Lewis Chinery

Happy to share that Humatch, @OPIGlets new humanisation tool, is now available at mAbs - doi.org/10.1080/194208… Please also check out the code on GitHub (web server coming soon!) - github.com/oxpig/Humatch Highlights🧵in previous tweet - x.com/LewisChinery/s…

~60% of therapeutic antibodies are not genetically human in origin so there is great demand for trusted humanisation tools. Please give Humatch a try and let us know what you think!

…while also considering the fact that immunogenic epitopes could be formed between heavy and light chains.

Prefer your antibodies to be human? Then please check out Humatch! Preprint - doi.org/10.1101/2024.0… Code - github.com/oxpig/Humatch Data - zenodo.org/records/137647… @OPIGlets

@JulianFaus13563 We didn't try the EGNN classifier with ABB structures as this would have required redocking with HER2 (as you mentioned) which would have been v. computationally expensive for 500k complexes & introduced more uncertainty. I would also be curious if others have tried this though

Happy to share our recent work on antibody library design, affinity prediction, and optimisation - 'Baselining the Buzz. Trastuzumab-HER2 affinity, and beyond!' Preprint (inc. SI) - doi.org/10.1101/2024.0… Code - github.com/oxpig/Tz_her2_… Data - doi.org/10.5281/zenodo… 🧵

@Ella_Maru Thanks, Ella! Though I'm a fan of the ML, for many research questions we're lacking enough experimental data for accurate/generalisable ML predictions. Wet lab experiments are therefore still absolutely essential for training and validating the latest AI models

@LewisChinery This is the innovative use of standard computational methods! How does the balance between computational prediction and experimental validation shift as we move towards more sophisticated AI models?

@JulianFaus13563 I'm certainly a believer in structure-based design too (10.1093/bioinformatics/btac732)! In this paper though we aim to show that for certain "simple" applications (single target) with limited data availability, less complex classifiers can perform better

@JulianFaus13563 FoldX is only used for the EGNN classifier, though I can definitely add an extra line to the SI to make this clearer once the experimental validation results come in!

@JulianFaus13563 Hi Julian, for ProteinMPNN we use ABodyBuilder2 to model each SI positive variant, similar to the main text - that's why ProteinMPNN designs different sequences here, affecting the final predicted enrichments

Image credits: Mr Potato Head - deviantart.com/zacktv321 Lotso - pixar.fandom.com/wiki/Pixar_Wiki Original screenshots from Toy Story 1 and 3!

[1] 10.1073/pnas.89.22.10915 [2] 10.1093/bioadv/vbac046 [3] 10.1126/science.ade2574 [4] 10.1126/science.add2187 [5] 10.1002/pro.4205