Target-aware latent diffusion model for design of apoptosis-inducing anticancer peptides.

Therapeutic peptides offer high specificity and low toxicity, but rational design remains difficult due to their vast chemical diversity and conformational flexibility. In oncology, B-cell lymphoma-extra large (BCL-xL), a key anti-apoptotic protein, is overexpressed in various cancers, making it an important target for peptide-based therapeutics. Existing generative models for anticancer peptide design often neglect detailed receptor structural context, limiting biological plausibility. We present a target-aware latent diffusion framework (T-LDM) that conditions sequence denoising on receptor pocket context encoded by a graph-based structural model. Conditioning is fused into the U-Net denoiser via lightweight modulation and cross attention, enabling the Latent Diffusion Model to align sampling with the local interaction topology of the binding groove while preserving protein language model priors. We evaluate distributional calibration with Jensen-Shannon divergence across key physicochemical properties, sequence quality with perplexity, motif overlap with BLEU, and structural compatibility via protein-peptide docking. A controlled ablation was performed that varies the scope of structural conditioning (none, global, pocket) and the fusion into the denoiser, letting us assess the marginal effect of pocket-aware guidance and fusion on generation quality. Overall, the study indicates that explicit pocket conditioning yields peptides that are syntactically fluent, physicochemically realistic, and more compatible with the intended BCL-xL target, while maintaining diversity and novelty. Source code is available at https://github.com/tiaranatashasayuti/T-LDM.