Abstract Body

Background: In HIV-1 treatment, clinicians still have to face drug-resistant viral strains, while the amount of available drugs and drug targets remains limited. Recently, combination therapy with broadly neutralizing antibodies (bNAbs) was introduced as a potential HIV-1 treatment that is capable to reduce viral load under detectable levels for up to 60 days in humanized mice and primates. However, similarly to HAART, the emergence of resistant strains is a major problem when selecting an efficient combination therapy of bNAbs for an individual patient. Prior to the administration of a bNAb combination therapy to a patient, it has to be ensured that the patient’s viral strains are susceptible to the particular bNAb.

Methods: For seven different bNAbs we trained a classifier for each using support vector machines. The prediction models are able to determine the neutralization susceptibility of unseen viral strains to the specific bNAb based on the viral envelope sequence (Env). Different string kernels as well as the polynomial kernel and the Gaussian RBF kernel were tested by 10 times nested cross-validation. In addition, we introduce new visualization techniques (e.g., motif of most important residues) to increase model interpretability of non-linear classifiers.

Results: Among all considered kernels, the oligo string kernel performed best. High prediction performance could be traced back to learnt discriminant features that are supported by literature. State of the art performance of the classifiers can be seen in the table. For the V3-loop directed bNAbs the N-glycosylation site at position 332 was found as the most discriminant residue for neutralization susceptibility whereas position 334 confers neutralization resistance. The N-glycosylation site at position 160 was identified to be the most discriminant residue for the classifiers targeting the V1/V2 -loop indicating neutralization susceptibility. The classifiers for the CD4 binding site-directed bNAbs recognized different known binding sites to the CD4 molecule on the gp120 Env subunit to be associated with neutralization susceptibility of the viral strains. In addition to the already known epitopes, we also found other discriminant residues that might be interesting for follow-up structural studies.

Conclusions: The good classfier performances motivate their use in the selection of bNAb combination therapy. The robustness of the models implies that models with similar accuracy and interpretability can also be learnt for additional bNAbs.

viewer?TAG_ACTION=SECURE_FILE_DOWNLOAD&V
AUC (area under the curve) performance of each classifier assessed by 10 times 5-fold nested cross-validation.