BioMap’s “Life Science Leaderboard Unlocks the Potential of Life Science AI Models

BioMap, a company with an AI foundation model-powered life science platform, today announced the official launch of the “Life Science Leaderboard” – a living resource for industry professionals and researchers to evaluate the performance of life science AI models. AI is undergoing a remarkable paradigm shift, with pre-trained large models evolving at an astonishing rate and demonstrating unprecedented capabilities. It is therefore becoming increasingly important to establish objective and accurate evaluation methods to provide reliable performance benchmarks and identify areas for improvement.

Currently, BioMap integrates a disruptive life science AI foundation model and super-scale computing with cutting-edge biotechnology to discover novel targets and design novel functional proteins. The company was co-founded by Robin Li, the Founder and CEO of Baidu, in late 2020.

In response to this unmet need, BioMap has officially launched the Life Science Leaderboard – the first comparison tool that allows users to identify recent progress and gaps in AI capabilities currently available from industry leaders.

BioMap invites professionals and researchers to utilize the Life Science Leaderboard to evaluate life science AI models and to collaborate in accelerating the growth and prosperity of the industry. By visiting http://www.biomap.com/sota/, users can access detailed information about the leaderboard, request task datasets, and participate in challenges.

The CEO of BioMap, Wei Liu, commented, “Our aim is to foster a collaborative environment for experts in large models by providing an open communication platform where they can share tasks and data. Together, we can unlock the potential of AI in the field of life sciences.”

Since its inception, BioMap has been dedicated to developing its life science AI foundation model “xTrimo,” which stands for Cross-model Transformer Representation of Interactome and Multi-Omics. xTrimo is a 100B-parameter cross-modal pre-trained large language model that integrates billions of multi-dimensional data from the amino acid level to the bio system level to understand and predict the behavior of life at multiple scales of complexity. It is composed of a pre-trained life science AI model as well as multiple downstream task models, utilizing a multi-layer nested structure design architecture.

The xTrimo model learns the key rules of how proteins are composed, function, interact, combine, and regulate cellular functions in order to generate new insights into the “natural language” of life. As well as successfully completing basic tasks such as structure prediction, xTrimo can also generate de novo protein designs to solve specific biological challenges.

Over the past two years, BioMap has conducted thorough and consistent evaluations of the xTrimo model as it has evolved to more than one hundred billion parameters.

Building on this, BioMap developed the Life Science Leaderboard, which encompasses more than 25 of the most important, representative, and cutting-edge tasks across six distinct protein design categories. The tasks cover various areas, including antibody structure, antibody function, drug development, disease treatment, and cell research, among others, and hold significance in advancing research and practical applications in life sciences.

Currently, the “xTrimo” model has attained SOTA (state-of-the-art) results in more than 25 downstream protein design tasks and is still undergoing continuous iteration and evolution.

Dr. Le Song, the CTO of BioMap, commented, “The birth of billion-scale models and the integration of cutting-edge AI technology and biotechnology signifies the beginning of a new golden age for the life science industry, creating hope of a brighter future with greater understanding and novel therapies for currently untreatable diseases.”