0.0.3

What's Changed

• Add Docker Image by @colbyford in #4.
• Add revisions for version 0.0.3 of FlowDock by @amorehead in #9.

Revisions for version 0.0.3:

1. Added CASP16 structure prediction benchmarking results and analysis notebook, which include new single and multi-ligand protein complexes!
2. Identified a version 0.0.2 bug in the PoseBusters pip package that allowed predictions for only 14 of DockGen-E's 122 protein-ligand complexes to be scored. Addressed by accordingly renormalizing each method's average number of correct DockGen-E predictions across all three of its repeat runs.
3. Updated FlowDock's overview figure to better illustrate the network's information flow (no pun intended).
4. Added results and analysis notebook for a PoseBusters Benchmark subset case study investigating how well FlowDock can predict PoseBusters Benchmark complexes that contain multiple distinct (Tanimoto similarity < 0.6) ligands binding to the same protein chain.
5. Added support for fine-tuning FlowDock's pretrained weights using new datasets such as PLINDER! Preliminary results suggest that (minimally) doing so with PLINDER increases FlowDock's average number of correct DockGen-E predictions by a few percentage points. Neat!
6. Added results for FlowDock when providing it with Chai-1's predicted (holo-like, ligand-bound) protein structures for the PoseBusters Benchmark dataset. Surprisingly, FlowDock's performance in this setting is nearly identical to its performance when providing it with ESMFold's (apo-like, unbound) predicted protein structures.
7. Corresponding updates to the arXiv preprint should be live by 03/27/2025.

New Contributors

• @colbyford made their first contribution in #4.

Full Changelog: 0.0.2...0.0.3