genai-lab: A Cutting-Edge Lab Reconstructing Computational Biology with Generative AI

genai-lab is a systematic open-source project aimed at applying generative AI technologies such as VAE, diffusion models, and Transformer to core scenarios in computational biology, covering areas like single-cell analysis, gene expression prediction, and drug perturbation response modeling. Positioned as an end-to-end research and application platform, it focuses on key applications like Perturb-seq perturbation prediction to facilitate drug discovery and life science research.

Generative AI is reshaping life sciences (e.g., AlphaFold solved protein structure prediction). genai-lab is not limited to reproducing a single model; its goal is to build a complete chain covering theoretical derivation, model implementation, and practical biological problems. Its flagship application is Perturb-seq perturbation prediction, which can simulate cell expression changes under drug or gene intervention, significantly reducing experimental costs.

1. VAE family: CVAE_NB (modeling gene expression count characteristics with negative binomial distribution), CVAE_ZINB (zero-inflation handling for scRNA-seq zero values), conditional design (injecting drug/cell type information);
2. Diffusion models: DDPM, Latent Diffusion (latent space diffusion reduces computational cost), DiT (Transformer replaces U-Net), Score/Flow Matching;
3. Foundation model adaptation: LoRA fine-tuning, Adapter insertion, hierarchical freezing strategy for adapting pre-trained models.

Industry Benchmarking: Benchmarked against platforms like Synthesize Bio (gene expression synthesis), Arc Institute (DNA sequence modeling), and Geneformer (single-cell analysis), providing a migration path from academia to industry; Documentation System: Includes mathematical derivations (VAE ELBO, diffusion process), architecture design (DiT/JEPA adaptation), application guides (Perturb-seq tutorials), and dataset descriptions, serving both as a codebase and a learning resource.

Completed: Theoretical documentation system, core model implementation, standardized data preprocessing; In Progress: Improving Perturb-seq application, public method benchmark comparison; Planned: Integrating causal inference (collaboration with causal-bio-lab), hybrid predictive generative models, biology-aware synthetic data pipelines.

Practical Significance: Provides researchers with a systematic tech stack, production-grade code, cutting-edge method tracking, and open collaboration; accelerates PoC transformation for drug discovery in industry; Summary: genai-lab embodies the "domain knowledge-driven" AI for Science paradigm, deeply adapts to biological data characteristics, and is a noteworthy open-source project in the intersection of AI and computational biology.