AiChemy: How MCP-Powered Agents Cut Drug Discovery Research from Weeks to Minutes

Summary

AiChemy is a multi-agent research assistant that connects external MCP servers for databases like OpenTargets, PubChem, and PubMed with proprietary chemical libraries on the Databricks Data and AI platform, enabling researchers to find drug targets and candidate compounds in minutes instead of days.
Model Context Protocol (MCP) acts as a standardized, USB-style interface for connecting tool kits to agents without moving data, and Databricks assets like Genie, Vector Search, and Functions are automatically exposed as MCP servers for seamless agent integration.
This video includes a live demo building a fully functional research agent with a front-end application in under five minutes using Databricks Agent Bricks, with zero code written.

AiChemy: How MCP-Powered Agents Cut Drug Discovery Research from Weeks to Minutes

Drug discovery demands sifting through millions of compounds, thousands of papers, and dozens of databases—a process that costs the industry an average of $2.6B and over 10 years per approved drug. Researchers spend days cross-referencing OpenTargets for genetic evidence, PubChem for chemical properties, and PubMed for supporting literature. What if agents could do it in minutes?
Meet AiChemy, a multi-agent assistant that combines external MCP servers like OpenTargets, PubChem, and PubMed with your own chemical libraries on Databricks. Add Agent Skills to find druggable targets, drug candidates and their chemical properties and safety profiles, complete with supporting evidence.
You can build your own multi-agent assistant in just 5 minutes using our no-code Agent Bricks. Plus, Databricks assets like Genie, Vector Search, and Functions are automatically turned into MCP servers for seamless integration. Stop scrolling through endless articles and databases—let your agents do it for you!
Talk By: Yen Low, Sr Specialist Solutions Architect, Databricks ;

Chapters

FAQs

What is AiChemy and what problem does it solve in drug discovery?

AiChemy is a multi-agent research assistant built on the Databricks Data and AI platform that coordinates specialized agents across genetic evidence databases, chemical property sources, and scientific literature. It addresses the fact that drug discovery requires researchers to spend days cross-referencing sources like OpenTargets, PubChem, and PubMed manually—a process AiChemy can complete in minutes.

What is Model Context Protocol (MCP) and why does it matter for research agents?

MCP is an open standard described by its creators, Anthropic, as the USB interface for connecting tool kits to AI agents without needing to move data between systems. With MCP, an agent can query a database like PubMed or a custom chemical library through a standardized connection, eliminating the need to build bespoke API integrations for each data source.

How are Databricks platform assets exposed as MCP servers?

Databricks automatically turns platform assets—including Genie Spaces, Vector Search indexes, and registered Functions—into MCP servers that agents can call directly. This means data teams can add their own curated knowledge and models to an agent's toolkit without writing custom integration code, making the Databricks marketplace a growing library of ready-to-use agent tools.

How quickly can a new drug discovery agent be built on Databricks?

Using Databricks Agent Bricks in the no-code Playground, a fully functional multi-agent research assistant with a front-end application can be built and deployed in approximately five minutes without writing any code. The demo in this video walks through the complete workflow from a blank canvas to a deployed agent ready for research queries.

Full transcript

[00:07] So, welcome everyone. I hope you're having a great time here at the summit. I'm Yen Low. Was a farmer former farmer chemist turned AI specialist at Databricks. I focus on our health and life sciences
[00:19] customers. So, I'm going to show you AI Chemist today. It's our research agent. Um where you can connect various MCPs, skills, and custom data.
[00:32] And then we are going to build one together in just 5 minutes. So, let's get started. Now, those of you in drug discovery will probably know this very well that it is
[00:44] a long and difficult process with high attrition. Takes 10 to 15 years to go from molecule to market and 90% are going to fail. So,
[00:57] and it takes teams of highly specialized experts, your chemists, your biologists, your informaticians, regulatory specialists all coming together to discover and develop the drug.
[01:10] But now we can have agents to help coordinate these multidisciplinary interactions because it can have more data at its fingertips and therefore generate faster, possibly better
[01:24] hypothesis, or even go on to validate them. And this accelerates the scientific inquiry process because you can make more informed bets leading to more successful
[01:37] experiments and therefore saving you time, labor, and valuable resources. And data integration is made easier now with MCP. No more wrestling with APIs,
[01:51] moving gigabytes of data. Um so, what is MCP, uh model context protocol? It's an open standard for tool kits to be connected to agents. Um the
[02:05] makers, Anthropic themselves, call it the USB interface, where you can just plug in a tool kit to your agent. Um some also call it the new API for agents, and many MCPs are actually wrapped
[02:19] around existing APIs. Um and the key thing is you don't have to move your data. You just connect to it. And all these MCPs are now available on
[02:34] our marketplace. So, if you go to the marketplace on your workspace, you'll see the following: PubMed, the US Census, as well as uh MCPs with specialized skills from our
[02:47] market from our partners. We have Arterys Health and Kythera. And I'm going to do a demonstration where we're going to build an agent with these MCPs and a front end in 5 minutes.
[03:01] So, Icomi is an example of a research agent with built with Data Bricks tools. It connects the various MCPs. It doesn't
[03:14] need to be this this combination of MCPs. You can mix and match. Um and since I don't have real chemical data or lab data, right? I'm going to use
[03:26] public sources like DrugBank, um as well as the ZINC chemical library of a quarter million molecules to stand in for a chemical library.
[03:41] Now, these structured tables like DrugBank, um you can make them Genie agents with text-to-SQL abilities. And then you have unstructured data like text, or in this case, molecular
[03:53] structures, you can put them in a vector search and that's what we did with a zinc chemical library of a quarter million molecules. And then if your MCP servers have certain tool gaps, right?
[04:05] You can actually write custom Python function as tools to plug those gaps. So here we going to we have a couple. One is to compute the molecular embeddings and another one to do model inferencing.
[04:17] We're going to predict the admin properties from the smiles representation of a molecule. So let's check out I can be itself. So this is the app.
[04:30] I'm going to just run the first query. Now this is a data bricks app wrapped around a land graph agent.
[04:42] And you can see the there are many model choices. So I go went with sonnet cuz it's fast. And then because we're going to ask a bunch of drug chemical questions, I selected
[04:54] pubchem as well as the zinc library in the vector search as well as some chemical utility functions that I wrote. So you saw what was returned, right? I asked
[05:06] for the molecular image of aforglipron and it returned this and the LM sonnet 4.5 is not making it up. It's actually >>
[05:19] Campbell ID and then going to using another tool called Campbell get molecule to bring up based on that Campbell ID to bring up the following information about the molecule. So all this is grounded
[05:32] because we provided it with the right tools to use. Next, I'm going to call um I'm going to do a similarity search. I'm going to try and find molecules that
[05:44] are most similar to aforglipron. And the reason for doing this is because um orforglyprone, off better known as found AO, is the latest GLP-1 drug just
[05:56] approved in April, right? And there's a lot of excitement over it because it is the only small molecule GLP-1, right? All the others have been peptides which had to be injected. They were difficult
[06:08] and expensive to make, but not this small molecule version that looks like this. So, researchers are studying its structure very closely to see if they can find similar with similar
[06:21] therapeutic properties. And that's what it's doing now. It's searching through our vector search. Um so, give it a few more seconds and should find the following. While we wait for it, I'm
[06:33] going to take you to um MLflow, which is tracing all these calls we've just made for this agent, right? So, this was the call we just made um at 2
[06:45] 2:05, right? It's already here and you can see the step-by-step agent flow, right? It made several two calls, uh Campbell search molecules and
[06:58] Campbell get molecule. You can see the actual content in itself, which we then have which we have passed uh into this right-hand side over here. So, all this comes out of the box,
[07:11] right? Uh you don't have to set up additional instrumentation, get another vendor for that level of observability. So, these from the second query, right? We've got the following results back. It found three zinc molecules um and
[07:26] provided the smiles, and then we can go on and do further investigation. And in this case, I'm going to try and predict the admin properties for the most similar molecule. So, this is a fairly
[07:38] quick query. It's going to call a model on Databricks and do model inferencing, right? So, it it computed all the ADME um
[07:50] pharmacokinetic properties. And so, here you can see that it actually predicted the Caco-2 permeability, blood-brain barrier penetration, all the CYP450 inhibition.
[08:05] This can be a model on Databricks itself. It can be an external API. So, once you have all these building blocks, right? You just have your agent orchestrate, you know, the various way
[08:17] of calling the tools, and you can build something like this. So, okay. So, next, let's go back to our slides.
[08:33] So, next, I want to convince you why you should build your agents on Databricks, right? You already have your data and models there on Databricks. You should use them to make our agents smarter. Next, we support any LLM on with any
[08:45] agent framework on any cloud. Um and also, there's Unity AI Gateway to secure all your MCPs, the LLMs, the agents, um and the AI
[09:00] traffic. We've got rules, budget policies, and so on. And finally, it's really easy. We have all the following coding options, including no-code options. And um and if you're a
[09:14] die-hard, you know, IDE or cloud code user, that's also a possibility because there's the Python SDK, we have our APIs, we have the Databricks MCP, as
[09:26] well as the um Databricks CLI. So, you could code locally and then push it to the Databricks workspace from your laptop. So, later, we are actually going to build an agent together using the second
[09:40] option. We're going to build it on our playground. It's going to generate code. And we're going to deploy on Databricks apps. So, this is what we're going to do in the next 5 minutes. So, let's let's try it out.
[09:55] >>
[10:07] interested in claims or real-world data, I'm going to do claims data set. That's already a genie space. So, you can do text to sequel of the claims data set. And I'm going to connect our partner MCPs, Artifice, as well as Kythera.
[10:28] And I'm going to copy up my prompts first. Go back to playground. Okay. And then I'm going to give it a system prompt, which are routing instructions so that it knows which MCP to call for what kind
[10:42] of use cases. Save it. And I'm going to start with our first query.
[10:55] So, I'm going to ask from my own claims data set, what's the percentage of diabetic patients are on GLP-1 drugs? Um this should take about 30 seconds. And in first couple of seconds, you already see it hitting our genie uh our
[11:09] claims genie data set. So, what it's doing is generating the sequel code, and then it's going to execute that sequel to return us our answers.
[11:23] So, um yep, you can see the the sequel being generated. And then it's executing it. It's Give it a few more seconds to summarize the results.
[11:35] Should come up fast. Okay, there we go. We have about 50% of this cohort being on GLP-1 drug. Now, that's a really high number. Imagine 50% of your patients
[11:48] being on GLP-1. Right? So, I'm going to ask I'm going to check real-world evidence, right? To see, you know, what what's the prevalence in other studies? So, I'm going to run this other query.
[12:01] What does the evidence say about the prevalence of GLP-1 users among type 2 diabetic patients? And you can see it's using the Ask Evidence tool from Artichoke
[12:14] MCP. Um this query is probably going to take about a minute to run. So, while that's running, let's review the SQL from the previous call. So, I can go to view trace.
[12:28] And I go to details. You can see the timeline, the the time that each step took. Uh I know the SQL query is in this poll response uh step. And then you can see the SQL
[12:40] query here and check to see if it makes sense. So, everything is transparent. Everything is traced. Um So, go back to our playground.
[12:52] Uh next, I'm going to deploy it to our agent uh Databricks apps. Um I'm going to give it a name, the agent life. And then I'm going to save the
[13:05] responses, the traces to this MLflow experiment called life agent. And then I'm going to export it. So, what it's doing now is generate the code, the
[13:17] agent code. And let's take a look at the code. So, it has generated the agent code here. You can actually
[13:29] open it in workspace. Um it's in the agent server. There's a agent.py. And you can actually
[13:42] edit it as you like, change the problem, whatever. Um you can also use Genie code to vibe code and edit it. Um going back to our apps page, right? So, once you have the code, it's going to deploy into
[13:54] an apps container. So, it's takes taking a couple of minutes to spin up the container and so on. So, let's let that uh process run and go back to our playground. We should now have our answers back from
[14:07] the Ask Evidence tool. And let's see what it returned. So, it found that from other studies, right? And it cited the, you know, PubMed ID number and so on
[14:19] with and you can actually cross-link to it and check out that study yourself. Uh which reported that the prevalence is actually a lot lower, like 1 to 13%, which actually makes more sense, right?
[14:31] And there's also a report generated by Arcturus itself. Um
[14:49] So, Okay, I should have signed up ahead of time, but it should take me there. So, here, right? It found a study on what's the prevalence or proportion of patients with GLP-1. Um it cited its own
[15:02] study as well as some from PubMed. And that's where it pulled its numbers from. It's not making it up, right? But you saw what our agent found. Okay. So, now I'm going to ask the third query.
[15:20] The last query. Here, I want to do a breakdown of the different GLP-1 drugs, right? So, I need the NDC codes, right? And what it's going to do is call our other partner, MCP Kythera,
[15:34] to actually find these NDC codes. So, there's no need to actually download, you know, all the different ontologies, look up tables, and then they go stale, and so on. Now, you can call an ontology
[15:48] like Kythera, right? And it's going to find all the different GLP-1 drugs, um get their RxNorm code, and then try to map them to the various NDC codes of
[16:00] the different formulations. So, you saw that. And then once you have these NDC codes, you can feed it to Genie, and then get Genie to do that breakdown for you, right? So much easier now. So, let's go back to our apps page.
[16:15] Um Okay, it's still deploying the code. It's going to take a few more minutes or so, and in the meantime, let's go back to our slides, and then we can come back to that later.
[16:28] So, if you like I- come E, you can get it from our blog. Here's a QR code, or Google I- come E Databricks, and it should take you to the blog. It has a public white link, so
[16:40] you can play with the app as I've shown you earlier. It also has the GitHub code, and we have seen our customers take it and make it, you know, build an even better version.
[16:56] It also comes with a config YAML file. So, if you have a different set of MCPs, models, skills, and so on, you can just change the config YAML file, and you pretty much have an app ready to go.
[17:13] And do check out our other pharma R&D solutions. There is I- come E, of course. It's a research agent, so it's agent-based. And then there's also a model inferencing platform. Um you can do model training, model
[17:26] fine-tuning, and so on. We call it Genesis Workbench, and it has the latest AlphaFold, ESMFold, uh as well as single-cell transcriptomics models, like the predict ADMET model that I just called earlier.
[17:39] Uh I the AI Chemist called is actually one of the QSAR models in this Genesis Workbench. Um and then thirdly, we also have this protein explorer. It's both a dashboard
[17:51] and a genie, so you can actually chat with your protein library. It also has models doing classification, annotating your your proteins, so you have richer metadata in your protein library. So,
[18:04] check these out, and these are just a few that I highlight. We actually have a lot more on our blogs and GitHub repo. So, let's go back to the app that we
[18:17] just built, and let's see how it's doing. So, I'm going to go to the app. I'm going to authorize. I use pass-through authentication, so it knows that I'm Yen logging in, and the
[18:31] uh permissions will be based on um the access controls that I have. Okay, let's see how they works. So, you just saw, right? What we just
[18:43] did was go to playground, plug in a few MCPs, and we use that get code button, right? We got the code, we deployed it to Databricks apps, and we got the following. Zero code written, we built
[18:57] an agent together with a front end in just 5 minutes. So, it's that easy. So, please go back, try it out. Um try out with the new um MCPs on our marketplace, and uh thank you for your
[19:10] time. And I hope you enjoy the rest of your summit.

Learn more about the Databricks Data and AI platform.

The information provided herein is for general informational purposes only and may not reflect the most current product capabilities or configurations.