Perturbation Heart completes tens of millions of yuan in financing, releases virtual cell model AlphaCell 1.0 Pro

Zhihui Dong (Smart Things) October 30 news: Yesterday, domestic AI for Science emerging company PerturbMind AI announced the completion of tens of millions of yuan in angel and angel+ funding, led by Qingsong Capital, with Tongyang Fund and several pharmaceutical companies participating, officially emerging from stealth mode.

The company focuses on a next-generation AIDD (AI drug discovery) innovation paradigm driven by geometric world models, exploring an R&D model centered on the inference of life system states.

In this announcement, PerturbMind AI also released AlphaCell 1.0 Pro, the world's first virtual cell state inference model under its digital life world model ecosystem PerturbSeek.

The model uses a geometric world model as its underlying architecture, with a standard parameter count of 1B that can be adjusted according to application needs: it can be trimmed down to a 300M lightweight model, AlphaCell 1.0 Flash, or scaled up to a 50B ultra-large-scale model, AlphaCell 1.0 Mega, adapting to the different deployment needs of research institutions, small and medium-sized pharmaceutical companies, and large multinational pharmaceutical companies.

Unlike traditional AI drug discovery, PerturbMind AI hopes to use geometric world models to learn the dynamic changes of life systems under different intervention conditions, infer disease state transitions and system responses after drug intervention, and thereby provide technical support for the analysis of complex disease mechanisms, new target discovery, and R&D of novel-modality drugs.

1. Shifting from target screening to state inference, releasing a virtual cell model

AlphaCell 1.0 Pro released by PerturbMind AI this time is an upgraded version of the company's virtual cell model AlphaCell 1.0 Preview.

On March 5 this year, the PerturbMind AI team led by founder and CEO Chuo Guohui published a preprint paper related to AlphaCell on bioRxiv, introducing the method by which the model simulates dynamic cellular changes caused by perturbations based on a geometric world model architecture.

▲The PerturbMind AI founding team published the AlphaCell 1.0 Preview paper, source: bioRxiv

The standard parameter count of AlphaCell 1.0 Pro released this time is 1B, and it has elastic parameter adjustment capability, allowing the model scale to be dynamically adjusted according to customers' computing power conditions and R&D needs.

Among them, the model can be trimmed down to 300M parameters to form the lightweight version AlphaCell 1.0 Flash, meeting the lightweight and private deployment needs of hospitals, research institutions, and small and medium-sized pharmaceutical companies; it can be scaled up to 50B parameters as AlphaCell 1.0 Mega, providing large multinational pharmaceutical companies (MNCs) with system-level perturbation trajectory inference capability in complex scenarios.

On the commercialization side, PerturbMind AI plans to provide on-device service subscriptions, private deployments, and joint drug R&D pipeline development around models of different scales.

On the technical route, PerturbMind AI hopes to change the traditional target-centric R&D model of AI-driven Drug Discovery (AIDD).

Traditional AIDD usually focuses on known targets to find candidate drug molecules that can bind to them, similar to finding the right key for a known lock. But for complex diseases, a single target may not explain the entire process of disease development, and drug intervention may also trigger changes at other biological system levels.

PerturbMind AI said that disease is essentially a state shift produced after the original homeostasis of the life system is disrupted. Therefore, drug R&D not only needs to find molecules that can act on specific targets, but also needs to understand how disease states form and what systemic changes different interventions may bring.

Based on this idea, the company has shifted its R&D focus from individual targets to state changes across the entire life system, hoping to use model inference to identify interventions that can restore disease states to normal.

2. Geometric world models build the technical foundation, exploring life system inference

Around the above technical route, PerturbMind AI has built the PerturbSeek digital life world model ecosystem, with perturbation omics, geometric deep learning, and geometric world model theory as its three major technical pillars.

Among them, perturbation omics mainly studies the changes produced in biological objects such as genes and proteins after different interventions, providing a data foundation for understanding the dynamic responses of life systems.

Geometric deep learning is used to learn the complex structures behind biological data. PerturbMind AI said that by building geometric deep learning algorithms, it learns and models the Biological Latent Manifold, mapping complex biological data into latent space to represent different cell states and their intrinsic relationships.

On this basis, geometric world models further learn the states of life systems and the patterns of their changes. It incorporates cell states and external interventions into a unified model framework, simulates changes in cell states under the action of different factors such as genes and drugs, and then infers the dynamic evolution process of life systems.

Simply put, this technical route hopes that its model can not only identify the association between a certain intervention and biological changes, but also learn how life systems change under different conditions and infer the possible outcomes of specific interventions.

On this basis, PerturbSeek hopes to combine specific diseases and clinical constraints to identify interventions with potential therapeutic value, rather than merely screening candidate molecules around known targets.

For model training, Perturbation Heart does not advocate simply expanding parameter scale and data volume. The company mentioned that model training in the life sciences field may not be able to directly apply the scaling laws of the general AI field, and that the number of model parameters needs to match actual task requirements, predictive performance, and computational cost.

Therefore, Perturbation Heart hopes to find the balance point between model performance and R&D cost, namely the parameter sweet spot, and improve the model's generalization ability under limited sample conditions by optimizing the underlying architecture and training efficiency.

III. A startup founded by a Tongji research team, working with multiple pharmaceutical companies to promote technology implementation

Dr. Chuo Guohui, founder and CEO of Perturbation Heart, graduated from the Digital Life Agent Laboratory at Tongji University. Chuo Guohui said that the company's underlying technology innovation is to some extent inspired by the work of Google's AI scientific research model DeepMind Alpha series, and hopes to transform cutting-edge world model technology into platform solutions that pharmaceutical companies and clinical settings can use.

▲ Chuo Guohui, founder and CEO of Perturbation Heart, source: Tongji University official website

In his view, living systems have complex dynamic change characteristics, and models not only need to understand the relationships between different biological objects, but also need to characterize the state transitions of the system after external intervention. Perturbation Heart hopes to establish state representations and state transition models of digital life systems through geometric world models, and transform inference results into drug R&D leads.

The company's chief scientist and head of the Digital Life Agent Laboratory at Tongji University, Professor Liu Qi, said that R&D of new modality drugs requires holistic understanding and modeling of complex living systems. Traditional AIDD has accumulated considerable technology in molecular design for known targets, but still faces challenges in analyzing complex disease mechanisms and discovering new targets.

▲ Liu Qi, chief scientist of Perturbation Heart and head of the Digital Life Agent Laboratory at Tongji University, source: Tongji University official website

Liu Qi said that Perturbation Heart hopes to advance drug R&D from mere molecular-level modeling to understanding and inference at the level of living system states, and then form a closed loop through experimental validation and data feedback, establishing a connection between digital models and real living systems.

In terms of scientific research accumulation, the Perturbation Heart team has cooperated with domestic and international pharmaceutical companies and contract research organizations (CROs) such as Roche, AstraZeneca, and WuXi AppTec. Relevant research results from the team have been published in journals such as Nature Methods, Nature Machine Intelligence, Nature Computational Science, and Cancer Cell, and have received the WuXi AppTec Life Chemistry Research Award and the Wu Wenjun Artificial Intelligence Natural Science Award, among others.

In terms of industrial cooperation, Perturbation Heart said that the company has signed a strategic cooperation agreement with a leading domestic chip company, and the PerturbSeek engine has completed full-stack adaptation to domestic GPU chips.

In addition, the company has signed cooperation agreements with drug R&D companies in specific domestic fields, is developing and deploying multiple prototype systems, and is building data and experimental feedback loops around specific application scenarios.

Conclusion: As the life world model moves toward industry, when will its actual R&D value be realized?

However, whether world models can truly understand complex living systems still needs to be tested through continuous experimental validation and actual drug R&D projects.

For Perturbation Heart, how to transform model predictions into verifiable drug R&D leads and further promote new target discovery and new drug R&D is a question it needs to answer for subsequent commercialization.

Source: Perturbation Heart