Seqcore

_{Reading one letter at a time, versus reading the whole batch at once. Watch the full 90-second explainer.}

High-performance biological sequence analysis library for Python.

A unified, NumPy-vectorized library for genomics, proteomics, structural biology, and drug design.

Note on GPU support: Seqcore ships CuPy device-management utilities (gpu_available, gpu_info, device, set_memory_limit, clear_gpu_cache), but the analysis kernels themselves still execute on NumPy. GPU dispatch for the compute functions is planned, not implemented -- see Roadmap.

Installation

pip install seqcore

With GPU support:

pip install seqcore[gpu]

With all optional dependencies:

pip install seqcore[full]

Quick Start

import seqcore as sc

# DNA sequences - efficient 2-bit encoding
dna = sc.DNAArray("ACGTACGTACGT" * 1_000_000)

# Batch operations
sequences = sc.DNAArray([
    "ACGTACGT",
    "TGCATGCA",
    "GGGGCCCC",
])

# Vectorized operations
gc = sc.gc_content(sequences)
lengths = sc.length(sequences)
rev_comp = sc.reverse_complement(sequences)

# Translation
proteins = sc.translate(sequences)

Features

Sequence Operations

# GC content, molecular weight, length
gc = sc.gc_content(dna)
mw = sc.molecular_weight(protein)

# Transcription and translation
rna = sc.transcribe(dna)
protein = sc.translate(dna, frame=0)

# K-mer operations
kmers = sc.extract_kmers(sequences, k=21)
kmer_counts = sc.count_kmers(sequences, k=21)

Sequence Alignment

# Pairwise alignment
result = sc.align(query, reference)
print(result.score, result.identity, result.cigar)

# Distance matrices
dm = sc.pairwise_distance(sequences, metric="edit")

# Pattern matching
matches = sc.find_pattern(sequences, "ATG[ACGT]{30,100}TAA")

File I/O

# Auto-detect format
data = sc.read("sequences.fasta")
data = sc.read("structure.pdb")
data = sc.read("reads.fastq.gz")

# Streaming for large files
for batch in sc.read_stream("huge.fastq.gz", batch_size=100_000):
    results = process(batch)

# Database fetching
seq = sc.fetch("NP_000509")      # NCBI/UniProt
structure = sc.fetch("1ABC")     # PDB

Structural Biology

# Load structure
structure = sc.read("protein.pdb")

# Access data
print(structure.chains)      # ['A', 'B']
print(structure.n_residues)  # 265

# Distance matrix
dm = sc.distance_matrix(structure, selection="CA")

# Find contacts
contacts = sc.find_contacts(structure, cutoff=4.0)

# RMSD calculation
rmsd = sc.rmsd(structure1, structure2, align=True)

# Surface analysis
sasa = sc.sasa(structure)
surface = sc.surface_residues(structure, threshold=25.0)

# Binding pockets
pockets = sc.find_pockets(structure)

Drug Design

# Small molecules
mol = sc.Molecule.from_smiles("CCO")

# Molecular properties
# Note: sc.molecular_weight is the sequence version. For molecules, import the
# molecules one directly -- the two share a name and the sequence one wins.
from seqcore.molecules import molecular_weight
mw = molecular_weight(molecules)
logp = sc.logp(molecules)
hbd = sc.h_bond_donors(molecules)

# ADMET filters
passes_lipinski = sc.lipinski_filter(molecules)
bbb_permeable = sc.bbb_filter(molecules)

# Fingerprints and similarity
fps = sc.morgan_fingerprint(molecules, radius=2)
similarity = sc.tanimoto_similarity(fps)

# Substructure search
matches = sc.substructure_search(molecules, "c1ccccc1")

Phylogenetics

# Tree construction
tree = sc.neighbor_joining(sequences)
tree = sc.upgma(sequences)

# Tree operations
print(tree.newick())
dist = tree.distance("Species_A", "Species_B")
subtree = tree.prune(["A", "B", "C"])

Population Genetics

# Variant analysis
variants = sc.read("variants.vcf")
af = sc.allele_frequency(variants)
maf = sc.minor_allele_frequency(variants)

# Population statistics
fst = sc.fst(pop1, pop2)
pi = sc.nucleotide_diversity(sequences)
d = sc.tajimas_d(sequences)

# Linkage disequilibrium
ld = sc.linkage_disequilibrium(variants)

GPU Utilities

These manage CuPy devices and memory. They do not move Seqcore's analysis functions onto the GPU -- those run on NumPy today.

# Check GPU availability
if sc.gpu_available():
    print(sc.gpu_info())

# Select the active CuPy device for your own CuPy code
with sc.device("cuda:0"):
    xp = sc.core.device.get_array_module()  # cupy when a GPU is active

# Memory management
sc.set_memory_limit("8GB")
sc.clear_gpu_cache()

# Timing (works on any code)
with sc.timer() as t:
    result = sc.align(sequences, reference)
print(f"Completed in {t.elapsed:.2f}s")

Interoperability

# NumPy
arr = sequences.to_numpy()
sequences = sc.DNAArray.from_numpy(arr)

# pandas
df = sequences.to_dataframe()
df = structure.to_dataframe()

# Biopython
bio_seq = sequences[0].to_biopython()
sc_seq = sc.DNAArray.from_biopython(bio_seq)

# RDKit
rdkit_mol = molecule.to_rdkit()
sc_mol = sc.Molecule.from_rdkit(rdkit_mol)

Performance

Seqcore stores a batch of sequences as one padded integer matrix, so batch-wide operations are single NumPy calls rather than per-sequence loops. That wins decisively where work amortizes across the batch, and loses where a compiled per-element implementation already exists. Both cases are reported below.

100,000 sequences x 1000 bp. Speedup is Seqcore vs the faster of Biopython and idiomatic pure Python; values below 1.0 mean Seqcore is slower.

Both machines run Python 3.12, NumPy 2.5.2, Biopython 1.88; best of 5 runs, each implementation in a fresh subprocess. Raw results in benchmarks/results/.

Every conclusion holds on both machines — the same operations win and lose, in the same order — but the magnitudes differ, so treat any single speedup figure as a point estimate rather than a property of the library.

The Mac mini is 1.7–2.2x faster than the cloud instance in absolute terms. We still quote the cloud instance in the paper, because it is the reproducible one: its ratios agree to within 2% across runs made 14 hours apart, whereas the same ratios on a desktop that is also being used for other work moved by 35–52% between an idle session and a busy one. If you benchmark on your own laptop, measure it twice.

Reproduce with:

python benchmarks/benchmark_suite.py

Domain modules

The molecules, structural biology and phylogenetics modules are benchmarked separately by benchmarks/benchmark_modules.py against RDKit, SciPy, Biopython and a NumPy reference.

The molecule functions wrap RDKit and are not expected to beat it — the goal is that they no longer charge a large multiple for the convenience.

python benchmarks/benchmark_modules.py

When to use something else

• Pairwise alignment is a NumPy anti-diagonal wavefront. Far faster than a scalar Python loop, but still 15–29x slower than Biopython's C aligner. For alignment-bound work, use a dedicated aligner.
• Reverse complement is memory-bandwidth bound and sits at rough parity with Biopython's str.translate. No vectorization advantage is available.
• k-mer counting switches to a Counter over strings once 4**k exceeds the number of windows, because past that point almost every k-mer is unique.
• Tree building is still an O(n³) Python loop. With metric="hamming" that, rather than the distance matrix, is the bottleneck above a few hundred taxa.
• nucleotide_diversity and tajimas_d are O(n²) over sequence pairs (0.21s for 160 sequences).
• GPU: not used by any compute path. See below.

Choosing a phylogenetic distance metric

neighbor_joining() and upgma() default to metric="identity", which aligns every pair with Needleman–Wunsch. That is correct for unaligned sequences and costs O(n²L²).

If your input is already an alignment, pass metric="hamming":

tree = sc.neighbor_joining(aligned, metric="hamming")   # 48 taxa: 6.2s -> 0.016s

The two are not interchangeable — once sequences diverge enough for the aligner to open gaps they give different distances (up to 0.24 apart in our tests), which is why the faster path is opt-in rather than the default.

Requirements

• Python 3.9+
• NumPy 1.22+

Optional:

• CuPy (GPU acceleration)
• Biopython (interoperability)
• RDKit (molecular operations)
• MDAnalysis (structure analysis)

Roadmap

• GPU dispatch for the core kernels (gc_content, translate, reverse_complement, k-mer counting) via CuPy, wired through get_array_module.
• Compiled pairwise alignment kernel to replace the current pure-NumPy dynamic programming implementation.
• Published API reference on Read the Docs.

Contributing

Contributions welcome. See CONTRIBUTING.md.

License

MIT License. See LICENSE.

Author

Dr. Pritam Kumar Panda Stanford University Email: pritam@stanford.edu

Citation

If you use Seqcore in your research, please cite:

@software{seqcore,
  author = {Panda, Pritam Kumar},
  title = {Seqcore: High-performance biological sequence analysis},
  url = {https://github.com/pritampanda15/seqcore},
  version = {0.5.0},
  year = {2026},
  institution = {Stanford University}
}