Supercharging Rowan’s tools with AI agents
We’re working with Rowan to make computational chemistry easier.
Computational chemistry is harder than it needs to be. There is a lot of invisible work that goes into maintaining and running software that computational scientists need to use. The development of tools in computational chemistry remains decentralized and fragmented. Individual academic groups write—or rewrite, as the case may be—tools to run, parse, and organize different types of atomistic simulations, cheminformatics data, or machine learning (ML) model architectures. Often, the lifecycles for these specialized pieces of software begin and end with the academic stays of the PhD students and postdocs who write them.
The bottleneck is not just software complexity. In many computational chemistry workflows, expert time is the limiting factor: experienced practitioners have to decide how to set calculations up, monitor failures, interpret outputs, and iterate on the next step. That makes scaling difficult, because throughput depends not only on available compute but also on the availability of skilled specialists.
Foundational large language models (LLMs) offer a new solution to this issue: In the right harness, LLMs can easily digest documentation and code for valuable tools and translate requests to viable inputs, so users can quickly run what they need and focus more on science than on software. Axon is the right harness, purpose-built for drug discovery, equipped with the tools, infrastructure and intelligence to accelerate and multiply the value of computational workloads while still showing every step. By allowing more of the computational chemistry workflow to be executed and iterated through an LLM with structured tooling, Axon helps scale the science more directly with compute rather than just experts’ time. That lowers the barrier to deploying these methods in real research settings and reduces the transaction costs of early-stage scientific exploration. Axon handles this work while preserving auditability so that experts can still inspect and verify its approaches and conclusions.
We’re fortunate to share the goal of facilitating computational chemistry with the folks at Rowan. They’ve made using existing tools—like those for quantum chemistry and protein simulations—much easier to use, developed their own models for chemistry (like Egret-1 and Starling), and provided useful datasets and benchmarks for ML interatomic potential (MLIP) models (namely Wiggle-150 and NNP Arena). Most recently, they’ve released a free energy perturbation (FEP) tool for easier and more accurate simulations of ligand-protein interactions (we plan to discuss this new tool later in a dedicated post). Rowan’s workflows include a lot of infrastructure that organizes and analyzes the outputs of these calculations so that users don’t have to worry about writing their own code and can instead focus on using computational tools for scientific discovery.
Given the alignment in our missions, and the incredible quality of Rowan’s platform, we were excited to build native integrations with Axon, so scientists can access Rowan calculations with the ease and intelligent direction of AI agents. Axon takes scientific questions, high-level objectives or computational directives, and helps users submit, monitor, and analyze calculations to accomplish the goal. In this way, Axon complements Rowan’s abilities by making them more accessible to more users and reducing the amount of expert orchestration required to use them productively at scale.
In this blog, we’ll highlight a couple of examples of leveraging Rowan through Axon to show the ease and value the combination offers.
Specialized molecular structure generation
Rowan has several tools for generating structures, but Axon users might be particularly interested in their conformer search tools. Rowan can do conformer searches both with and without solvent and rank conformer outputs using a plethora of semi-empirical, density functional theory (DFT), and MLIP methods. They’ve provided access to new methods that improve on other common approaches, such as Monte Carlo Multiple Minimum (MCMM) and a generative method based on flow matching called Lyrebird. We’re excited about the ways that some of these newer methods can be used with larger molecules that are difficult to study with older approaches but that are still relevant to computational drug discovery, such as macrocyclic peptides.
Calculating molecular properties
Redox potential
Another important consideration for molecules is how easily they can accept or donate electrons, as measured by the molecule’s redox potential. Rowan has incorporated a redox potential workflow to calculate these values and to do it quickly with relatively good accuracy.
One example is metronidazole, a prodrug that relies on anaerobic respiration mechanisms in certain microbial species to produce a radical that damages their DNA. We asked Axon to use Rowan’s redox potential workflow to estimate how some modifications to metronidazole change its redox potential. We asked it to produce a report comparing the outputs of the simulations and a figure showing the molecules and the differences in redox potential relative to metronidazole, which you can see below. Here is the full example of Axon testing these modifications if you’re curious to see how Axon works with these tools.
This example illustrates how Axon is designed to be auditable. The value is not that users have to trust an LLM’s intuition about chemistry; it is that Axon can call a real computational workflow, organize the results, and let users verify the conclusion directly. In this case, the predicted redox potentials can be compared explicitly across the tested analogs, so the output is checkable rather than merely persuasive.
Macroscopic pKa
Chemists typically determine a molecule’s acidity using pKa, which measures the equilibrium between protons that can move between the solution and attachment to a molecule. Rowan provides two workflows to estimate pKa: macroscopic pKa, which describes the overall protonation equilibrium, and microscopic pKa that estimates the pKa values of individual atoms that can be (de)protonated. Their macroscopic pKa approach uses their Starling model, a physics-informed neural network based on the Uni-pKa model. These pKa values can be important to know during drug development because they can affect a molecule’s membrane permeability, solubility, and ionization state in different physiological environments. Axon can now use both macroscopic and microscopic pKa calculations and help users analyze their outputs.
Identifying atomic properties within a molecule
Rowan has several tools on their platform for calculating properties of particular atoms or moieties within molecules. If you’re using Axon for drug discovery, you might be particularly interested in identifying parts of a molecule that are most likely to react or that interact strongly with other particular species. Rowan has implemented a handful of particularly useful workflows to identify just these things, which we describe a bit more in the subsections below.
Reactivity metrics: Fukui indices and bond-dissociation energy
One particularly useful workflow that Rowan has implemented calculates Fukui indices for atoms within a molecule. Briefly, Fukui indices are calculated by adding or removing an electron to a molecule and calculating how its electron distribution changes. These Fukui indices can indicate both which atoms are reactive and how they might react—either with nucleophiles or electrophiles. This information is useful for medicinal chemists trying to understand how a drug molecule might be metabolized, among others. Before Rowan introduced this workflow, it might take a computational chemist significantly more time to set up, run, and analyze the calculations required to determine these values—or the use of some bespoke internal code; Rowan’s workflow makes it possible to do in much less time. Axon can now run this workflow to calculate Fukui indices, identify reactive atoms in a molecule, and analyze its outputs.
Rowan also has workflows available to calculate bond dissociation energies (BDEs)—both “traditional” approaches using quantum chemistry or semi-empirical methods and newer machine-learned interatomic potential approaches, which they’ve carefully benchmarked. Calculating BDEs for key bonds is another way to estimate how easily a reaction can occur for an atom in a molecule. Repeating these calculations for a series of bonds in one molecule can be cumbersome, but Rowan’s tools make it relatively easy to do and handle the multitude of calculations for their users.
Let’s see how this workflow looks in Axon. The drug diclofenac has a chlorinated aniline ring with C-H bonds at the 3’, 4’, and 5’ positions. We asked Axon to fluorinate this ring in diclofenac and use Rowan’s tools to test how that adjustment changes BDEs for nearby H atoms, particularly the more reactive C-H bond at the 4’ position. Axon produced an annotated diagram showing the modifications that it tested and how they changed the BDEs of the C-H bonds, which you can see below. You can also read the full transcript of this Axon session.
Microscopic pKa and H-bond basicity
We introduced pKa above when introducing Rowan’s macroscopic pKa calculations and want to dedicate some space to discussing their microscopic pKa approach, too. Microscopic pKa estimates the pKa values of individual atoms that can be (de)protonated. These values might be particularly useful for understanding the likelihood of particular microstates (combinations of protonated and deprotonated atoms within a molecule) that can help understand how a molecule might look under different conditions.
We asked Axon to test how modifications to the drug chloroquine change the pKa values for atoms with each modification using Rowan’s microscopic pKa workflow. Chloroquine is a common anti-malarial drug with multiple basic N atoms, and in particular a tertiary amine near the end of a side chain. We provided a list of proposed modifications and asked Axon to run and analyze the calculations, the output of which you can see below. Some of the proposed changes seem to substantially reduce the pKa of the target atom, in line with the request we submitted to Axon. The full transcript for this example is available here.
Another important factor to consider in drug discovery is the ability for certain atoms to form H-bonds with polar species. Predicting how strong an H-bond is without explicitly calculating interactions between two molecules can be a little bit tricky, but fortunately, Rowan has introduced their own method for quickly estimating H-bond basicity with their pKBHX workflow. Their approach identifies the areas with the minimum electrostatic potential (Vmin) wherever there are lone electron pairs in a molecule. Most conveniently, they’ve set up their workflow to automatically handle all of the steps for identifying these Vmin from a single DFT calculation. Rowan highlighted an interesting example from medicinal chemistry on their blog introducing this workflow that illustrated how modulating H-bond acceptor strength (as measured by pKBHX) can affect things like membrane permeability and solubility.
🪩 ➕ 🌳 🟰 💊
The folks who started Rowan have been explicit about their goal to make it easier to run high-quality computational chemistry calculations across different disciplines from their beginning. We think this mission is critical, particularly as computational tools become faster and more accurate and ever more essential in chemical and materials design. Using Axon not only makes computational chemistry work more convenient, but removes major bottlenecks for experts to manage these workflows. Axon makes Rowan’s curated capabilities available with the expert guidance of state-of-the-art AI, so scientists can focus more on novel hypotheses and critical discoveries. We hope that their availability through Axon further streamlines computational chemistry and makes Rowan’s incredible work accessible to even more people!
We’re excited to partner with Rowan to bring these tools to our users, and we’re excited about the future of computational chemistry. If you’re interested in leveraging Rowan’s tools with Axon, give it a try! If you have other questions or comments, I’d love to hear from you at alex@mirrorphysics.com.