Data noise caps accuracy of cyclodextrin binding models; a diffusion-based size descriptor adds nothing over McGowan volume
- Reproducibility of compiled cyclodextrin inclusion constants was assessed at three levels: within a single calorimetric study, across laboratories and methods, and among replicated host–guest pairs in a standard compilation, which showed a heavy tail of large outliers.
- Propagating this noise through a small modeling set implies an out-of-sample R2 ceiling well below published values; response noise explains most of the leave-one-out error.
- A size descriptor from measured diffusion coefficients was statistically indistinguishable from McGowan volume (each with log P), and both degraded when whole chemical families were held out.
💡 Points Based on the abstract
Before we trust any R2 from a literature-derived affinity set, I'd want the replicate-based ceiling computed for it. Do it on your own assay-derived data: take the duplicate measurements, estimate the noise variance, and report the achievable R2 next to every model. Also, a 'new' descriptor that ties McGowan volume tells you to stop, not tune. Only family-held-out splits are honest.