Applying machine learning to explore correlates of protection for a vaccine against Trypanosoma cruzi

GAMBA J.C.; BORGNA E.; PROCHETTO E.; DÍAZ G.; PÉREZ A.R.; MARCIPAR I.; GERARD M.; CABRERA G.

San Luis

Congreso; LXXI Annual Meeting of the Argentine Society of Immunology (LXXI SAI 2023); 2023

Sociedad Argentina de Inmunología

Introduction: Chagas disease, caused by the protozoan parasite Trypanosoma cruzi (T. cruzi), is a tropical neglected disease for which a vaccine has yet to be developed. We previously described a vaccine candidate composed of a trans-sialidase fragment (TSf), and a cage-like particle adjuvant (ISPA). Correlates of protection (CoPs) are immunological biomarkers which can be used to predict the efficacy of a vaccine. Machine learning algorithms can be used as an important tool to establish CoPs. Objective: to apply machine learning to search for potential CoPs in the development of a vaccine against T. cruzi. Methods: BALB/c mice that received a protocol of vaccination with TSf-ISPA were included in the study. IgG antibodies anti-TSf were measured by ELISA. Delayed hypersensitivity reaction (DTH) was measured 48 h post-inoculation of 5 ug TSf in the footpad of TSf-ISPA treated and PBS-control inoculated mice. Vaccinated and control mice were challenged intraperitoneally with 1000-2000 T. cruzi. (n=20 per study). Parasitemia were measured at day 15 post-infection (p.i.) and survival was recorded until day 40 p.i. Python’s scikit-learn library was used to construct machine learning classification models. Results: Logistic regression models were generated to assess the use of optical density (OD 450 nm) as a CoP for the survival of vaccinated mice challenged with 1000 or 2000 T. cruzi. In all cases, the results were obtained using k-fold methodology. In order to consider not only the death-live difference but also the time of survival, a criterion was developed. A classification label of “1” was assigned to each mouse that died before day 21 p.i. (coincident with the peak of parasitemia), and a label of “0” was assigned to each mouse which died after day 21 p.i. or even did not die. This approach yielded a substantial predictive capability. For instance, confusion matrix analysis returned average values of sensibility of 80,5%, specificity of 86%, and an area under ROC curve of 0,84 for data from mice infected with 1000-2000 parasites. Regarding DTH as a CoP, when logistic regression was used to analyze data after the challenge with 1000-2000 T. cruzi, confusion matrix returned average values of 83% sensitivity, 90,5% specificity, and a value of 0,87 for the area under the ROC curve. For parasitemia studies, to develop the models, the following criterion was used: if parasitemia value of each mice/mean PBS parasitemia>1, a classification label of “1” was assigned to that mouse. If parasitemia value of each mouse/mean PBS parasitemia