Nuclear magnetic resonance (NMR) spectroscopy is a powerful technique for characterising the complex chemistry of soil organic carbon (SOC), but is prohibitively expensive, time-consuming and technically-demanding. Diffuse reflectance mid-infrared (MIR) spectroscopy is an attractive alternative because it is a high-throughput, cost-effective and easy-to-use technique that provides information on the amount and nature of soil mineral and organic components. However, interpretation of complex MIR spectra can be challenging due to difficulties with distinguishing SOC peaks from overlapping mineral-related peaks. We present a novel approach to predict the entire NMR spectra of SOC from corresponding MIR spectra using partial least-squares regression (PLSR) in an R environment. We developed a multi-response MIR–PLSR prediction model by regressing corresponding NMR and MIR spectra of 99 HF-treated <50μm fractions of soils using the pls package. The model was validated using (set-aside) test sets in four model iterations. The model provided accurate predictions of the entire average NMR spectra. Average Euclidean distance values between spectra in the training set were at least 3.5 fold greater than those between average reference and predicted NMR spectra, indicating that prediction errors were small relative to between-soil variation. Our approach accurately predicted intricate NMR spectra, demonstrating new potential for routine analysis of complex SOC chemistry. Display omitted •We present a novel application of the pls package in R to predict NMR spectra of SOC.•We develop multi-response calibrations using MIR spectra and partial least-squares.•Method validation reveals accurate predictions of intricate NMR spectra of SOC.•This approach offers enhanced capability for routine analysis of complex SOC chemistry.