Article title: BioMap Officially Launches AI Drug Safety Evaluation System, Moving from Virtual Cells to Virtual Monkeys to Bring New Momentum to Drug Development
Body text: On November 19, BioMap officially launched its AI Drug Safety Evaluation System (hereinafter referred to as the "AI safety evaluation system"), combining virtual cell models, toxicity prediction capabilities, and nonclinical safety evaluation services. Relying on safety prediction models such as "virtual monkeys," it extends AI prediction from the cellular level to the animal body.
The official release of the AI safety evaluation system also marks that the company's virtual cell capabilities are beginning to move from public benchmarking into real drug development workflows, further extending from cell state prediction to empowering in vivo animal safety evaluation, helping R&D teams determine, before investing in animal experiments, which candidate molecules are worth prioritizing and which risks need to be verified first.
Currently, BioMap has established partnerships with more than 30 CROs, experimental institutions, and other partners, providing integrated services for drug development covering exploratory toxicology, regulatory toxicology, toxicokinetics, histopathology, and regulatory support.

At the launch event, Zhang Xiaoming, Senior Vice President and Head of AI R&D at BioMap, introduced the AI Drug Safety Evaluation System publicly for the first time
From virtual cells to in vivo animal prediction, enabling key judgments to happen earlier
In recent years, artificial intelligence has already been widely adopted in areas such as target discovery and molecular design, significantly expanding the scope of exploration in early-stage drug development. However, when candidate molecules enter the animal experiment stage, problems such as long R&D cycles, high costs, and slow feedback on results still persist.
Taking experimental monkeys as an example, the purchase price for a single cynomolgus monkey has reached as high as 300,000 yuan, and the dosing phase of some long-term toxicity studies alone requires about six months. For the large number of molecules awaiting screening in early-stage R&D, the cost and time of conducting animal experiments one by one are difficult to bear. The U.S. Food and Drug Administration (FDA) is also promoting the use of new approach methodologies (NAMs) such as computational models for drug safety evaluation, to reduce animal testing in some antibody development.
Particularly in the safety evaluation stage, from dose design and dosing observation to pathology slide reading, data review, and report issuance, all tasks are closely interlinked. If some risks are only discovered late in the study, they may lead to additional studies, data rework, or even adjustments to the R&D plan. BioMap launched the AI safety evaluation system in the hope of moving key judgments forward to before deciding which molecules enter animal experiments. In this workflow, virtual cells can predict the responses that candidate molecules may induce and provide early toxicity clues.
In 2025, BioMap's xTrimoSCPerturb defeated more than a thousand teams from 114 countries around the world to win the first Global Virtual Cell Challenge, validating the model's ability to predict changes in cell states under unknown perturbations. On this basis, BioMap further combined existing animal experiment data to explore virtual dosing through safety prediction models such as "virtual monkeys." Pharmaceutical companies can use this to prioritize more promising candidate molecules, direct limited experimental resources to the molecules most in need of verification, reduce late-stage rework and the time and cost it brings, and gain time for subsequent development and commercialization.
After entering the actual dosing stage, partner CROs verify the risk predictions indicated by the models through animal experiments. BioMap matches the prediction results for the same molecule with real experimental data to validate and iterate the models, and accumulates safety evaluation data that can be used for subsequent R&D. With the accumulation of data and the advancement of model validation, BioMap hopes to gradually expand toxicity prediction capabilities covering cells, organs, and animals, cross-validate prediction results at different levels with real experimental data, provide pharmaceutical companies with a more comprehensive chain of safety evidence, and gradually reduce reliance on animal experiments.

The AI safety evaluation system launched this time does not simply provide a toxicity prediction result, but embeds AI capabilities into key nodes of nonclinical safety evaluation, enabling model prediction, experiment execution, data analysis, and report delivery to form a continuous workflow.
Before a study begins, the model can integrate in vitro cytotoxicity, small animal in vivo data, and historical results for large animals and similar molecules to provide references for species selection, dose grouping, safety windows, and judgment of the first toxicity target organ, helping R&D teams identify some risks before formal animal experiments start.
During the study, the model can continuously analyze in-life data such as body weight, food consumption, clinical tests, and toxicokinetics, identify abnormal trends, deviations in dose-exposure relationships, and potential data gaps, and assist research teams in timely verification and handling of issues.
After entering the necropsy and reporting stage, the model can also complete prescreening of whole pathology slides and annotation of suspicious areas, mark areas requiring focused attention, and conduct consistency checks on raw data, report tables, and regulatory materials, allowing pathologists and researchers to devote more time to professional judgment.
By moving some analysis and verification work earlier, the AI safety evaluation system aims to reduce waiting, serial processing, and late-stage rework in the workflow, enable toxicity signals to be identified earlier than formal reports, and allow reports to enter the review and issuance stage earlier, so that key judgments happen earlier.
Leading results on multiple benchmarks, with models identifying drug toxicity earlier
One of the core models supporting the AI safety evaluation system is xTrimoTox, a toxicity prediction model built by BioMap based on its cell world model.
BioMap conducted a systematic evaluation of xTrimoTox on the TDC public benchmark around key safety evaluation tasks such as hepatotoxicity, cardiotoxicity, genotoxicity, and acute toxicity. The results showed that, among the test tasks selected this time, xTrimoTox achieved results higher than publicly leading models. Among them, the model's AUROC on drug-induced liver injury and hERG cardiotoxicity tasks reached 0.9637 and 0.8918, respectively, and reached 0.9076 in a blinded test of liver injury for 47 clinical drugs. AUROC is a commonly used metric for measuring a model's ability to distinguish between high-risk and low-risk samples; the closer the value is to 1, the stronger the model's discriminative ability. This indicates that the predictive capability of xTrimoTox is not limited to a single toxicity endpoint, but can integrate molecular structure and cellular response information and maintain relatively stable performance across different types of safety evaluation tasks.

BioMap further conducted independent testing of the model on the original Tox21 Challenge dataset. Tox21 is a representative public evaluation system in the field of toxicity prediction, mainly examining whether compounds affect nuclear receptors and cellular stress pathways. In evaluations on 645 official test molecules and 12 toxicity endpoints in Tox21, xTrimoTox achieved an average ROC-AUC of 0.8598, higher than the classic champion model DeepTox (0.8470), as well as publicly leading models such as CLAMP+RF (0.8440), SNN (0.8440), and CheMeleon (0.8380).

Collaborating with CROs and industry experts to build an AI safety evaluation collaboration system
For AI to enter the safety evaluation stage, it requires not only model capabilities but also a mature experimental system, professional talent, and project delivery capabilities.
Around the AI safety evaluation system, BioMap has established a collaboration network with more than 30 domestic CROs and experimental institutions with professional execution capabilities. Partners will provide support in nonclinical experiments, toxicokinetics, bioanalysis, and histopathology according to the study type, technical requirements, and regulatory objectives of different projects. BioMap is responsible for AI models, protocol support, institution matching, process tracking, and technical review, jointly connecting prediction, experiments, data analysis, and regulatory delivery.
This collaboration model not only preserves the professional capabilities of CRO institutions in experimental execution and quality systems, but also enables BioMap's AI models to enter real experimental workflows. Risk alerts provided by the models can be verified through CRO experiments, and the data generated by experiments can in turn feed back into the models, driving continuous iteration of toxicity prediction capabilities.
On this basis, BioMap has also gradually introduced senior consultants and industry experts in the field of safety evaluation to participate in technical system development, assessment of complex projects, and review of model results. Currently, BioMap has launched cooperation with institutions such as the China Academy of Information and Communications Technology to jointly build a life science AI evaluation standards system around practical application scenarios such as AI drug safety evaluation. By bringing together the capabilities of AI R&D teams, CRO partners, and industry experts, BioMap hopes to form an AI safety evaluation collaboration system that combines model innovation, experimental validation, and professional judgment. At present, the relevant capabilities have already entered actual project operations.
Zhang Xiaoming, Senior Vice President and Head of AI R&D at BioMap, said: "After winning the first Virtual Cell Challenge last year, we have been thinking about whether we can turn the model's prediction of cellular responses into a basis for judgment that pharmaceutical companies can use in real R&D. In the past, when discussing the dry-wet loop in AI drug R&D, the industry focused more on molecular design and in vitro validation. Through the AI safety evaluation system, we hope to push this loop one step further, from molecules to cells and organs, and then to in vivo animal safety evaluation, combining predictions at the cell and organ levels with in vivo data, and exploring virtual dosing with 'virtual monkeys.' Models provide clues for experiments, and experimental results in turn test and improve the models, bringing new momentum to drug development."
Liu Tengfei, Deputy Director of the Future Industry Department of the Institute of Industrial Internet and Internet of Things at the China Academy of Information and Communications Technology, said: "Whether AI safety evaluation can create value depends on whether model prediction, experimental validation, and comprehensive evaluation can form a reviewable chain of evidence. BioMap has a globally leading virtual cell model and also has the ability to conduct dry-wet loop validation based on its self-built high-throughput experimental platform. The Institute of Industrial Internet and Internet of Things of CAICT will leverage its accumulated experience in standards development, testing and evaluation, and industry collaboration to work with BioMap on data quality, model performance, and trusted validation, jointly building life science AI evaluation capabilities, and promoting reproducible and credible AI-assisted safety evaluation results so that they can be better integrated into drug development and biomanufacturing workflows."
In the future, BioMap will continue to expand cooperation with CRO institutions, research teams, and industry experts, integrate data at different scales such as cells, organs, and animals, explore the application of life science foundation models in toxicity mechanism research, organ response prediction, and in vivo safety evaluation, and gradually establish a safety evaluation system that can be experimentally validated and can also continuously feed back into experiments. BioMap hopes to use this as a starting point to enable AI to cover more R&D stages of drugs, from molecular design and cellular responses to in vivo evaluation, and provide earlier and more reliable information for candidate molecule screening, risk assessment, and R&D decision-making.
Leiphone (WeChat official account: Leiphone)