chemexp 0.2.4

An explainer for Chemprop based on LIME and PathExplain.

Chemexp

A Chemprop explainer based on LIME and PathExplain.

Installation

• Make sure your Python version is >= 3.8
• conda is required for ChemProp. If it is not installed on your machine, you can take a look at miniconda

1. Install Chemprop
2. Then simply run pip install chemexp
3. You're good to go!

Explanation

Using Python

import chemexp
from chemprop.utils import load_checkpoint

model = chemexp.ExplanationModel(load_checkpoint("models/fusion_db_1.pt"))
exp = model.explain_molecule("COC(=N)Cc1ccccc1")
chemexp.exp_to_png(exp, "mol.png")

A little more detail is available in test/test.py.

Web interface

Although you can use this Python module in your scripts, you can also experiment a user-friendly interface.
⚠️ Warning: this only works on Linux for the moment (this is due to the linux-specific paths used to save some files, but it can be adapted for Windows or Mac).

Use the following command to run the web server:

python -m chemexp <path>

where path is a folder containing chemprop models / checkpoints (.pt files)

Authors

• BELLANGER Clément
• BERNIER Fabien
• MAIGRET Bernard
• NAPOLI Amedeo

To cite our work:

@mastersthesis{bernier:hal-03371070,
  TITLE = {{A Study about Explainability and Fairness in Machine Learning and Knowledge Discovery}},
  AUTHOR = {Bernier, Fabien},
  URL = {https://hal.inria.fr/hal-03371070},
  PAGES = {58},
  SCHOOL = {{TELECOM Nancy}},
  YEAR = {2021},
  MONTH = Sep,
  KEYWORDS = {machine learning ; explainability ; fairness ; antibiotic classification ; machine learning ; explicabilit{\'e} ; fairness ; classification antibiotique},
  PDF = {https://hal.inria.fr/hal-03371070/file/M_moire_ing_nieur_Fabien.pdf#chapter.5},
  HAL_ID = {hal-03371070},
  HAL_VERSION = {v1},
}