luca0621/polyedit-mot-retrained

PolyEdit Molecular Optimization Transformer (polymer-retrained)

This checkpoint adapts MolecularAI's released Molecular Optimization Transformer to two-anchor PSMILES and the eight PolyEdit DFT properties. It is not directly comparable to the authors' original LogD/Solubility/Clint model without this adaptation.

The original transformer body is retained: 6 encoder/decoder layers, hidden size 256, 8 heads, and feed-forward size 2,048. All 258 shape-compatible upstream tensors were loaded. Four vocabulary-dependent embedding/generator tensors were initialized for a new vocabulary in which padding and the * polymer anchor have distinct IDs. Training used 152,036 component-held-out examples; the best validation checkpoint was epoch 7 of 10.

On the balanced 8,176-request test set, the checkpoint obtains 99.217% RDKit+TDC validity, 99.168% two-anchor polymer validity, 99.083% changed outputs, 8.892% strict MIPS-retrained full-edit hits, and 10.127% strict DFT full-edit hits at 95.034% DFT coverage. Only 383 unique raw outputs were generated, so the strong validity comes with substantial mode collapse and must not be reported without the diversity result.

Upstream code: <https://github.com/MolecularAI/deep-molecular-optimization>

Upstream checkpoint: <https://doi.org/10.5281/zenodo.5571965>

PolyEdit implementation and record-level evaluation: <https://github.com/promotion-kim/POLYEDIT/tree/tsyou/balanced-polymer-baseline-eval>