Co-Diffusion: An Affinity-Aware Two-Stage Latent Diffusion Framework for Generalizable Drug-Target Affinity Prediction

2026-03-13 · Source: stat.ML updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computational Drug Discovery · Depth: Expert, extended

Summary

Co-Diffusion is a novel affinity-aware latent diffusion framework designed to improve drug-target affinity (DTA) prediction, especially in challenging cold-start scenarios where data is scarce or exhibits domain shifts. The framework redefines DTA prediction as a constrained latent denoising process, employing a two-stage training paradigm. Stage I establishes an affinity-steered latent manifold by aligning drug and target embeddings using a supervised objective. Stage II then introduces modality-specific latent diffusion as a stochastic perturb-and-denoise regularizer, forcing the model to recover consistent affinity semantics from noisy structural representations. This approach mitigates the common "reconstruction-regression conflict" found in other generative DTA models. Co-Diffusion significantly outperforms state-of-the-art baselines on the KIBA and Davis datasets, demonstrating superior zero-shot generalization on unseen molecular scaffolds and novel protein families, with an 18.5% MSE improvement over PAIR-VAE in out-of-sample PDBbind evaluation.

Key takeaway

Research Scientists developing DTA models should consider adopting Co-Diffusion's two-stage, affinity-aware latent diffusion approach. This framework offers robust generalization in cold-start scenarios by effectively decoupling affinity alignment from generative refinement, thereby overcoming the reconstruction-regression conflict common in other generative models. Implementing this method could lead to more accurate and reliable predictions for novel drug-target pairs, accelerating early-stage drug discovery and virtual screening.

Key insights

Co-Diffusion enhances drug-target affinity prediction in cold-start scenarios using a two-stage, affinity-aware latent diffusion framework.

Principles

• Decouple affinity alignment from generative refinement.
• Latent diffusion regularizes against distribution shifts.
• Maximize variational lower bound for probabilistic coherence.

Method

Co-Diffusion uses a two-stage process: Stage I aligns drug and target embeddings to an affinity-steered latent manifold; Stage II applies modality-specific latent diffusion as a stochastic perturb-and-denoise regularizer to refine affinity semantics.

In practice

• Use GatedConv extractors for sequence feature extraction.
• Employ a 1-D UNet for latent space denoising.
• Optimize diffusion steps (T=1000) and noise magnitude (beta in [1e-4, 4e-4]).

Topics

• Drug-Target Affinity Prediction
• Latent Diffusion Models
• Cold-Start Generalization
• Computational Drug Discovery
• Generative Models

Best for: Research Scientist, AI Researcher, AI Scientist, AI Data Scientist

Related on AIssential

• Treatment-Conditioned Diffusion for Forecasting Neurodegenerative Disease Progression — A treatment-conditioned diffusion framework accurately forecasts neurodegenerative disease progression by integrating imaging and pharmacological data, preserving anatomical detail.
• A Triple-Modal Contrastive Learning Framework with Sequence, Graph, and 3D Features for Drug-Target Interaction Prediction — TriMod-DTI uses triple-modal contrastive learning with 1D, 2D, and 3D features for superior drug-target interaction prediction.
• 🔬 The Coolest Diffusion Research Isn't in LLMs — Evan Feinberg & Sergey Edunov, Genesis Molecular AI — Diffusion models, enhanced with physical priors and synthetic data, are achieving critical resolution in 3D protein-ligand prediction, transforming drug discovery.
• MicroscopyMatching: Towards a Ready-to-use Framework for Microscopy Image Analysis in Diverse Conditions — MicroscopyMatching unifies diverse image analysis tasks into a matching problem using latent diffusion models.
• Finetuning-Free Diffusion Model with Adaptive Constraint Guidance for Inorganic Crystal Structure Generation — Adaptive constraint guidance in diffusion models enables finetuning-free generation of inorganic crystal structures with targeted properties.
• Two-dimensional geometric template diffusion for boosting single-sequence protein structure prediction — TDFold uses 2D geometric template diffusion for efficient, accurate single-sequence protein structure prediction.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by stat.ML updates on arXiv.org.