solver.press

← All replicated papers

Inconclusive under our protocolOct 27, 2026

Conditional GraphGANFed: Optimizing Graph-Structured Molecule Generation in Federated Generative Adversarial Networks

Manu and Alazzwi · arXiv 2608.24610

The outcome is the verdict on the one claim below, under the protocol we fixed in advance. It is not a verdict on the paper as a whole.

Claim tested

The conditional model (cGraphGANFed) generates molecules with about 10% higher drug-likeness (QED) than the GraphGANFed baseline.

What we ran and what came back

The released code, run for one round with three clients on a T4 GPU, with the small patches listed in the protocol. The pre-registered three seeds are Inconclusive: in two of them the conditional model's generator produced no valid molecules, so no interval could be computed. The baseline trained in all three. We then ran three more seeds under an amendment that was written after seeing that result, and is labelled as such. All three trained. Over the four usable seeds the conditional model's drug-likeness score (QED) was 0.8% above the baseline's (interval -3.4% to +5.2%), against the +10% the paper reports. Under the decision rule that pooled reading is Did not replicate, conditional on training success, and the two failed runs are reported beside it. In the released code a round trains one client's generator on the full data, so nothing here tests federation, and the paper's 100-round figures were not attempted.

What this result supports

The released code does not reproduce a +10% QED gain at one round on usable seeds.

What it does not support

Anything about the paper's 100-round numbers or a corrected federation loop.

Record

Protocol, deposited before the run: 10.5281/zenodo.23327505, 10.5281/zenodo.23333253

The authors were sent this result. They have not reviewed this listing. A first note went on 13 October 2026 and the result with a two-page report on 27 October 2026. No reply had been received when this was listed.

How the protocol works is on the method page. If you are an author of this paper and something here is wrong, write to customers@aegismind.app and we will correct it on this page.