DREAM-B3P: Dual-Stream Transformer Network Enhanced by Feedback Diffusion Model for Blood-Brain Barrier Penetrating Peptide Prediction

Introduction: The blood-brain barrier (BBB) protects the central nervous system but prevents most neurotherapeutics from reaching effective concentrations in the brain. BBB-penetrating peptides (BBBPs) offer a promising strategy for brain drug delivery; however, the scarcity of positive samples and severe class imbalance hinder the reliable identification of BBBPs. Objectives: Our goal is to alleviate class imbalance in BBBP prediction and to develop an accurate, interpretable classifier for BBBP prediction. Methods: We propose DREAM-B3P, which couples a feedback diffusion model (FB-Diffusion) for data augmentation with a dual-stream Transformer for classification. FB-Diffusion learns the BBBP distribution via iterative denoising and uses an external analyzer to provide feedback, generating high-quality pseudo-BBBPs. The classifier contains a sequence stream that extracts structural features from peptide sequences and a physicochemical stream that captures physicochemical features such as hydrophobic surface area, molecular charge, number of rotatable bonds, and polarizability. Combining the two features leads to superior BBBP predictive performance. Results: On a benchmark test set containing equal numbers of BBBPs and non-BBBPs, DREAM-B3P surpasses baseline methods (Deep-B3P, B3Pred, BBPpredict and Augur), improving AUC/ACC/MCC by 4.3%/17.8%/14.9%, respectively, over the second-best method. Conclusion: By integrating feedback diffusion with a dual-stream Transformer classifier, DREAM-B3P effectively mitigates data scarcity and imbalance and achieves state-of-the-art performance.

💡 Research Summary

The blood‑brain barrier (BBB) is a critical protective interface that severely limits the delivery of therapeutic agents to the central nervous system. BBB‑penetrating peptides (BBBPs) have emerged as promising vectors for brain drug delivery, yet the scarcity of experimentally validated BBBPs creates a pronounced class‑imbalance problem that hampers the development of reliable predictive models. Existing approaches—ranging from traditional machine‑learning classifiers based on amino‑acid composition and physicochemical descriptors to deep learning architectures such as CNNs, RNNs, and Transformers—have made progress but still suffer from two major limitations: (1) the synthetic positive samples generated by generative models often lack diversity and quality, and (2) many deep models focus exclusively on either sequence information or physicochemical properties, missing the synergistic information contained in both.
To address these challenges, the authors propose DREAM‑B3P, a framework that couples a feedback‑guided diffusion model (FB‑Diffusion) for data augmentation with a dual‑stream Transformer classifier. FB‑Diffusion builds on the recent diffusion‑based generative paradigm, which learns data distributions by iteratively adding Gaussian noise (forward process) and then denoising (reverse process). Because the BBBP dataset is extremely small (428 positive sequences), the diffusion model is prone to under‑fitting. The authors therefore integrate an external BLAST analyzer that scores each generated peptide for similarity to known BBBPs. Only sequences exceeding a predefined similarity threshold are retained as pseudo‑BBBPs and added back into the training pool. This feedback loop is repeated throughout 1,000 training epochs, with 6,000 pseudo‑BBBPs generated every 200 epochs, progressively refining the generator’s output distribution. Empirical results show that pseudo‑BBBPs from FB‑Diffusion are of higher quality than those produced by a previously used FB‑GAN, leading to consistent improvements in downstream classification performance.
The classifier itself consists of two parallel streams. The sequence stream receives 50‑amino‑acid peptides encoded as one‑hot vectors and processes them through a multi‑head self‑attention Transformer, capturing residue‑level motifs and long‑range dependencies. The physicochemical stream encodes four quantitative descriptors—hydrophobic surface area, net molecular charge, number of rotatable bonds, and polarizability—through a simple fully‑connected network. The embeddings from both streams are concatenated and fed to a final prediction head, enabling the model to exploit complementary information from both modalities.
Benchmark experiments were conducted on a balanced test set (50 BBBPs vs. 50 non‑BBBPs) while keeping the training set identical across all methods (378 BBBPs + 6,900 non‑BBBPs). DREAM‑B3P achieved an AUC of 0.951, sensitivity of 0.912, specificity of 0.858, accuracy of 0.886, and MCC of 0.773, outperforming the next best method by 4.3 % (AUC), 17.8 % (ACC), and 14.9 % (MCC). Ablation studies demonstrated that using either the sequence‑only or physicochemical‑only stream alone yields lower performance, confirming the benefit of the dual‑stream architecture. Moreover, as FB‑Diffusion training progressed, the classifier’s AUC and ACC steadily improved, plateauing around epoch 800, indicating that the quality of generated samples directly correlates with downstream gains.
In summary, DREAM‑B3P effectively mitigates data scarcity and class imbalance by generating high‑quality synthetic BBBPs via a feedback‑controlled diffusion process, and it leverages a novel dual‑stream Transformer to integrate sequence and physicochemical cues. The framework not only sets a new state‑of‑the‑art for BBBP prediction but also offers a generalizable strategy for other peptide‑function prediction tasks under limited‑sample, imbalanced‑data conditions. Future work may incorporate additional feedback sources (e.g., structural modeling) and expand the set of physicochemical descriptors to further boost generalization.