ValueError: b'Extrapolation not allowed with blending'

[FlowerPurr]

hi,
A bug occurred when I was using scanpy for high variant gene arithmetic. I learned that the bug originated from scikit-misc, so I came to you for help.

my code:

adata.layers["counts"] = adata.X.copy() sc.pp.normalize_total(adata, target_sum=1e4) adata adata.raw = adata # keep full dimension safe sc.pp.highly_variable_genes( adata, flavor="seurat_v3",# n_top_genes=2000, layer="counts", batch_key="orig.ident", subset=True, span=1 )

error output

`ValueError Traceback (most recent call last)
Cell In[13], line 3
2 adata.raw = adata # keep full dimension safe
----> 3 sc.pp.highly_variable_genes(
4 adata,
5 flavor="seurat_v3",#
6 n_top_genes=2000,
7 layer="counts",
8 batch_key="orig.ident",
9 subset=True,
10 span=1
11 )
13 filename = 'melanoma_sw_high_var.h5ad'
14 adata.write(filename)

File ~/miniconda3/envs/scanpy/lib/python3.12/site-packages/scanpy/preprocessing/_highly_variable_genes.py:441, in highly_variable_genes(adata, layer, n_top_genes, min_disp, max_disp, min_mean, max_mean, span, n_bins, flavor, subset, inplace, batch_key, check_values)
439 sig = signature(_highly_variable_genes_seurat_v3)
440 n_top_genes = cast(int, sig.parameters["n_top_genes"].default)
--> 441 return _highly_variable_genes_seurat_v3(
442 adata,
443 layer=layer,
444 n_top_genes=n_top_genes,
445 batch_key=batch_key,
446 check_values=check_values,
447 span=span,
448 subset=subset,
449 inplace=inplace,
450 )
452 if batch_key is None:
453 df = _highly_variable_genes_single_batch(
454 adata,
455 layer=layer,
(...)
462 flavor=flavor,
463 )

File ~/miniconda3/envs/scanpy/lib/python3.12/site-packages/scanpy/preprocessing/_highly_variable_genes.py:87, in _highly_variable_genes_seurat_v3(adata, layer, n_top_genes, batch_key, check_values, span, subset, inplace)
85 x = np.log10(mean[not_const])
86 model = loess(x, y, span=span, degree=2)
---> 87 model.fit()
88 estimat_var[not_const] = model.outputs.fitted_values
89 reg_std = np.sqrt(10**estimat_var)

File _loess.pyx:927, in _loess.loess.fit()

ValueError: b'Extrapolation not allowed with blending'`

[has2k1]

This error is makes the most sense when making predictions outside the fitted data (x, the independent observations). There the solution is to use surface="direct" instead of interpolate. i.e. loess(x, y, span=span, degree=2, surface="direct")

In this case it happens when fitting the data which is unexpected. I would need to see the x and y values.