Genesis Molecular AI has unveiled Pearl, a groundbreaking generative foundation model for biomolecular structure prediction that sets a new industry standard in drug-protein interaction modeling. The model, developed using a novel architecture, advanced training methodology, and large-scale synthetic data, outperforms even AlphaFold 3 in accurately predicting how small molecules bind to proteins—a critical challenge in drug discovery often referred to as the “holy grail” of biopharma. Unlike large language models that benefit from vast internet-sourced data, AI in biochemistry faces a severe shortage of high-quality experimental structural data. Pearl overcomes this limitation by leveraging physics-based simulations to generate massive amounts of synthetic training data. This approach enables the model to scale performance as more simulated data is introduced, offering the first evidence of synthetic data scaling laws in AI-driven drug discovery. This breakthrough allows Pearl to generalize far beyond the constraints of limited public datasets. Pearl is an end-to-end diffusion model that integrates physical principles into its input, architecture, and output design. A key innovation is its use of physics-informed synthetic data, which enhances accuracy and physical plausibility. According to Aleksandra Faust, Ph.D., Chief AI Officer at Genesis, this dual strategy—combining synthetic data with highly efficient training methods—represents a major leap forward in applying AI to low-data domains like molecular biology. In rigorous benchmarking against AlphaFold 3 and open-source cofolding models such as Boltz-1, Boltz-2, Chai-1, and Protenix, Pearl demonstrated superior performance in both accuracy and structural validity. Notably, Pearl excels not only in standard evaluation settings but also in real-world deployment scenarios, where it can incorporate expert knowledge during inference to further improve predictions on complex and flexible protein targets. Evan Feinberg, Ph.D., founder and CEO of Genesis, emphasized that while AlphaFold 3 was a landmark achievement, Pearl is the first model to surpass it. He highlighted that many existing cofolding models struggle with generalization and sometimes produce physically implausible results—limitations that Pearl is designed to overcome. Pearl is a core component of Genesis’s GEMS (Genesis Exploration of Molecular Space) platform, which integrates AI and physics to accelerate drug discovery across challenging therapeutic targets. The announcement follows a strategic collaboration with NVIDIA, announced in November 2024, involving an investment from NVentures and technical integration of NVIDIA’s cuEquivariance kernels. These optimizations have accelerated Pearl’s training by 15% and boosted inference speeds by 10% to 80%, enabling scalable deployment across Genesis’s internal programs and partner initiatives. Anthony Costa, Director of Digital Biology at NVIDIA, praised the synergy between physics and AI in models like Pearl, noting that accelerated computing platforms are essential for advancing next-generation molecular AI. For more details, including the technical report and research paper, visit genesis.ml/pearl_technical_report/. Genesis Molecular AI, headquartered in the San Francisco Bay Area with offices in San Diego and New York, has raised over $300 million from top-tier investors in AI, technology, and life sciences. The company has established multiple research partnerships with major pharmaceutical firms and is actively using its GEMS platform to develop novel therapeutics for high-impact diseases.