Research map / Probabilistic and scientific modeling / Molecules and proteins
Enhancing Chemical Explainability Through Counterfactual Masking
Łukasz Janisiów, Marek Kochańczyk, Bartosz Michał Zieliński, Tomasz Danel
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace feature zeroing with counterfactual masking (CReM) of non-shared fragments | GIN solubility regressor · Common substructure pairs, single anchor | |delta y| (lower better): 2.08 -> 1.68 | Improved |
| replace feature zeroing with CM (CReM) | GIN solubility regressor · Common substructure pairs, multiple anchors | |delta y| (lower better): 1.60 -> 1.32 | Improved |
| replace feature zeroing with CM (DiffLinker) | GIN solubility regressor · Common substructure pairs, combined | |delta y| (lower better): 1.80 -> 1.55 | Improved |
| use generative CM instead of feature zeroing | masking pipeline · Common substructure pairs | coverage (validity): 100% -> 53% (multiple anchors, CReM) | Hurt |
| mask with CM (CReM) instead of feature zeroing | embeddings of masked molecules · Solubility test distribution | Wasserstein distance to test set | Improved |
| use CM (CReM) instead of MMACE for counterfactual generation | GIN classifiers on CYP3A4, CYP2D6, hERG · Counterfactual generation | synthetic accessibility (lower better): 3.69 -> 3.10 (hERG) | Improved |
| use CM (CReM) instead of MMACE | counterfactual generation · CYP3A4, CYP2D6, hERG | similarity to original: 0.65 -> 0.51 (hERG) | Hurt |
| use DiffLinker instead of CReM as the generator (ablation) | CM counterfactuals · CYP3A4, CYP2D6, hERG | synthetic accessibility: 2.71 -> 3.91 (CYP3A4) | Hurt |
| use CM (CReM) instead of feature zeroing for benchmarking Grad-CAM | XAI evaluation · Consistency score across 5 TDC datasets | consistency score: 84 -> 71 (CYP2D6) | Hurt |
| use CM (DiffLinker) for benchmarking explainers | XAI evaluation · Consistency score across 5 datasets | consistency score: 56 ± 28% (Grad-CAM CYP2D6) vs 50 ± 28% random | No clear effect |
Closest papers by meaning
- Learning Over Molecular Conformer Ensembles: Datasets and Benchmarks (ICLR 2024 · similarity 0.86)
- Faithful Model Explanations through Energy-Constrained Conformal Counterfactuals (AAAI 2024 · similarity 0.86)
- Instructor-inspired Machine Learning for Robust Molecular Property Prediction (NeurIPS 2024 · similarity 0.86)
- Multimodal Molecular Pretraining via Modality Blending (ICLR 2024 · similarity 0.86)
- Swallowing the Bitter Pill: Simplified Scalable Conformer Generation (ICML 2024 · similarity 0.86)
- Representing Molecules as Random Walks Over Interpretable Grammars (ICML 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
