Self-Driving\nSynthesis of Protein Nanoparticles by\nActive Transfer-Learning-Assisted Autonomous Flow Platform

The ability to precisely engineer protein nanoparticles\n(P-NPs)\nwith tailored size and low polydispersity index (PDI) unlocks their\npotential for a wide range of biomedical and material applications.\nHowever, conventional P-NP synthesis heavily relies on trial-and-error\napproaches, hindering efficient optimization. To address this challenge,\nwe introduce a novel self-driving synthesis platform that leverages\nthe power of active transfer learning and multitask Bayesian optimization\n(MTBO). This approach enables the intelligent transfer of knowledge\nfrom previous, even seemingly unrelated, synthesis tasks to accelerate\nthe optimization of new P-NPs. In this study, comprehensive in-silico\nsimulations demonstrate the significant acceleration achieved by pretrained\nMTBO models in optimizing the synthesis of four distinct P-NPs (BSA,\nZein, Gelatin, and Hemoglobin). We further validate these findings\nby developing an autonomous flow microfluidic platform for algorithm-driven\nsynthesis. Experimental case studies with BSA and Zein illustrate\nMTBO’s superiority over single-task BO in minimizing both size\nand PDI. Crucially, we introduce a mutual active learning platform\nwhere MTBO dynamically leverages data from the parallel synthesis\nof two P-NP systems. This strategy addresses the challenge of limited\npre-existing data sets, further expanding MTBO’s practical\napplications in data-scarce scenarios. Our work highlights the transformative\npotential of active transfer-learning concepts in accelerating the\nsynthesis and development of P-NPs and potentially a wide array of\nother advanced materials.