Machine Learning-Guided Discovery of 19F MRI Agents Enabled by Automated Copolymer Synthesis

Abstract

Modern polymer science is plagued by the curse of multidimensionality; the large chemical space imposed by including combinations of monomers into a statistical copolymer overwhelms polymer synthesis and characterization technology and limits the ability to systematically study structure-property relationships. To tackle this challenge in the context of 19 F MRI agents, we pursued a computer-guided materials discovery approach that combines synergistic innovations in automated flow synthesis and machine learning (ML) method development. A software controlled, continuous polymer synthesis platform was developed to enable iterative experimental-computational cycles that resulted in the synthesis of 397 unique copolymer compositions within a six-variable compositional space. The non-intuitive design criteria identified by ML, which was accomplished by exploring less than 0.9% of overall compositional space, upended conventional wisdom in the design of 19 F MRI agents and lead to the identification of >10 copolymer compositions that outperformed state-of-the-art materials.

Featurization strategies for polymer sequence or composition design by machine learning

“…43, dataset C from ref. 48, and dataset D from ref. 74); these datasets feature different property labels, design spaces, and CU metadata.…”

“…Dataset C is sourced from ref. 48, which uses a computerguided materials discovery approach to design statistical copolymers of methacrylates to serve as high contrast 19 F MRI agents. 48 There are six possible CUs that can be combined in varying proportions and degrees of polymerization, but the polymer sequences are unknown.…”

“…48, which uses a computerguided materials discovery approach to design statistical copolymers of methacrylates to serve as high contrast 19 F MRI agents. 48 There are six possible CUs that can be combined in varying proportions and degrees of polymerization, but the polymer sequences are unknown. Of the 397 unique copolymers reported in the study, we use 271 copolymers that were labeled with the signal-to-noise ratio (SNR) from NMR experiments; the SNR is always treated as the target output for our regression task.…”

“…In (B), the property label is an experimentally determined signal-tonoise ratio (SNR) as reported in ref. 48. Both panels examine the effect on MAE using one-hot encoding (purple), molecular fingerprints (green), and descriptor vector (orange) approaches to CU fingerprinting.…”

“…29,35,36 Presently, this challenge is overcome by either (i) laborious, brute-force data sourcing and curation or (ii) in-house data generation. The former approach has been largely useful for polymer informatics in the space of homopolymers, [37][38][39][40] while the latter has been typically necessary to design systems with sequence [41][42][43][44] or compositional control [45][46][47][48] over multiple monomers or constitutional units 49 (CUs). In the near term, advancements in automated polymer synthesis 50 and in hierarchical polymer simulation, 33 coupled with efficient data acquisition schemes, 34 are likely to substantially enhance capabilities to generate requisite data for training ML models on-the-fly.…”