Former ByteDance AI scientist touts drug-discovery model that tops world leaders

A former ByteDance artificial intelligence scientist’s new venture has released a life sciences AI model that the firm claims outperforms state-of-the-art systems, intensifying the increasingly crowded competition in AI-driven drug discovery.

Geodesic Intelligence’s NovaAtom-Lite, an AI model designed to predict the structures and interactions of biomolecules, topped a third-party benchmark in several drug discovery tasks, founder Quanquan Gu said on social media on Wednesday.

On the FoldBench benchmark, he said, the model correctly predicted 79.1 per cent of antibody-antigen interfaces – the mechanisms by which an immune system proteins bind to specific targets.

That placed it ahead of IsoDDE, a platform developed by Google DeepMind spin-off Isomorphic Labs, a widely recognised leader in AI-driven drug design. IsoDDE scored 75.6 per cent on the same FoldBench task, according to a technical report published in February.

Gu said Geodesic’s new model also far surpassed DeepMind’s AlphaFold 3, a pioneer in AI-based protein design, which had a success rate of 47.9 per cent on the same task.

Gu previously led the AI drug design team at TikTok parent ByteDance, which in January this year released the SeedFold biomolecular structure prediction model. He was also a key member behind the Chinese tech giant’s flagship Seed large language models.

The AI scientist, also an associate professor of computer science at the University of California, Los Angeles, announced his departure from ByteDance in June.

Gu announced the formation of Menlo Park, California-based Geodesic in September, saying the start-up’s goal was to “build artificial general intelligence for drug discovery and find the shortest path from biology to medicines”.

Geodesic’s new model comes as frontier AI developers and global pharmaceutical companies double down on bets that large language models can transform the discovery and design of new medicines.

Anthropic said in its initial public offering prospectus in late October that powerful AI can accelerate drug discovery and disease detection, possibly “compressing decades of scientific progress into years”. Major drug makers including Eli Lilly, Roche and Novo have all made major moves to integrate AI across their operations.

In China, 44 AI drug discovery start-ups raised 15.8 billion yuan (US$2.4 billion) across 57 deals within the first nine months of this year, according to September data from CVSource, a Chinese private investment data platform.

BioMap, an AI drug discovery start-up backed by Chinese search giant Baidu, said in an event in Hong Kong last week that it was working with Hong Kong’s Baptist University to establish a lab using AI to study the mechanisms of traditional Chinese medicine.

However, many remain sceptical about how soon the massive investments in AI drug discovery will yield tangible results.

An AI model’s accuracy in benchmark tests did not guarantee actual efficiency improvements in real-world drug research and development, Northeast Securities analysts said in a research note published earlier this week.

Bruce Booth, a partner at US venture capital firm Atlas Venture, said in a blog post last month that AI today was “predominantly democratising design around known targets, known drug scaffolds and known biology”.

This could in the near term exacerbate the biopharma industry’s problem of being crowded with “me-too” drugs, he said.