Thread by @ideasbyjin

Exciting work from @KathyYWei1, @AmeyaHarmalkar & @proteinrosh on predicting scFv developability. TLDR: transformers and CNNs can potentially help prioritise mutations sites for enhancing stability (1/5) #antibodies #machinelearning #proteineng

https://twitter.com/KathyYWei1/status/1533235027110113280

Context: a single-chain Fv (scFv) is an antibody construct whose heavy and light chains are linked. It's not the conventional "Y" shape molecule, and is useful for engineering / phage display, etc. See @AlissaHummer's post blopig.com/blog/2021/07/a… (2/5)

Thermostability (measured by TS50, the temperature when scFv loses binding) is weakly predicted by 0-shot and fine-tuning via transformers (ESM-1v + ESM-1b). CNNs using sequence and structural (energy) convolutions perform better (?) [hard to tell, sorry!🙈] (3/5)

The crazy result in my mind is how both models did a good job at picking out the positions that enhance thermostability. The correct mutation isn't always predicted, but this can dramatically reduce the search space. Could be a game changer. (4/5)

Several Qs remain: which germlines were predicted better? Were they the ones naturally stable to begin with? Where are the ROC curves for the transformer models? Will the models perform better for IgGs vs scFv? Will CNNs increasingly replace LMs? (5/5)