發生了什麼事
Deep Origin and the PREDICTS consortium announced that 62 of their 77 in-development machine learning models for predicting drug absorption, distribution, metabolism, excretion, and toxicity (ADMET) are now being implemented by four organizations in a pilot program. These models, built on Deep Origin's foundational chemistry model Togo, are being used by ImmVue, Sanford Burnham Prebys, Synko, and SyzOnc to filter millions of compounds and prioritize drug candidates. The initiative is part of the ARPA-H CATALYST program, which awarded the consortium up to $31.7 million in 2025 to develop human-based safety prediction models.
Deep Origin and the PREDICTS consortium, a group including Ginkgo Bioworks and Sanford Burnham Prebys, announced that their AI-powered ADMET models are now in active use by four partner organizations. The pilot program involves ImmVue, Sanford Burnham Prebys, Synko, and SyzOnc, which are using the tools to assess and filter millions of compounds across antiviral, antibacterial, oncology, and immune-mediated disease programs.
The models are part of the ADMET-NOW platform, which consists of 77 machine learning-based predictors. Currently, 62 of these models are available to pilot participants. These models are built on Togo, Deep Origin’s foundational chemistry model, which is trained on a wide range of molecular and protein-ligand tasks. Togo allows for the training of new property models on fewer than 1,000 data points, addressing the challenge of small public datasets for many ADMET endpoints.
According to the source, 89% of the 62 available models outperform the best-performing models identified in the literature for specific endpoints, including genotoxicity, human ether-à-go-go-related gene inhibition, and drug-induced liver injury. The consortium was awarded an up to $31.7 million Other Transaction Agreement from ARPA-H in 2025 to develop these in silico models under the CATALYST program.
為什麼這很重要
This deployment marks a practical shift from theoretical AI drug discovery to active implementation in preclinical workflows. By providing models that outperform existing literature benchmarks for key safety endpoints like genotoxicity and liver injury, the consortium addresses the critical bottleneck of small public datasets in toxicology. The feedback loop, where partners return experimental results to improve training data, creates a scalable framework for safer and faster drug development, particularly for rare diseases where traditional clinical trial data is scarce.
The deployment of these models represents a significant step in applying AI to preclinical drug safety, a stage where high failure rates often stall development. By accurately predicting toxicity and safety profiles, the models help prioritize candidates that are more likely to succeed in later stages, potentially reducing the time and cost of drug development.
The initiative is particularly relevant for rare disease populations, where traditional clinical trial data is limited. The CATALYST program aims to create human-based models that capture more representative physiologies, aligning with the goals of the U.S. FDA’s Modernization Act to improve the efficiency and safety of drug approval processes.
The collaborative feedback loop, where partner laboratories share experimental results back to the consortium, ensures that the models continue to improve with real-world data. This approach addresses the limitation of small public datasets and creates a more robust and accurate prediction system over time.
互動機制：它實際上是如何運作的
以互動方式探索這項發展背後的基礎技術。
In AI, what are a model's "parameters"?
接下來看什麼
Monitor the performance validation results as partners share experimental data back to the consortium. Watch for the expansion of the pilot to additional organizations and the development of the 'Virtual Human Avatars of Toxicology,' which aim to simulate complex physiological states and comorbidities to improve safety predictions for special populations.
The next phase of the project involves the development of 'Virtual Human Avatars of Toxicology.' These avatars are designed to simulate special populations and biological situations, including physiological traits, comorbidities, and mutations, to increase the human-relevance of safety predictions.
Deep Origin is developing organ models for the liver, kidney, intestine, blood coagulation, and bone marrow that are connected to simulate the physical and chemical events inside a body. This mechanistic approach to toxicology is distinct from statistical 'digital twins' and aims to predict drug safety by tracking pharmacokinetics and metabolic interactions.
The performance of the models will be further validated as partners share their experimental results. This data will be used to assess model performance and continue adding to the training data, potentially leading to the expansion of the pilot program to more organizations and therapeutic areas.