Sci-kit Learn Invalid Parameter(prevents Hessian and ENM models from being produced)

[JoseBernardS]

Issue description

While running the run_py_capsid_report function from pyCapsid package, an invalid parameter error related to the radius_neighbours_graph function of the scikit-learn package is thrown, as follows:

Error message:
Fetching file 1 / 1 (4oq8)...
Done
Number of protein chains in full structure: 60
Number of protein chains in asymmetric unit: 1
# of residues: 8640
Building hessian for model preset:  U-ENM
Model parameters:  {'cutoff': 7.5, 'gnm': False, 'fanm': 0.1, 'wfunc': 'power', 'base_dist': 1, 'd_power': 0, 'backbone': False, 'k_backbone': 1, 'l_backbone': 1, 'chain_starts': None, 'save_hessian': True, 'save_kirchhoff': False, 'save_cg_path': './4oq8/'}
Finding neighbors within 7.5Å
Traceback (most recent call last):
 File "<string>", line 1, in <module>
 File "/Users/selasisedalo/anaconda3/envs/pycapsid/lib/python3.10/site-packages/pyCapsid/__init__.py", line 149, in run_capsid_report
   kirch, hessian = buildENMPreset(coords, **params_dict['CG'])
 File "/Users/selasisedalo/anaconda3/envs/pycapsid/lib/python3.10/site-packages/pyCapsid/CG.py", line 31, in buildENMPreset
   return buildENM(coords, cutoff=cutoff, fanm=fanm, **kwargs)
 File "/Users/selasisedalo/anaconda3/envs/pycapsid/lib/python3.10/site-packages/pyCapsid/CG.py", line 80, in buildENM
   distGraph = radius_neighbors_graph(tree, cutoff, mode='distance', n_jobs=-1)
 File "/Users/selasisedalo/anaconda3/envs/pycapsid/lib/python3.10/site-packages/sklearn/utils/_param_validation.py", line 203, in wrapper
   validate_parameter_constraints(
 File "/Users/selasisedalo/anaconda3/envs/pycapsid/lib/python3.10/site-packages/sklearn/utils/_param_validation.py", line 95, in validate_parameter_constraints
   raise InvalidParameterError(
sklearn.utils._param_validation.InvalidParameterError: The 'X' parameter of radius_neighbors_graph must be an array-like or an instance of 'sklearn.neighbors._base.RadiusNeighborsMixin'. Got <sklearn.neighbors._ball_tree.BallTree object at 0x11ac41210> instead.

Steps to reproduce the error (prior commands)

Installation process

conda create -n pycapsid -y
conda activate pycapsid
conda install -c luque_lab -c conda-forge pycapsid

Installation proceeded without errors

config file content

[PDB]
pdb = '4oq8' # PDB ID of structure
save_all_path = './4oq8' # where to save the results
[CG]
preset = 'U-ENM' # Model Preset To Use
save_hessian = true # Whether to save the hessian matrix
[NMA]
n_modes = 200 # Number of low frequency modes to calculate
eigen_method = 'eigsh' # eigen method to use
[b_factors]
fit_modes = true # Whether to select the number of modes used to maximize correlation
[QRC]
[VIS]
method = 'chimerax'
chimerax_path = 'C:\Program Files\ChimeraX\bin\ChimeraX.exe'

pyCapsid execution

python -c “from pyCapsid import run_capsid_report;run_capsid_report(‘config.toml’)”

Software versions

Software	Version
Python	3.10.0
Conda	24.1.0
Scikit-learn	1.4.1
MacOs	13.0.1

[colintravisbrown]

Looks like there is a difference in versions of sci-kit-learn. See the difference in parameter options in the function between these two versions:
https://scikit-learn.org/1.3/modules/generated/sklearn.neighbors.radius_neighbors_graph.html#sklearn.neighbors.radius_neighbors_graph
https://scikit-learn.org/stable/modules/generated/sklearn.neighbors.radius_neighbors_graph.html#sklearn.neighbors.radius_neighbors_graph

I suggest either locking the scikit-learn version to <1.4 or refactoring how the neighbor graph is computed to use scikit-learns API in a way that isn't version dependent.