EGFR potency prediction - an inspectable pilot

Research question

How much predictive performance survives separation by publication and by distant chemistry?

Determine whether public human EGFR activity measurements support a useful small compound-potency model. Assemble reasonably comparable biochemical IC50 observations, evaluate predictions for compounds withheld from training, challenge apparent performance with related-compound and source/assay holdouts, and retain enough evidence to inspect and rerun the calculations. This is an exploratory historical pilot, not a validated predictor for new EGFR chemistry or an assay-specific decision tool.

Data and comparison design

An ordinary compound holdout can retain assay and publication context shared with training. The combined holdout makes that transfer harder. Compare absolute pIC50 error with within-assay rank correlation so that apparently acceptable pooled error does not obscure poor ordering within an assay.

Input or source	Role in this case	Source context
ChEMBL human EGFR target CHEMBL203	Public activity measurements and source identifiers	Credit ChEMBL; this run used a frozen, ordered convenience sample, not the complete target dataset

Recorded findings

Evaluation	Recorded result	Interpretation
Curated pilot	404 compounds	Convenience sample from the first 1,000 records.
Ordinary compound holdout MAE	0.502 pIC50	Easier retrospective partition.
Publication + distant-chemistry MAE	0.788 pIC50	Harder transfer across publication and chemistry.
Median within-assay rank correlation	0.179	Weak ordering can coexist with a usable-looking pooled error.

The saved report describes six prespecified five-fold holdout protocols, including random compounds, scaffolds, chemical-similarity groups, assays, publications, and publications with close training analogues removed. The ridge model's mean absolute error was 0.502 pIC50 units for ordinary compound holdouts and 0.788 for the combined publication-plus-distant-chemistry holdout. In that combined test it improved over two simple baselines, but median within-assay Spearman correlation across 14 sufficiently populated assays was only 0.179. Thus the saved evidence is more encouraging for broad potency trends than reliable fine ranking within an unfamiliar assay or chemical series. The first-1,000-record convenience sample, missing wild-type annotation and incomplete assay protocols further limit generalization.

What the comparison shows

The harder holdout and weak within-assay ranking expose limits of the pilot. Inspectability makes those limits easier to examine; it does not itself make the model ready for compound-selection decisions.

Assumptions and limits

The first-1,000-record convenience sample and incomplete assay conditions limit generalization. This is an inspectable retrospective pilot, not a prospective predictor, mutant-EGFR validation or assay-specific decision tool.

Sources and research materials

The unchanged research package comes from upstream Case 14. It includes the archived session, analysis materials, results and figures. The poster uses an AI-generated illustrative photograph. This introduction summarizes archived findings; the analysis was not rerun.