Boltz-2 and AEV-PLIG tested on new 15-PGDH lead optimization data
- Two recent ML affinity predictors, Boltz-2 and AEV-PLIG, were assessed in a prospective-style test on newly disclosed data from four related 15-PGDH ligand series, judged by whether they could guide synthetic prioritization.
- AEV-PLIG ranked compounds poorly in most series, and supplementing its training with literature 15-PGDH inhibitors brought no improvement.
- Boltz-2 cofolded poses and ranked compounds well for series resembling published inhibitors in its training data.
- For a series probing a new pocket region, Boltz-2 posed some ligands less accurately and underestimated potency gains, showing limits where guidance is needed most.
💡 Points
Boltz-2 doing well on chemistry that resembles published inhibitors is the least informative result, since a team can already guess those trends. The telling case is the new-pocket series, where it misposed ligands and undersold potency gains. I would use it to sanity-check familiar series, not to decide which risky analog gets made. And is AEV-PLIG, unhelped by literature data, worth another look?