Genesis Molecular AI, the company pioneering the world’s leading molecular AI models for drug design and development, has unveiled Pearl, a generative foundation model for biomolecular structure prediction. According to new research co-authored with experts from NVIDIA, Pearl outperforms all existing models — including AlphaFold 3 — in predicting how small molecules bind to proteins, a long-standing challenge in biochemistry and pharmaceutical R&D.
A Leap Beyond AlphaFold 3
Drug-protein binding prediction has long been considered the holy grail of drug discovery. Solving it enables scientists to design more precise and effective medicines for patients with previously untreatable conditions. While traditional large language models have benefited from abundant internet-scale data, biomolecular AI faces a severe data scarcity problem — limited by the cost and complexity of generating experimental structural data.
Pearl overcomes this limitation by introducing a novel diffusion architecture and a training methodology that integrates physics-based synthetic data at scale. This approach allows the model to learn efficiently in low-data regimes and generalize to novel molecular structures far beyond its training set.
“One of the biggest roadblocks in applying AI to drug discovery is the lack of high-quality biomolecular data,” said Aleksandra Faust, Ph.D., Chief AI Officer at Genesis. “Inspired by how autonomous vehicle AI used simulated data to train safely, we applied a similar principle to biomolecular systems. By combining physics-generated synthetic data with improved training sample-efficiency, we’ve made a significant leap toward unlocking AI’s full potential in drug discovery.”
Benchmarking Performance: Pearl vs. AlphaFold 3
In a head-to-head evaluation against AlphaFold 3 and several open-source cofolding models (including Boltz-1, Boltz-2, Chai-1, and Protenix), Pearl achieved the highest overall accuracy and physical validity across all benchmarks:
- Up to 40% higher relative performance than AlphaFold 3 on external benchmarks
- Stronger generalization across real-world drug discovery datasets
- Superior results on Genesis internal drug program benchmarks
Unlike traditional cofolding models, Pearl was engineered for deployment. It can integrate expert conditioning during inference, enabling researchers to use domain-specific data and enhance accuracy for complex, flexible protein targets.
“AlphaFold 3 was a historic, Nobel-worthy breakthrough, and Pearl is the first model to surpass it,” said Evan Feinberg, Ph.D., founder and CEO of Genesis. “Pearl doesn’t just predict structures — it enables drug design. Integrated into our GEMS platform, Pearl gives our scientists and partners the ability to tackle previously undruggable targets and design a new generation of therapeutics.”
Scaling AI for the Future of Biotech
Pearl also represents one of the first proven cases of synthetic data scaling laws in drug discovery — demonstrating that performance continues to improve as simulated data grows.
As part of its collaboration with NVIDIA, Genesis integrated cuEquivariance kernels into Pearl’s architecture, resulting in a 15% training speedup and 10–80% faster inference. The partnership will further optimize large-scale inference for real-world deployment across Genesis’s internal and partner programs.
“Next-generation foundation models like Pearl, which combine the power of physics and AI, are opening new frontiers in molecular understanding,” said Anthony Costa, Director of Digital Biology at NVIDIA. “NVIDIA’s accelerated computing platform, featuring libraries like cuEquivariance, is essential for scaling these innovations.”
Power Tomorrow’s Intelligence — Build It with TechEdgeAI
