Implemented early stopping by specified score and threshold

[gezmi]

added flags to implement early stopping, as an example:
--early_stop_metric iptm --early_stop_threshold 0.85

Setting the metric will also change this metric to be the one used for ranking_score

In case someone wants to change the ranking score, without early stopping, one can use for example:
--early_stop_metric ptm --early_stop_threshold 1.0 => this will result in ptm being used for ranking amongst the models

[hdieckhaus]

@gezmi thanks for the PR! Looks like the CI tests are good, I just want to understand what this PR is doing before merge.

When running with multiple seeds, this stops after any seed that hits a certain score threshold? So then it won't run any further seeds for that sequence? I am curious the reasoning for doing this, rather than doing the early stopping after a certain number of recycles.

[gezmi]

Hi, Thanks for reviewing :) Yes, exactly, it breaks the for loop that goes over the seeds. As far as I understood, if I stop the recycling, the next seed would still be run. I am now running complexes sometimes up to 20 or even 100 seeds, but they can result in a "good enough" structure even with 1 seed. In these cases, I don't want to spend more resources on the run.

…

On Wed, 4 Feb 2026 at 00:29, Henry Dieckhaus ***@***.***> wrote: *hdieckhaus* left a comment (Kuhlman-Lab/alphafold3#32) <#32 (comment)> @gezmi <https://github.com/gezmi> thanks for the PR! Looks like the CI tests are good, I just want to understand what this PR is doing before merge. When running with multiple seeds, this stops after any seed that hits a certain score threshold? So then it won't run any further seeds for that sequence? I am curious the reasoning for doing this, rather than doing the early stopping after a certain number of recycles. — Reply to this email directly, view it on GitHub <#32 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ADHCIIYYZD7DIBSPD5AZOID4KEOKXAVCNFSM6AAAAACS6TJNLCVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZTQNBUGA2TQOJSGU> . You are receiving this because you were mentioned.Message ID: ***@***.***>

[hdieckhaus]

Sorry for the delay on this, merging now.