Huayuan Zhiyin Secures Million-RMB Seed Round to 'Replicate' Human Cells and Predict Drug Efficacy with AI 用AI“复刻”人类细胞、预判药效，「华源智因」获千万级人民币种子轮融资｜36氪首发

TL;DR

• Huayuan Zhiyin secured tens of millions in RMB seed funding led by Shuimu Venture Capital to iterate multimodal sequencing technologies and expand hospital partnerships.
• The company focuses on AI Virtual Cells (AIVC) via its self-developed Wise-Perturb model, targeting human drug efficacy prediction to bridge the gap between animal models and clinical trials.
• Wise-Perturb utilizes a three-layer pyramid data foundation integrating DNA, RNA, and protein omics from patient cohorts, overcoming limitations of traditional single-RNA cell lines.
• The model features zero-shot cross-cell, cross-cancer, and cross-drug generalization capabilities, validated by high consistency with real-world efficacy in DS-8201 and Osimertinib cases.
• Early commercialization includes service fees from top-tier hospitals and pharma companies for preclinical evaluation and patient stratification, with plans for risk-sharing joint R&D models.

Why It Matters

This development addresses the critical bottleneck in drug discovery where high attrition rates stem from the disconnect between animal experiments and human physiology. By providing a computational method to predict human-specific drug responses and toxicity early in the pipeline, it offers a potential pathway to reduce the decade-long, billion-dollar cost of bringing new drugs to market. For the industry, it highlights a shift from purely generative AI in target identification to predictive simulation of complex biological systems using multi-omics data.

Technical Details

• Wise-Perturb Architecture: A specialized model for simulating cellular drug perturbations that integrates multi-modal omics data (DNA, RNA, protein) rather than relying solely on transcriptomic data.
• Three-Layer Data Pyramid: The model is trained on a hierarchical dataset consisting of: bottom layer (massive static single-cell sequencing datasets), middle layer (hundreds of millions of in vitro cell line/PDX drug-gene perturbation pairs), and top layer (scarce paired pre/post-treatment clinical tumor cohort data).
• Cell-Type Specificity & Generalization: Incorporates a cell type recognition architecture to account for physiological differences across tissues (e.g., lung vs. liver), enabling zero-shot prediction for new diseases or drugs without extensive retraining.
• Clinical Validation: Demonstrated accuracy in predicting ovarian cancer response to DS-8201 using breast cancer training data, and successfully stratified non-small cell lung cancer patients treated with Osimertinib based on baseline sequencing data alone.
• Multimodal Fusion Algorithm: Proprietary algorithm designed to synchronize and integrate heterogeneous data types captured from the same single cells to reconstruct true regulatory logic.

Industry Insight

• Data Moats via Clinical Partnerships: Success in this field depends heavily on access to high-quality, paired clinical longitudinal data. Companies that establish deep collaborations with top-tier hospitals to build specialized disease models will gain significant competitive advantages over those relying on public or synthetic datasets.
• Shift to Predictive Validation: The industry is moving beyond "finding" targets to "validating" human relevance computationally. AI tools that can reliably predict clinical outcomes before Phase I trials will become essential infrastructure for reducing R&D risk and attracting pharma investment.
• Commercial Viability of Early-Stage AI Bio: This case demonstrates that AI biotech firms can achieve revenue and validate product-market fit early by offering specific, high-value services like patient stratification and preclinical evaluation, rather than waiting years to develop proprietary drug assets.