1/6 Can #AlphaFold learn from + & - labeled data to predict peptide binding better? w/ Phil Bradley @JustasDauparas @minkbaek @Mohamad_Abedi David Baker we present joint structure prediction-classification fine-tuning of #AF2 to classify pMHC binding.@UWproteindesign @fredhutch 
2/6 We developed a template-based method for accurately predicting pMHC structures & fed confidence scores to a simple classifier to fine-tune with structure+classification loss on a self-distillation set of + and - pMHC interactions & reach SOTA in Class I & II classification.
3/6 We show that our method is potentially orthogonal to NetMHCpan and can correctly label peptides where NetMHCpan fails.
4/6 We also extend our template-based modeling approach to SH3 and PDZ domains. We construct accurate specificity sequence logos in silico and show that the fine-tuned model on pMHC performs better than #AF2 on this task.
5/6 All of our code for combined structure prediction-classification #AF2 fine-tuning (in JAX) is available at github.com/phbradley/alph… (currently minimal version; updates coming soon). We believe this method is powerful for systems where less experimental data is available.
6/6 Find the preprint here: biorxiv.org/content/10.110…
• • •
Missing some Tweet in this thread? You can try to
force a refresh
