<p>No-Reference Point Cloud Quality Assessment (NR-PCQA), which automatically evaluates the perceptual quality of distorted point clouds without a reference, is crucial for optimizing point cloud applications. However, it is hampered by data complexity and subjectivity in quality perception. Current deep learning methods typically regress quality to a single score, failing to capture the distribution of subjective ratings. To address this limitation, we propose <b>DPR-Net (Dual-branch Probabilistic Regression Network)</b>, which explicitly models subjective uncertainty in NR-PCQA. DPR-Net features two parallel pathways: a regression branch predicts the Mean Opinion Score (MOS), while a probability branch represents quality as a distribution over adaptively determined anchors using Gaussian soft assignment. We extract features from orthogonal projections using a Swin Transformer backbone, feeding them into both branches. Joint optimization with a combined loss function encourages learning features sensitive to both average quality and its distribution. Experiments on SJTU-PCQA and WPC datasets show our method significantly outperforms state-of-the-art NR-PCQA approaches, validating the effectiveness of our dual-branch probabilistic approach in handling subjective variance for point clouds.</p>