Co-folding models, such as AlphaFold 3, have not yet attained practical precision in real-world drug development settings. This is mainly because publicly accessible data on protein-drug molecule pairings is limited, and there's the challenge of effectively leveraging the proprietary data of pharmaceutical companies while preserving confidentiality. Five pharmaceutical companies, AbbVie among them, employed Apheris's federated learning framework, grounded in the OpenFold3 Preview 2 model, and integrated it with over 20,000 private protein-ligand structural data entries to train the AISB-1-Fed model. This model surpasses public models and those refined by individual pharmaceutical companies in key metrics, especially excelling in predicting protein-protein interfaces. Tests verified that the model did not compromise any data, with its enhanced performance primarily stemming from the utilization of proprietary data. At present, most co-folding models are trained on PDB data, which leads to problems like 'rote learning' and limited generalization. To tackle this, the industry is seeking data solutions: beyond the federated learning method embraced by AISB, Eli Lilly has introduced the TuneLab platform, and new public data initiatives like OpenBind and LIGAND-AI have also surfaced. The industry perceives federated learning as merely a transitional solution, with freshly generated data carrying more significant long-term worth. Moreover, federated learning can only offer partial support for drug development; the upcoming stage of co-folding model advancement demands a transition from static structure prediction to dynamic modeling, necessitating both existing proprietary data and newly created data.