Tackling Fragmentation in AI Drug Discovery

In modern drug discovery, scientists often grapple with fragmented workflows that hinder innovation. Critical tasks, such as hypothesis generation, molecular design, and multidimensional data analysis, are often performed in silos using disparate tools, making it difficult to iterate quickly and reproduce results. To address these challenges, Insilico Medicine, a clinical-stage biotechnology company specializing in generative artificial intelligence (AI), has collaborated with Microsoft on a demonstration of its Nach01 foundation model deployed on the Microsoft Discovery science-focused platform for AI-accelerated research and development (R&D). Together, these capabilities demonstrate Microsoft Discovery’s extensibility with third-party models that could be used to create an AI-native workflow for drug discovery that spans biology and chemistry. This approach is designed to improve transparency, reproducibility, and scalability in R&D organizations.

Nach01 is a multimodal foundation model developed and provided by Insilico Medicine for accelerating drug discovery. It leverages language model architecture and molecular point cloud encoder to process both chemical structural and spatial data and was built to handle hundreds of diverse tasks in molecular design and prediction. Nach01 combines techniques from its predecessors (such as textual and 3D spatial representation learning) into a single model intended to address molecular property prediction, compound design, and other complex chemistry challenges.

Microsoft Discovery: An Enterprise Agentic AI Platform for Research and Development

Microsoft Discovery is Microsoft’s enterprise agentic platform designed to accelerate research and development. Built on Azure, it provides an environment where R&D teams can orchestrate multi-step investigations, integrate AI models with data and tools, and manage end-to-end workflows seamlessly.

Key capabilities of Microsoft Discovery include:

• Cognition engine: This AI-powered system that oversees complex projects, manages tasks, and breaks down scientific challenges into traceable, actionable steps actions.
• Knowledge generation: Internal documentation, Electronic Notebooks, and results of previous Design Make Test Analyze (DMTA) cycles are indexed with graph-based retrieval (GraphRAG), giving researchers access to the corpus of organizational data.
• Scalable compute orchestration: Microsoft Discovery leverages Azure ML Workspace to deploy and scale AI models, such as Nach01, on elastic infrastructure. This approach ensures that compute-intensive workloads such as generative chemistry and large-scale inference are executed efficiently across CPU and GPU clusters.
• Data connectivity: Microsoft Discovery integrates with Azure’s Identity and Access Management (via Microsoft Entra ID, formerly Azure Active Directory) and data services. By offering these capabilities, Microsoft Discovery transforms ad-hoc modeling scripts into managed, repeatable processes, allowing Scientists and engineers to focus on scientific logic and experimentation while Azure handles the underlying infrastructure.

Integration: Orchestrating Nach01 in Microsoft Discovery

In the future, when customers decide to make use of Nach01 from Insilico and deploy it on the Microsoft Discovery platform, they will be able to integrate Nach01 seamlessly into broader discovery workflows. In practical terms, this means that such customers could extend the functionality of Microsoft Discovery to potentially do the following:

• Compose AI-Driven Workflow: Using Microsoft Discovery, researchers could develop an end-end computational workflow using Nach01 and other computational tools and data sources as appropriate. For example, an investigation might start with a structure-based design step (using Nach01 to propose novel molecular structures given a biological target), then feed those structures into a prediction step (also powered by Nach01 or another specialized model) and finally pass the results into a data analysis or visualization tool of their choice. All these steps are configured in one workflow using Microsoft Discovery cognition engine.
• Execute and Iterate: Once the investigation is set up, it can scale based on demand, Microsoft Discovery manages the execution, spinning up compute resources as needed and routing data between steps. After execution, researchers can review the outcomes (e.g., list of candidate molecules with their predicted properties) and, if needed, adjust parameters or swap components to refine the investigation. They can then re-run the updated investigation, being confident that the process is still traceable and consistent with the previous runs for comparison.
• Reproducibility and Handoff: The entire configuration (which models were used, with what versions and settings) is saved in Microsoft Discovery. This means another team member or even another team could reuse the investigation blueprint. For example, a scientist in translational chemistry could take an investigation developed initially by an AI engineer and rerun it on a new target or new data.

Here is a walk-through of a drug discovery project aiming to find a new inhibitor for a protein, illustrating how customer extensibility of Microsoft Discovery with Nach01 can be used to streamline each step of the workflow.

1. A target identification step might start the process (e.g., confirming the disease relevance of the protein via knowledge graphs or literature mining – this step could involve other AI services outside Nach01’s scope).

2. Next, Nach01 is invoked for hit generation: it generates a set of novel chemical structures predicted to bind the protein’s active site (using its generative chemistry capability).

3. Each proposed molecule is then evaluated by Nach01’s predictive modules for properties like solubility and toxicity (ADMET predictions).

4. The investigation might then use a third-party QSAR model or docking tool (also orchestrated by Microsoft Discovery) for additional scoring of each molecule.

5. Finally, the results are compiled, and an automated report or visualization is created as shown in Figure 1.

Fig 1: Top 10 proposed virtual structures ranked using Nach01 on Microsoft Discovery.

This entire multi-step workflow can be run on the Microsoft Discovery platform, with Nach01 extending the functionality of the platform. If any part of the process produces suboptimal results, researchers can tweak and rerun quickly – speeding up the DMTA cycle that is central to drug discovery.

Conclusion

This example demonstrates how the extensibility of Microsoft Discovery with cutting-edge third-party generative chemistry models can be used by customers to tackle one of the biggest challenges in computational R&D, the end-to-end integration of many complex steps in a secure, efficient manner. This also demonstrates how researchers in biotechnology and pharmaceutical organizations can experiment with AI-driven drug discovery ideas faster and with greater confidence that their results are reproducible and scalable. By using Nach01 within Microsoft Discovery, teams can focus on scientific questions (e.g., finding the right molecule for a given target) rather than technical logistics, thereby accelerating innovation (as evidenced by the rapid hypothesis testing capabilities) and fostering better collaboration between AI specialists and bench scientists. Microsoft Discovery’s orchestration, combined with third party domain specific models, such as Nach01, has the potential to enable a new class of solutions for drug discovery, where AI is deeply woven into the scientific process, from start to finish. “From our foundational patents on mutual information in chemistry and biology in 2018 to our deep, ongoing research in multimodality, Insilico has consistently pushed the boundaries of generative AI. While Nach01 is a small, specialized model, its successful launch demonstrates the immense power and scalability of the Microsoft platform. We are now applying these learnings to develop large, highly multimodal models aimed at delivering SOTA and SOTA+ results in every aspect of drug discovery—a critical step toward scientific superintelligence”, says Insilico Founder and CEO Alex Zhavoronkov, PhD. Combining advanced AI models with robust cloud agentic platform, the industry moves closer to a future where AI is a first-class citizen in drug discovery, accelerating the journey from scientific insight to life-saving medicine.