGr library, which died of bitrot. I have made it less memory-consuming and created wrapper functions which allows batch-querying the NCLS for further speed gains.

It was implemented to be the cornerstone of the PyRanges project, but I have made it available to the Python community as a stand-alone library. Enjoy.

Original Paper: https://academic.oup.com/bioinformatics/article/23/11/1386/199545 Cite: http://dx.doi.org/10.1093/bioinformatics/btz615

Install

pip install ncls

Usage

from ncls import NCLS

import pandas as pd

starts = pd.Series(range(0, 5))
ends = starts + 100
ids = starts

subject_df = pd.DataFrame({"Start": starts, "End": ends}, index=ids)

print(subject_df)
#    Start  End
# 0      0  100
# 1      1  101
# 2      2  102
# 3      3  103
# 4      4  104

ncls = NCLS(starts.values, ends.values, ids.values)

# python API, slower
it = ncls.find_overlap(0, 2)
for i in it:
    print(i)
# (0, 100, 0)
# (1, 101, 1)

starts_query = pd.Series([1, 3])
ends_query = pd.Series([52, 14])
indexes_query = pd.Series([10000, 100])

query_df = pd.DataFrame({"Start": starts_query.values, "End": ends_query.values}, index=indexes_query.values)

query_df
#        Start  End
# 10000      1   52
# 100        3   14


# everything done in C/Cython; faster
l_idxs, r_idxs = ncls.all_overlaps_both(starts_query.values, ends_query.values, indexes_query.values)
l_idxs, r_idxs
# (array([10000, 10000, 10000, 10000, 10000,   100,   100,   100,   100,
#          100]), array([0, 1, 2, 3, 4, 0, 1, 2, 3, 4]))

print(query_df.loc[l_idxs])
#        Start  End
# 10000      1   52
# 10000      1   52
# 10000      1   52
# 10000      1   52
# 10000      1   52
# 100        3   14
# 100        3   14
# 100        3   14
# 100        3   14
# 100        3   14
print(subject_df.loc[r_idxs])
#    Start  End
# 0      0  100
# 1      1  101
# 2      2  102
# 3      3  103
# 4      4  104
# 0      0  100
# 1      1  101
# 2      2  102
# 3      3  103
# 4      4  104

# return intervals in python (slow/mem-consuming)
intervals = ncls.intervals()
intervals
# [(0, 100, 0), (1, 101, 1), (2, 102, 2), (3, 103, 3), (4, 104, 4)]

Benchmark

Test file of 100 million intervals (created by subsetting gencode gtf with replacement):

Library	Function	Time (s)	Memory (GB)
bx-python	build	161.7	2.5
ncls	build	3.15	0.5
bx-python	overlap	148.4	4.3
ncls	overlap	7.2	0.5

Building is 50 times faster and overlap queries are 20 times faster. Memory usage is one fifth and one ninth.

Cite

http://dx.doi.org/10.1093/bioinformatics/btz615

Original paper