Genesis Molecular AI, a leader in molecular AI for drug design, has unveiled Pearl, a cutting-edge foundation model that excels in biomolecular structure prediction. Announced on October 28, 2025, in Burlingame, California, Pearl utilizes a novel architecture and advanced training methods combined with extensive synthetic data.
A recent study co-authored with experts from NVIDIA reveals that Pearl outperforms existing models, including AlphaFold 3, in accurately predicting the binding of small molecules to proteins. This capability addresses a critical challenge in drug discovery, often termed the “holy grail” of the field, promising to enhance the design of new medicines for patients with significant unmet medical needs.
Unlike large language models that benefit from abundant online training data, AI applications in biochemistry face a scarcity of high-quality experimental data. Pearl, a fully integrated diffusion model, significantly enhances existing methodologies by leveraging Generative AI techniques even in low-data scenarios. The model innovatively incorporates physics into its input, architecture, and outputs.
A major advancement in Pearl”s performance stems from its training on vast synthetic datasets generated from simulations, which mitigates the lack of high-quality experimental structural data. The findings demonstrate that Pearl”s capabilities improve as more simulated data is utilized, marking a notable discovery in the scaling properties of synthetic data within AI-driven drug discovery.
“The primary challenge in applying AI to drug discovery has been the limited availability of high-quality biomolecular data, which are both costly and time-consuming to gather,” stated Dr. Aleksandra Faust, Chief AI Officer at Genesis. “Our approach, inspired by the success of autonomous vehicle models using simulated data, integrates physics-generated data into Pearl”s training while improving sample efficiency in data-scarce environments. This dual strategy represents a substantial advancement for the future of drug discovery.”
Rigorous evaluations of Pearl against AlphaFold 3 and various open-source cofolding models—such as Boltz-1, Boltz-2, Chai-1, and Protenix—utilized standardized protocols to assess accuracy and the physical validity of predicted structures. Pearl is designed to enhance the workflows of drug discovery scientists. Not only does it outperform other models in pure cofolding scenarios, but it also integrates expert conditioning during inference, allowing for the incorporation of target-specific data to further optimize performance in practical applications.
“AlphaFold 3 represented a monumental achievement, and Pearl is the first model to exceed its performance,” remarked Dr. Evan Feinberg, founder and CEO of Genesis. “It is widely recognized in the biopharma industry that while cofolding models may excel in certain metrics, they often fall short in real-world applications, occasionally producing significant physical inaccuracies. Our team has focused on developing Pearl as an essential component of our comprehensive GEMS platform, enabling the targeting of challenging drug candidates throughout various stages of our internal and partner drug development programs.”
In November 2024, Genesis secured additional funding from NVentures, the venture capital arm of NVIDIA, to enhance computational methods for AI-driven drug discovery. This collaboration has led to the integration of NVIDIA”s cuEquivariance kernels for triangle operations, resulting in a 15% relative speedup in training and a 10-80% speedup in inference. The partnership also aims to optimize inference operations for broader deployment of Pearl in Genesis”s drug programs and partnerships.
“Next-generation foundation models like Pearl, which merge physics and AI, are paving the way for new insights into molecular interactions,” stated Anthony Costa, Director of Digital Biology at NVIDIA. “The accelerated computing capabilities provided by NVIDIA, including libraries like cuEquivariance, are crucial for scaling these breakthroughs.”
For further information, readers can visit the Genesis website and access the complete technical report.
