Prasinezumab: A Bayesian Perspective on Its Efficacy

Abstract

This study employed a Bayesian approach to examine the impact of prasinezumab on the progression of PD symptoms and signs. We used the BF in hypothesis testing. The BF is inherently comparative: it weighs the support for one model against that of another. Moreover, BFs do so by fully conditioning on the observed data. Otherwise, the P value depends on hypothetical outcomes that are more extreme than those observed in the sample. Such practice violates the likelihood principle and results in inconsistent or paradoxical conclusions. The BF can quantify evidence in favor of the null hypothesis. In the Bayesian framework, no special status is attached to either of the hypotheses under test; the BF assesses each model's predictive performance and expresses a preference for the model that made the most accurate forecasts. The fact that the BF can quantify evidence in favor of the null hypothesis can be of substantive importance. For instance, the hypothesis of interest may predict the absence of an effect across a varying set of conditions. Quantifying the null hypothesis is also important to learn whether the observed data provide evidence of absence or absence.

Specifically, the possible outcomes of the BF can be assigned to three discrete categories: (1) evidence in favor of $H_1$ (ie, evidence in favor of the presence of an effect), (2) evidence in favor of $H_0$ (ie, evidence in favor of the absence of an effect), and (3) evidence that favors neither $H_1$ nor $H_0$. Instead, the P value cannot provide a measure of evidence in favor of the null hypothesis. Finally, the BF is not affected by the sampling plan, that is, the intention with which the data were collected. This irrelevance follows from the likelihood principle, and it means that BFs may be computed and interpreted even when the intention with which the data are collected is ambiguous, unknown, or absent. All these advantages are not available if a classical analysis is performed as was done for the PASADENA trial data.

Based on the findings shown in the first table of the source article²¹ (Table 1), a Bayesian analysis of the results obtained in these subpopulations was carried out. The results of the Bayesian analysis are shown in Table 2. The posterior probability column on the drug's effectiveness indicates no major and clinically relevant difference between the placebo and treated groups. The lack of efficacy applies to the population with and without rapid disease progression. The probability of efficacy is consistently less than 50% and near 50% only for the data-driven subphenotype diffuse malignant subgroup. The results support the notion that the drug's effects cannot be safely assessed.

The findings are robust, as depicted in Figure 2, which shows no variation between the two hypotheses as the a priori changes; instead, there is only an increase in the BF for the same hypothesis. Additional analyses are described in Supporting Information Data S1 (Additional Bayesian Analyses section) and support the main results. Thus, there is no substantial evidence of a difference in efficacy between subpopulations with or without rapid progression. Ultimately, these results are consistent with the phase two PASADENA trial, which did not reach the primary end point at 52 weeks. In conclusion, the exploratory analysis to assess whether prasinezumab generates greater benefits on motor progression in prespecified subgroups with faster motor progression using the BF resulted in no supporting evidence.

The negative findings related to the anti–α-synuclein approach confirm the many still unknown physiological and pathophysiological mechanisms controlled by the protein.

Incorporation of the Bayesian approach in analyzing the efficacy of prasinezumab for the treatment of PD offers a number of important strategies for ongoing and future clinical trials, like adaptive designs to allow real-time decision-making and optimizing the trial parameters.²³ The Bayesian framework also provides a probabilistic approach that quantifies uncertainty in treatment effects, thereby better informing decisions by stakeholders about modifications in trials, such as early stopping and expansion or changes in allocation ratios.²⁴ Furthermore, Bayesian hierarchical models enable the assessment of individual variability in treatment responses, providing the groundwork for personalized treatment strategies.²⁵

Beyond statistical and methodological issues, an additional point that should be considered when discussing the unsuccessful attempts to reach DMTs pertains to the disconnect between therapeutic strategies that are envisioned and tested in “clean and sanitized” clinical trials and the more difficult challenges offered by real-world settings. As recently discussed by Brett K. Beaulieu-Jones et al,²⁶ research populations are studied among actively recruited individuals who often receive earlier diagnoses and comply with more consistent follow-ups. In contrast, in real-world populations, patients are diagnosed later in life and often exhibit a more rapid progression because of a combination of selection bias, multiple-hit comorbidity, late access to care, and intrinsic population differences.²⁶ The study also highlighted bias in data collection. The mode of data gathering (actively recruited vs. passively recorded) introduces biases that must be carefully considered in clinical trial design and real-world analyses, ultimately affecting the validity of clinical outcomes.

Somehow, the PASADENA trial mirrors findings from the Alzheimer's field in which multiple clinical trials targeting a single protein, amyloid, generated modest or null effects, substantially failing as disease-modifying intervention.²⁷ In that regard, novel insights are promoting a reconceptualization of PD itself as more than a mere “synucleinopathy”, but rather a heterogeneous disorder arising from the convergence of multiple pathological processes but ultimately giving a variety of clinical manifestations. Most likely, “there is more than one Parkinson's Disease,”²⁸ and our collective efforts should focus on dissecting the convergent and divergent mechanisms that act inside or outside the central nervous system.

Research project: A. Conception, B. Organization, C. Execution; Statistical analysis: A. Design, B. Execution, C. Review and critique; Manuscript: A. Writing of the draft, B. Review and critique.

MR: 1A, 1B, 1C, 2C, 3A, 3B.

TC: 1A, 1C, 2A, 2B, 2C, 3A, 3B.

DC: 1A, 1C, 2C, 3A, 3B.

SLS: 1A, 1B, 1C, 2C, 3A, 3B.

S.L.S. is supported by research funding from the Italian Department of Health (RF-2013–02358785 and NET-2011-02346784-1), from the AIRAlzh Onlus (ANCC-COOP), from the Alzheimer’s Association—Part the Cloud: Translational Research Funding for Alzheimer’s Disease (18PTC-19-602325) and the Alzheimer’s Association—GAAIN Exploration to Evaluate Novel Alzheimer’s Queries (GEENA-Q-19-596282).