BGPT: Paper Review: Machine Learning-Driven Drug Repurposing for KRAS G12C and KRAS G12D Inhibition

Explore by Goal

Quick Explanation Copied

Paper focus:

This preprint uses two ML regressors (Random Forest + a neural network) trained on binding-affinity data for WT KRAS, KRAS G12C, and KRAS G12D to prioritize FDA-approved small molecules for repurposing. The paper’s main signal is the identification of candidate drugs with higher predicted pIC50, but it also reports weak consensus between model types and limited mutation-specific overlap in top lists.

Source:

Long Explanation

Machine Learning-Driven Drug Repurposing for KRAS G12C and KRAS G12D Inhibition — Skeptical Visual Review

Preprint being reviewed:

1) Visual map of the workflow (what the paper does)

Built from the preprint’s described ML/data workflow: training on binding data; RDKit descriptors + atomic hybridization features; RF/NN training with CV; FDA-approved filtering; ranking by predicted pIC50 and overlap/agreement; reporting performance, feature importance, and candidate lists.

2) Visual summary of how the paper evaluates models

The paper states that model accuracy is assessed using R² and Mean Squared Error, and it reports that RF and NN show similar R² but NN appears to overfit more (attributed to NN parameter count vs dataset size).
It further notes that performance differences across datasets are partly due to differing dataset sizes (with the G12D dataset described as having more entries, improving training/evaluation).

Schematic visualization only: the excerpt does not provide numeric R² values from the paper’s Figure 1, so numeric claims are avoided. The qualitative interpretation (“similar R²; NN greater overfitting”) is directly from the preprint text.

3) Candidate prioritization: where skepticism is warranted

3.1 Model consensus is limited (mutation specificity not consistently recovered)

The paper reports limited overlap between RF and NN top candidates on the same datasets, and it emphasizes that even at the top-30 level overlap is small and that the G12D dataset shows no overlap in the discussed comparison.
It also explains that some molecules are predicted across WT and mutant datasets, interpreting this as a sign of lack of selectivity (or potential pan-KRAS binding), and it gives examples such as kinase inhibitors predicted in multiple categories.

3.2 Unexpected predictions: known KRAS G12C inhibitors show up for WT

The paper describes an “unexpected” outcome: a WT-trained RF model predicted both adagrasib and sotorasib even though the dataset WT KRAS lacks the cysteine mutation necessary for their classic switch-II covalent binding strategy. The authors interpret this as possible reliance on features not impacted by that cysteine residue, but this is a major red-flag for mechanistic interpretability of the ML-to-biology mapping.

3.3 Why the above matters (skeptical logic)

If the ML model is learning proxies correlated with binding assay artifacts or non-covalent/orthogonal properties rather than the relevant mutation-dependent binding mode, then high predicted pIC50 may not imply mutation-specific inhibition. The authors themselves acknowledge that computational models may not capture biological complexity and that predictions require experimental validation and improved uncertainty quantification.

4) Feature importance: what drives the models (and what we can/can’t conclude)

The paper states that permutation importance identifies top ~10 features, including S sp3 and N sp3 counts and electronic/topological descriptors such as PEOE_VSA, SMR_VSA, and EState_VSA, linking them to atomic hybridization/electronic structure and the induced polarization hypothesis.
Limit of inference: because numeric permutation importance values and uncertainty aren’t included in the supplied excerpt, we should not treat the feature list as quantified causal drivers of mutation-specific binding—only as stated influential descriptors in the trained regression models.

5) Biological interpretation vs dataset artifacts (known/unknown)

What is known (from the preprint text)

The study trains models on binding affinity data for WT KRAS and KRAS mutants G12C/G12D and uses IC50→pIC50 and normalization steps as described.
Model consensus is limited: RF and NN top lists do not overlap for top-10 and have modest overlap at top-30; G12D shows no overlap in the described comparison.
The discussion highlights unexpected predictions for WT (e.g., adagrasib/sotorasib) and interprets them as suggesting model features may be mutation-insensitive in some ways.

Unknowns / disconfirming information that could change the view

Mutation-specific binding mode: the excerpt does not include experimental binding/functional validation for the prioritized candidates. Therefore, whether predicted candidates truly bind mutation-dependent pockets or covalent chemistries is unknown.
Dataset staleness/outdated toxicity annotations: the paper gives an example about methapyrilene being deemed highly toxic/revoked and suggests that database labeling may be incomplete/outdated; this implies broader dataset-quality uncertainty affecting ML ranking.

6) Reproducibility checklist (based on what’s present in the excerpt)

Item	Stated in excerpt?	Impact on reproducibility
Training targets (WT/G12C/G12D)	Yes	Good: defines modeling tasks.
Descriptor set & hybridization feature logic	Partially	Good direction, but exact feature engineering details (code/parameters) aren’t provided in the excerpt.
Cross-validation scheme & NN architecture sizes	Yes	Good: gives CV fold counts and high-level NN configuration.
Candidate ranking criterion	Yes	Ambiguity remains because “model agreement” selection thresholds and overlap construction are not fully enumerated in excerpt.
External validation / wet lab	No	Reproducibility of results depends on ML randomness/seeds and data snapshots; without experimental confirmation, scientific reproducibility is limited to computational replication.

7) Paper-to-next-step: what I would request before trusting candidates

Mutation-selectivity metrics computed explicitly (e.g., comparing predicted or measured binding on WT vs mutant for each candidate) rather than only overlap counts across ranked lists.
Uncertainty quantification and/or calibration of predicted pIC50 so top-ranked compounds have interpretable confidence.
Mechanistic docking/MD (if used) aligned to mutant-binding hypotheses, especially because the WT RF prediction of known G12C covalent inhibitors suggests mutation-insensitive proxies.

Author-specific deep-dives (run on BGPT)

Button targets use the author full-names as extracted from the provided preprint text block.

Feedback:

Updated: April 01, 2026