NanoNet: Rapid and accurate end-to-end nanobody modeling by deep learning

Abstract

Antibodies are a rapidly growing class of therapeutics. Recently, single domain camelid VHH antibodies, and their recognition nanobody domain (Nb) appeared as a cost-effective highly stable alternative to full-length antibodies. There is a growing need for high-throughput epitope mapping based on accurate structural modeling of the variable domains that share a common fold and differ in the Complementarity Determining Regions (CDRs). We develop a deep learning end-to-end model, NanoNet, that given a sequence directly produces the 3D coordinates of the backbone and Cβ atoms of the entire VH domain. For the Nb test set, NanoNet achieves 3.16Å average RMSD for the most variable CDR3 loops and 2.65Å, 1.73Å for the CDR1, CDR2 loops, respectively. The accuracy for antibody VH domains is even higher: 2.38Å RMSD for CDR3 and 0.89Å, 0.96Å for the CDR1, CDR2 loops, respectively. NanoNet run times allow generation of ∼1M nanobody structures in less than 4 hours on a standard CPU computer enabling high-throughput structure modeling. NanoNet is available at GitHub: https://github.com/dina-lab3D/NanoNet

tFold-Ab: Fast and Accurate Antibody Structure Prediction without Sequence Homologs

“…Evaluation. Baseline methods include antibody-specific structure prediction (ABodyBuilder [19], DeepAb [34], ABlooper [1], NanoNet [8] and IgFold [31]) and general protein structure prediction, either MSA-based (AlphaFold [15] and AlphaFold-Multimer [10]) or MSA-free (HelixFold-Single [11], ESMFold [20], and OmegaFold [40]).…”

“…DeepH3 [33] and its improved version, DeepAb [34], predict inter-residue geometric restraints with residual networks and then perform constrained energy minimization with Rosetta. NanoNet [8] firstly aligns all the training structures and then builds a 1D convolutional network to predict 3D coordinates of backbone and C β atoms. However, above methods only use neural networks to predict intermediate structures or restraints, and still rely on additional tools for full-atom structures, which limits the prediction accuracy.…”

Structural pre-training improves physical accuracy of antibody structure prediction using deep learning

“…As the basis architecture, we aimed for the simplest deep learning model that would not rely on improvements such as problem-specific loss functions or all-atom predictions. For this purpose, we re-created the architecture of NanoNet (Cohen, Halfon, and Schneidman-Duhovny 2022). The model is a small residual neural network (seven blocks, 1.9m parameters) predicting coordinates directly, in comparison to other networks using more advanced architectures and training regimes.…”

“…Model for pre-training. We employed the simplest deep learning architecture we could identify that reported heavy chain modeling of antibodies, which was NanoNet (Cohen, Halfon, and Schneidman-Duhovny 2022). In brief it consists of seven convolutional residual blocks.…”

“…To address this issue, but also to draw from AlphaFold2 and its derivatives (Ahdritz et al 2022; Lin et al 2022), many antibody-specific deep learning structure prediction methods have been introduced (Wilman et al 2022; Lin et al 2022). The algorithms started by addressing CDR-H3 loop prediction such as DeepH3 (Ruffolo et al 2020) and AbLooper (Abanades, Georges, et al 2022), after extending these to the entire variable domain via NanoNet (Cohen, Halfon, and Schneidman-Duhovny 2022) and the entire Fv molecule by DeepAb (Ruffolo, Sulam, and Gray 2022), IgFold (Ruffolo et al 2022), AbodyBuilder2 (Abanades, Wong, et al 2022), EquiFold (Lee et al 2022) and tFold-Ab (Wu et al 2022). As opposed to the homology methods that reported CDR-H3 root mean squared deviation (RMSD) accuracies in the region of 3-4Å (Almagro et al 2014), the deep learning methods achieve an RMSD of 2-3Å RMSD.…”