A proteomic survival predictor for COVID-19 patients in intensive care

Abstract

Global healthcare systems are challenged by the COVID-19 pandemic. There is a need to optimize allocation of treatment and resources in intensive care, as clinically established risk assessments such as SOFA and APACHE II scores show only limited performance for predicting the survival of severely ill COVID-19 patients. Comprehensively capturing the host physiology, we speculated that proteomics in combination with new data-driven analysis strategies could produce a new generation of prognostic discriminators. We studied two independent cohorts of patients with severe COVID-19 who required intensive care and invasive mechanical ventilation. SOFA score, Charlson comorbidity index and APACHE II score were poor predictors of survival. Plasma proteomics instead identified 14 proteins that showed concentration trajectories different between survivors and non-survivors. A proteomic predictor trained on single samples obtained at the first time point at maximum treatment level (i.e. WHO grade 7) and weeks before the outcome, achieved accurate classification of survivors in an exploratory (AUROC 0.81) as well as in the independent validation cohort (AUROC of 1.0). The majority of proteins with high relevance in the prediction model belong to the coagulation system and complement cascade. Our study demonstrates that predictors derived from plasma protein levels have the potential to substantially outperform current prognostic markers in intensive care.

Parenclitic and Synolytic Networks Revisited

“…This property itself (apart from other advantages of parenclitic approaches) can be valuable in biological and medical problems utilising omics data, where there can be a huge number of features with comparatively few patients. This is exemplified in Demichev et al (2021) , where the use of wSA was more effective than other ML methods, most likely because the sample size was small relative to the large number of features. It was also seen that for wLRPA, the quality of discrimination (compared to other ML techniques) decreased once the number of samples highly exceeded the number of features, whereas there was no such effect for wKDEPA and wSA.…”

“…Moreover, as we were able to show with synthetic data and then confirm on real data, the quality of the wSA as a classifier is higher for those datasets where the sample size is small in comparison with the features size. Using this advantage we have recently applied this approach to the analysis of proteomics data from a large cohort of CoVID-19 patients, in which this is the case ( Demichev et al, 2021) (manuscript submitted). In this analysis, we showed that wSA was able to produce accurate classifications, where other ML algorithms were not on the same data.…”

“…To summarise our approach as an instruction for multi-disciplinary researchers: • For specialists in the field of medicine and biology, using the wSA approach ∗ As a classifier, in situations where the number of samples is small in comparison with the dimension of analytes [when there are few patients, but there are many measurements of their states, see, for example, ( Demichev et al, 2021) (manuscript submitted)]; ∗ As a high-quality and simple data visualization, when a visual representation of the state of the system features of an individual in the aggregate is required (we assume that such a representation in the form of networks can give a new understanding of the relationship of features, both among the entire set of subjects, and with an indication of some of their individual properties subjects, as shown in Figure 2A ); ∗ In situations where it is required to determine the intermediate state of the points during the transition, for example, from a healthy to severely ill state. As we have shown (through the radii on the artificial data, see Figure 3A ), parenclitic approaches reflect the spatial state on a one-dimensional scale; ∗ When it is required to interpret the transition between two states with respect to some kind of continuous effect (for example, in the work Krivonosov et al (2020) we showed how the third groups of samples according to the characteristics of networks demonstrate age tendencies between the features selected in binary networks between case and control groups); • For specialists in the field of machine learning and network approaches, we recommend using the wSA approach, as an interesting method to get new representations of the data.…”

“…∗ As a classifier, in situations where the number of samples is small in comparison with the dimension of analytes [when there are few patients, but there are many measurements of their states, see, for example, ( Demichev et al, 2021) (manuscript submitted)];…”

“…Moreover, we showed that the characteristics of the constructed networks help to interpret the obtained signatures in relation to aging in individuals from non-Downs Syndrome and Down Syndrome populations. A further development came from not binary networks, but weighted networks, and this method was successfully applied to prediction of survival for severely ill Covid-19 patients ( Demichev et al, 2021) , and for prediction of prostate cancer progression in patients on active surveillance ( Sushentsev et al, 2021) . We used SVM as the core, and the probability of belonging to a group of cases as the weights of the edges.…”