KaKsTurbo

Model-averaged Ka/Ks estimation with CPU/GPU execution and explicit numerical diagnostics.

Source and documentation · Quick start · Raw CDS/PEP tutorial · Portable downloads

KaKsTurbo estimates nonsynonymous and synonymous substitution rates (Ka and Ks), and their ratio, for batches of homologous codon-aligned gene pairs. It combines structured FP64 matrix computation, boundary-constrained optimization and complete fourteen-model averaging. The statistical model families and averaging framework build on established KaKs methods; this is a computational implementation with numerical validation, not a new biological selection test.

Installation

This is a release candidate, version 1.1.0rc10. Use Python 3.11 or later on Linux x86-64 (glibc >=2.17) or Windows x86-64. ARM and macOS are not supported by these native wheels.

python -m pip install --only-binary=:all: kaksturbo==1.1.0rc10
kaksturbo --version
kaksturbo doctor --backend cpu

The wheel includes the pinned native CPU/GPU engine. CPU execution needs no CUDA installation. Existing portable GitHub downloads include their own Python runtime and remain available separately; this pip distribution uses the Python environment into which it is installed.

Analyze aligned pairs

kaksturbo validate --input aligned_pairs.axt
kaksturbo run --input aligned_pairs.axt --output results_cpu --backend cpu --threads 8
kaksturbo status results_cpu

Supply codon-aligned AXT, or aligned FASTA plus an explicit pair manifest. Raw CDS and protein files require alignment first: the Linux-first batch tutorial provides MAFFT/PAL2NAL preparation scripts, examples and validation. Download those companion files from the repository or portable archive; they are not installed as commands by pip.

Use a new output directory. kaksturbo resume results_cpu continues a compatible interrupted run. Terminal progress reports committed pairs and processing stages; --json provides machine-readable command output and --quiet suppresses progress.

GPU requirements

nvidia-smi
kaksturbo doctor --backend gpu --probe-gpu
kaksturbo run --input aligned_pairs.axt --output results_gpu --backend gpu --device cuda:0

One run uses one NVIDIA GPU. The compiled backend requires compute capability >=8.6, driver CUDA compatibility >=12.9 and >=8 GiB free memory, including a 2 GiB reserve. A separate CUDA Toolkit is not needed. auto selects CPU. An explicit GPU request reports unavailable devices, incompatible drivers or insufficient memory with corrective guidance; it never silently falls back to CPU. See device requirements.

Outputs and interpretation

Missing or unidentifiable estimates remain NA with reasons. Finite output and optimizer success are not convergence certificates. The projected-gradient diagnostic threshold remains 1e-8; unresolved numerical cases are documented. Ka/Ks >1 alone is not a significance test for positive selection. Current scope is MA under standard genetic code 1, with FP64 and all fourteen models preserved.

Optional compatibility export writes a KaKs_Calculator-style table without replacing the standard results:

kaksturbo export results_cpu --format kaks-calculator --output compatible_results

Unsupported or unvalidated statistics are NA with explanations. File-format compatibility does not imply estimator equivalence.

Performance, methods and support

The performance report contains measured full-workflow comparisons, replicate counts, physical-core allocations and limitations. Speedups depend on sequence characteristics and estimator policies; no universal acceleration or cross-platform GPU ratio is claimed.

Methods · Numerical validation · Input/output schema · Build and corresponding source

GPL-3.0-only; third-party license notices are included. Maintainer: Taishan Li, litaishan@caf.ac.cn. Report issues using minimal synthetic inputs. Cite the version and release URL; no paper DOI is currently assigned.