Abstract

Synthetic single-wall carbon nanotubes (SWCNTs) contain various chiralities, which can be sorted by DNA. However, finding DNA sequences for this purpose mainly relies on trial-and-error methods. Predicting the right DNA sequences to sort SWCNTs remains a substantial challenge. Moreover, it is even more daunting to predict sequences for sorting SWCNTs with target chirality. Here, we present a deep-learning (DL) enhanced strategy for the accurate prediction of DNA sequences capable of sorting target-chirality nanotubes. We first experimentally screened 216 DNA sequences using aqueous two-phase (ATP) separation, resulting in 116 resolving sequences that can purify 17 distinct single-chirality SWCNTs. These experimental results created a comprehensive training data set. We utilized the recently released 3D molecular representation learning framework, Uni-Mol, to construct a DL workflow that maps atomistic-level structural information on DNA sequences into the feature space. This information captures the structural features of DNA molecules that are crucial for their interactions with SWCNTs. This may account for the superior performance of our DL models. The models successfully predicted resolving sequences for (6,5), (6,6), and (7,4) SWCNTs with accuracy rates of 87.5, 90, and 70%, respectively. Importantly, the discovery of numerous resolving sequences for (6,5) SWCNTs allows us to systematically manipulate the sequence-dependent absorption spectral shift, photoluminescence intensity, and surfactant sensitivity of DNA-(6,5) hybrids and elucidate the underlying mechanisms.