Thanks for IARA — it is a genuinely useful pre-design signal. I have been using it to triage epitopes for de-novo binder campaigns, and wanted to share one licensing ask and two feature ideas grounded in that use.
1. LICENSE
There is currently no LICENSE file, which makes IARA hard to adopt or build on in downstream tooling (and to contribute back to, since contributions have no terms). Adding one (MIT / Apache-2.0 / BSD-3, etc.) would unblock a lot of reuse. Happy to help if useful.
2. Explicit aromatic feature
From predict.py, the per-residue node features are hydrophobicity, charge, a density-based pseudo-SASA + relative pseudo-SASA, Cα-density at 8/15 Å, and a local charge patch (a Cα-only GCN). Aromatics (F/W/Y) therefore only enter implicitly, through the hydrophobicity term.
In our campaigns, exposed aromatic anchors were one of the clearest discriminators of de-novo success — targets with an exposed aromatic loop were designable, while polar/anchorless surfaces of the same fold family failed de-novo. An explicit aromatic (or π-stacking) feature might sharpen the hotspot calls. This would need a retrain, so it is a bigger ask than #3.
3. Glycosylation awareness (no retrain)
IARA scores the bare, deglycosylated backbone, so it can flag a patch that is grippable in the model but carries an N-glycan in vivo. We hit exactly this: on Tenascin-C domain A1, IARA’s top-scoring internal patch overlaps N1119, a UniProt-annotated N-glycosylation site — so the "most designable" patch is actually glycan-occluded and unusable for a small binder.
A lightweight option would catch it without retraining: after smooth_predictions, flag or down-weight residues in N-X-[S/T] sequons (X≠P), optionally accepting a user-supplied glyco-site list. "Grippable but glycosylated" is a real failure mode a target-only model cannot otherwise see.
I would be glad to contribute a PR for #3 (and #2 if you are up for a retrain), especially once a license is in place. Thanks again for the tool.
Thanks for IARA — it is a genuinely useful pre-design signal. I have been using it to triage epitopes for de-novo binder campaigns, and wanted to share one licensing ask and two feature ideas grounded in that use.
1. LICENSE
There is currently no LICENSE file, which makes IARA hard to adopt or build on in downstream tooling (and to contribute back to, since contributions have no terms). Adding one (MIT / Apache-2.0 / BSD-3, etc.) would unblock a lot of reuse. Happy to help if useful.
2. Explicit aromatic feature
From
predict.py, the per-residue node features are hydrophobicity, charge, a density-based pseudo-SASA + relative pseudo-SASA, Cα-density at 8/15 Å, and a local charge patch (a Cα-only GCN). Aromatics (F/W/Y) therefore only enter implicitly, through the hydrophobicity term.In our campaigns, exposed aromatic anchors were one of the clearest discriminators of de-novo success — targets with an exposed aromatic loop were designable, while polar/anchorless surfaces of the same fold family failed de-novo. An explicit aromatic (or π-stacking) feature might sharpen the hotspot calls. This would need a retrain, so it is a bigger ask than #3.
3. Glycosylation awareness (no retrain)
IARA scores the bare, deglycosylated backbone, so it can flag a patch that is grippable in the model but carries an N-glycan in vivo. We hit exactly this: on Tenascin-C domain A1, IARA’s top-scoring internal patch overlaps N1119, a UniProt-annotated N-glycosylation site — so the "most designable" patch is actually glycan-occluded and unusable for a small binder.
A lightweight option would catch it without retraining: after
smooth_predictions, flag or down-weight residues in N-X-[S/T] sequons (X≠P), optionally accepting a user-supplied glyco-site list. "Grippable but glycosylated" is a real failure mode a target-only model cannot otherwise see.I would be glad to contribute a PR for #3 (and #2 if you are up for a retrain), especially once a license is in place. Thanks again for the tool.