Use CasesEnterpriseResearchBlog

Company

Pricing
Biomni Labarrow

Unlocking small molecule discovery with chemistry capabilities

By Matteo Aldeghi

Nov 15, 2026 · 6 min read

Biomni Lab now has a chemical structure editor, molecular design tools on the Biomni HPC, and expert chemistry skills.

Most of what Biomni does today sits on the biology side of a drug program: reading the literature, analyzing the data, putting together the evidence for and against a target. That work is crucial, but it stops well short of a molecule. An expert still has to find chemical matter, work out what it does, and decide what to make next.

We're excited to share that Biomni can now help computational and medicinal chemists make chemistry decisions. Our inaugural set of chemistry capabilities in Biomni Lab is aimed at small molecule discovery: a chemical structure editor, molecular design tools that run on the Biomni HPC, and skills that tell the agent how an experienced chemist would use those tools.

Where small molecule programs get stuck

Like biology, chemistry increasingly relies on computational tools and data science to make sense of patterns hidden across datasets, predict properties of interest, and design experiments. Today, researchers have access to a growing list of databases, methods, and tools for designing molecules, planning synthetic routes, and understanding how matter behaves at the atomic level. But as the tool list expands, keeping up with the field, understanding the tools' capabilities, and knowing how to apply them has become a challenge in itself.

This is particularly relevant to small molecule discovery as lead discovery programs now have more data and computational tools available than ever, thanks to automation, AI, and a growing network of CROs. But this does not automatically translate into faster decisions.

Making sense of the evidence requires human chemists who can connect what is known about a target, the compounds tested against it, and the assay results to determine what to make and test next. The supply of these specialized human experts has not kept pace with the growing volume of data and tools, making human expertise a critical bottleneck in the discovery process.

As a result, important decisions are often delayed while teams wait for the right expertise, or made with incomplete evidence, which creates longer cycle times between rounds and a lower chance the program succeeds.

Biomni now helps computational and medicinal chemists discover promising compounds, faster.

The layers supporting chemistry work

1. Chemical structure editor

AI chat interfaces are designed for conversations, but they are a poor way to look at a molecule. A SMILES string tells you almost nothing at a glance, and a hundred of them tell you less.

Biomni now renders molecular tables, with structures drawn inline beside their data, and grids you can scan by eye. A chemical structure editor lets you draw a molecule, or edit one the agent returned, and give it to Biomni as input.

2. Molecular design tools, ready to use

A compound library in Biomni, with every structure rendered inline beside its name and ID.

There are now several powerful physics-based and AI tools for molecular design, from binding free energy calculations to co-folding and generative chemistry models.

Using them requires time and effort wrangling software dependencies, setting up input files, and validating and interpreting the outputs.

Biomni comes with these tools out of the box, and the agent runs them automatically, with or without prompting.

Biomni agents can predict protein-ligand structures with OpenFold3, calculate binding affinities with OpenFreeEnergy, propose analogs with REINVENT4, and build predictive deep learning models with Chemprop. All of them run on the Biomni HPC, and additional tools will be added to the HPC over the coming weeks.

3. Skills: how a chemist would use them

While tools tell the agent what it can run, Skills tell it how an expert would: when a method applies and when it tends to be unreliable, how to interpret outputs and predictions, and how to reason the way a chemist in pharma R&D would.

Skills hold the knowledge scientists build up over years of practice and rarely write down, and Biomni uses them to carry out longer workflows accurately on its own.

We've built skills for chemical data analysis and molecular modeling, so Biomni can work as a peer to cheminformaticians and CADD specialists, and as a computational partner to medicinal chemists.

Biomni running the R-group decomposition skill on a 130-compound EGFR series, with the resolved scaffold and an SAR explorer across R1 and R2.
The decomposition it returned, with every structure, scaffold and R-group drawn inline beside the measured activity.

The new small molecule discovery workflow

Supporting small molecule discovery end to end takes all three layers. Here's the kind of workflow we're building toward.

In hit-to-lead, a chemist can ask Biomni to cluster high-throughput screening hits into candidate lead series, weighing potency, synthetic tractability, and expected bioavailability, and to build a preliminary SAR. After reviewing the series Biomni ranks, they can prompt the agent to support hit expansion by running binding free energy calculations to prioritize higher-potency candidates, whether Biomni proposed them or the chemist added them.

We've run that sequence end to end on 803 public COVID Moonshot compounds: fingerprint clustering into three candidate series flagged for potency, synthetic accessibility and oral risk, a preliminary SAR across 18 traceable comparisons, and a structure catalogue for the series the chemist picked.

Markush core and 50-molecule catalogue for the S02 series, from the hit-to-lead run on public COVID Moonshot data.

In lead optimization, a chemist can ask Biomni to analyze a lead series and find the vectors that could drive selectivity against a homolog. Starting from what's known about the target, the homolog, and the ADME/PK data, Biomni builds SAR and matched-pair views, correlates physicochemical properties with the bioavailability parameters to preserve, and surfaces activity cliffs. It returns an evidence-backed hypothesis about where to modify the molecule.

After confirming the vector and design constraints, the chemist can ask Biomni for design ideas. Biomni chooses REINVENT for the job, generates candidates against those constraints, and filters out molecules with structural alerts. Where the measured data supports a selectivity model, Biomni trains one and tests it against a held-out set before trusting it to rank. The chemist reviews the shortlist and adds, removes, or edits candidates as they see fit.

The new way chemists work

The interface, the HPC tools, and the first skills are the foundation. Next we're adding tools for molecular modeling, structure-based drug design, and synthetic chemistry, including protein dynamics simulation, machine-learned potentials, quantum chemistry packages, and synthesis planning. The tables, grids, and editor will get more interactive too: you'll be able to ask Biomni questions straight from a table or grid, and the agent will be able to work in the chemical structure editor itself.

Try it on your own data in Biomni Lab at biomni.phylo.bio. If you run a discovery program and want to learn what it would look like in your environment, reach us at contact@phylo.bio.