py-dyneval

A Python port of dynverse/dyneval (Saelens et al. Nat Biotechnol 2019) — the metric library used by the dynbenchmark trajectory inference benchmark.

• Pure NumPy / SciPy / NetworkX / scikit-learn — no R
• Depends on pydynwrap for trajectory wrappers + geodesic distances
• 8/14 dyneval metric IDs ported in v0.1

Install

pip install pydyneval-bio   # pulls in pydynwrap-bio

Quick-start

import pydyneval as de
import pydynwrap as dw
import numpy as np

gold = dw.wrap_data([f"c{i}" for i in range(30)])
gold = dw.add_branch_trajectory(gold,
    pseudotime=np.tile(np.linspace(0,1,10), 3),
    branch=np.repeat(["A","B","C"], 10))

pred = dw.wrap_data([f"c{i}" for i in range(30)])
pred = dw.add_branch_trajectory(pred,
    pseudotime=np.tile(np.linspace(0,1,10), 3),
    branch=np.repeat(["A","B","C"], 10))

scores = de.calculate_metrics(gold, pred)
print(scores)
#    correlation  him  edge_flip  isomorphic  F1_branches  F1_milestones ...
# 0          1.0  1.0        1.0         1.0          1.0           1.0

Function map

R parity on canonical fixture (3-branch, 30 cells × 50 genes)

| Metric | R | Py | |Δ| | |---|---|---|---| | correlation | 0.9812 | 0.9813 | 0.0001 | | him | 1.0 | 1.0 | 0.0 | | edge_flip | 1.0 | 1.0 | 0.0 | | isomorphic | 1.0 | 1.0 | 0.0 | | F1_branches | 1.0 | 1.0 | 0.0 | | F1_milestones | 0.6667 | 0.6667 | 0.0 | | featureimp_cor | 0.9622 | 0.9005 | 0.0617 | | featureimp_wcor | 0.9664 | 0.9548 | 0.0115 |

featureimp_* differs slightly because of Random Forest non-determinism between R randomForest and sklearn RandomForestRegressor — both implement Breiman 2001 with the same hyperparameters; tree-splitting RNG differs.

Citation

Saelens, W. et al. A comparison of single-cell trajectory inference methods. Nat Biotechnol 37, 547–554 (2019).

License

MIT.