AbAdapt: an adaptive approach to predicting antibody–antigen complex structures from sequence

Abstract

Motivation The scoring of antibody-antigen docked poses starting from unbound homology models has not been systematically optimized for a large and diverse set of input sequences. Results To address this need, we have developed AbAdapt, a web server that accepts antibody and antigen sequences, models their 3D structures, predicts epitope and paratope, and then docks the modeled structures using two established docking engines (Piper and Hex). Each of the key steps has been optimized by developing and training new machine-learning models. The sequences from a diverse set of 622 antibody-antigen pairs with known structure were used as inputs for leave-one-out cross validation. The final set of cluster representatives included at least one “Adequate” pose for 550/622 (88.4%) of the queries. The median (IQR) ranks of these “Adequate” poses were 22 (5 to 77). Similar results were obtained on a holdout set of 100 unrelated antibody-antigen pairs. When epitopes were re-predicted using docking-derived features for specific antibodies, the median ROC AUC increased from 0.679 to 0.720 in cross validation and from 0.694 to 0.730 in the holdout set. Availability AbAdapt is available at https://sysimm.org/abadapt/. Supplementary information Supplementary data are available at Bioinformatics Advances online.

Improved antibody-specific epitope prediction using AlphaFold and AbAdapt

“…The preparation of data for LOOCV and holdout tests was described previously [4]. In brief, crystal structures of antibody-antigen complexes were gathered from the Protein Data Bank (downloaded May.…”

“…AbAdapt was run using default parameters, as described previously [4]. In brief, antibodies were modeled using Repertoire Builder [12] and antigens were modeled using Spanner [13].…”

“…The rank-one epitope prediction from EpiPred was chosen for all subsequent analysis. For AbAdapt-AF and AbAdapt, a threshold of 0.5 for final epitope prediction was used [4]. The performance indices of epitope prediction were calculated using the Python package Scikit-learn (v 0.23.2) [26].…”

“…We first analyzed the effect of AlphaFold2 models on the frequency of "true" poses produced by Hex or Piper in the LOOCV set. Here, as in our previous work [4], a relatively loose cutoff value of 15 Å for the RMSD of the interface residues (IRMSD), along with epitope and paratope accuracies of 50%, was used to define a "true" pose. The median true pose ratio of the Hex engine increased modestly from 1.50% to 1.53% (Supplementary Fig.…”

“…AbAdapt is a pipeline that combines antibody and antigen modeling with rigid docking and re-scoring in order to derive antibody-antigen specific features for epitope prediction [4]. As has been reported by others, the rigid docking and scoring steps are sensitive to the quality of the input models [5, 6].…”