Fig. 4: Comparative analysis of 3CLpro from human-infecting betacoronaviruses bound to NEMO.

a–d Substrate-binding site of human-infecting betacoronaviruses with hNEMO_227–234. Contacts that persist for more than 70% of the simulation time are labeled in bold. e Ranking of predicted hNEMO_227–234-binding affinities to 3CLpro from betacoronaviruses computed with quantum mechanics (QM) and molecular dynamics/machine learning (MD/ML) approaches. SARS-CoV-2, SARS-CoV, HCoV-HKU1, and MERS-CoV are abbreviated as SARS2, SARS1, HKU1, and MERS, respectively. In the MD/ML approach, five machine learning methods were used to train the model, namely, support-vector machine (SVM), gradient-boosted trees (BT; scaled and unscaled*), and random forest (RF; scaled and unscaled*). The ranking displayed was consistent for nine of the five cases using MD conformers (five ML models applied to either all MD conformers or the three lowest-energy MD conformers). The exception was a boosted tree model trained on unnormalized features that yielded a ranking of SARS-CoV < MERS-CoV < HKU1-CoV < SARS-CoV-2 when considering just the conformers with the lowest energy of interaction with 3CLpro computed from MD simulations. *Unscaled refers to the fact that unscaled, or unnormalized, features were used in the training.