We teamed up with onepot to co-host our fourth and most ambitious hackathon yet: Ebolathon. At muni, we focus on helping you go from in silico to pre-clinical. onepot combines a 30-billion-molecule synthesis-accessible chemical space with AI-driven automated synthesis, enabling fast, cost-efficient, on-demand production of custom compounds.
Taking designs into the lab has been central to our hackathons and muni itself, so an invitation to “come to our Ebolathon” needed more context.
Our goal was to design selective small molecule antivirals against Bundibugyo ebolavirus (BDBV) without compromising cell viability.
On a Saturday in San Francisco, a room full of scientists and AI agents designed 140 molecules. onepot's own chemistry agents and laboratory automation support reaction execution, workup, purification, and analytical quality control, connecting molecular designs to physical compounds through an integrated workflow. Eleven lab days later, all 140 were in vials and ready to ship. And twenty-two lab days of testing later, we had a double-digit hit rate. This wouldn’t have been possible without our sponsors: a huge thank you to PostHog for hosting us, and to Rowan, Boltz, Modal, OpenRouter, and Anyscale for computational resources and support throughout.
Cases of BDBV are still on the rise across the Democratic Republic of the Congo, Uganda, Kenya, and now Nigeria. As of October 27th, there have been a total of 10,091 confirmed cases and 4,891 total confirmed deaths (CDC). These numbers are likely underreported, but it is still a spreading disease without an approved therapeutic.
With onepot CORE v1, our compounds were synthesized in eleven lab days, a step that commonly takes 3–6 weeks for comparable building-block chemistry. We then tested the compounds in a pseudovirus entry assay in 293T cells for the two-dose primary screen, a counterscreen, and a dose-response on selected compounds.
Assay details #
Pseudovirus. Entry was measured with single-cycle vesicular stomatitis virus (VSV) pseudotypes. The backbone, VSV-ΔG, lacks the viral glycoprotein gene and encodes firefly luciferase, so an infected cell reports entry as luciferase activity and does not release infectious virus. Particles for the primary assay were pseudotyped with full-length, human codon-optimized Bundibugyo ebolavirus (BDBV) glycoprotein (GP) from isolate 26FHV054, collected in Bunia, Ituri, on 7 May 2026. This GP differs from the 2007 isolate NC_014373.1 at 12 amino acids (98.2% identity).
Infection. 293T cells were seeded at 3.5 × 10⁴ per well and cultured for 24 h before infection. Compound and pseudovirus were preincubated for 30 min at 37 °C and then maintained on the cells for 24 h. Infection was quantified by firefly luciferase luminescence. Neutralization is the percentage loss of luciferase signal relative to vehicle-treated infected wells, after subtraction of background from uninfected wells. Mycophenolic acid (1 and 5 µM) was the positive control.
Viability. Uninfected 293T cells were incubated with compound for 24 h, and viability was measured with a TACS XTT assay and expressed as a percentage of vehicle. Cytotoxicity of less than 25% inhibition denotes viability of at least 75% of the vehicle control.
Primary screen. 140 compounds were tested against BDBV-GP pseudovirus at 2 and 10 µM, one well per concentration, with viability measured at 10 µM.
Counterscreen. Twenty-six compounds were retested at both concentrations against three pseudotypes of the same VSV-ΔG backbone: BDBV GP (26FHV054), Ebola (Zaire) GP, and VSV-G. VSV-G is the specificity control, so inhibition of BDBV and not VSV-G indicates a selective block of Bundibugyo GP-mediated entry. Zaire GP was included as a second ebolavirus species and was not an advancement criterion.
Dose-response. Ten compounds were advanced on the basis of BDBV neutralization greater than 90%, failure to inhibit VSV-G-mediated entry, and cytotoxicity of less than 25% inhibition. Each was tested at nine concentrations in twofold dilution from 20 to 0.078 µM, in duplicate or triplicate, with viability measured over the same range. Half-maximal inhibitory concentrations (IC50) and half-maximal cytotoxic concentrations (CC50) were obtained by nonlinear regression in GraphPad Prism. Where viability remained above 50% at 20 µM, CC50 is reported as greater than 20 µM.
18.6% hit rate against Bundibugyo ebolavirus #
Of the 26 preliminary hits, the 11 that made it to the leaderboard inhibited Bundibugyo GP-mediated entry and were specific in the counterscreen. Inclusion requirements are VSV-G neutralization under 25% at 10 µM and viability of at least 75%. Compounds with a dose-response curve are ranked by lowest IC50, and compounds without curves are ranked by Bundibugyo neutralization at 10 µM.
The top three on this leaderboard are endoxifen, tamoxifen (Z isomer), and N-desmethyltamoxifen, submitted by core-or-bust. Endoxifen and tamoxifen have published Ebola GP-entry data from assays that used Zaire glycoprotein. No prior Ebola GP-entry experiment was found for N-desmethyltamoxifen, and to our knowledge this is the first Bundibugyo GP-entry experiment for all three. They prioritized the GP1–GP2 interface pocket occupied by toremifene in EBOV GP (PDB 5JQ7, ligand residue T0R) and used onepot CORE to search from four Selective Estrogen Receptor Modulator (SERM) seeds – toremifene, tamoxifen, clomiphene, raloxifene. They then used RDKit physchem filtering (MW, HBD/HBA, TPSA, rotatable bonds; soft logP for SERMs), and finally Rowan analogue docking through muni to 5JQ7 using crystallographic toremifene (5JQ7:T0R) as a template. At 10 µM, Bundibugyo neutralization was 99.4% for endoxifen, 100% for (Z)-tamoxifen, and 99.5% for N-desmethyltamoxifen.
Repurposing: can anti-itch gel be used to end the Ebolavirus outbreak? #
The more unexpected results came from Grok 4.5 using our agentic loop harness. The agent designed two selective antivirals against BDBV – benproperine and bamipine – using muni to search the onepot CORE chemical space, run computational tools, and iterate on its candidates. We didn’t expect to be able to use any of Anthropic’s or OpenAI’s models,
but we tried anyway.
After multiple iterations on muni, Grok 4.5 submitted bamipine, a marketed drug in Europe and no FDA-approved product. Bamipine was synthesized and it successfully neutralized Bundibugyo GP-mediated entry by 92.2% at 10 µM with no loss in cell viability. To our knowledge, there is no paper that tests bamipine against an ebolavirus or any other filovirus. It’s sold almost entirely as an over-the-counter topical anti-itch treatment (Soventol gel in Germany), and is primarily used for insect bites and stings, sunburns, nettle rashes, and other itchy skin reactions.
Benproperine, also submitted by Grok 4.5 using muni, is a marketed drug as a non-opioid cough suppressant. It isn’t approved in the U.S., but is sold widely in Europe, China, and other Asian markets. Vanmechelen et al. reported IC50 values in Huh-7 cells for replication of Zaire Mayinga virus, but this is the first published data on Bundibugyo GP-entry on 293T cells, with a counterscreen against Zaire and VSV-G.
Another compound, Homochlorcyclizine, was submitted by a one-person team – Zayd – who configured Hermes agents (Nous) to dock CORE molecules into the toremifene pocket of Zaire GP. This compound is an antihistamine used in Japan as an oral tablet for itching from urticaria and allergic rhinitis. Zayd kept this one as a cationic amphiphilic drug, meaning that this shape can potentially collect in the acidic compartments a virus uses to enter a cell and disrupt lysosomal function, ultimately prohibiting viral entry. It’s also why they flagged this compound as needing the VSV-G counterscreen, just to ensure that this effect was specific to the Bundibugyo glycoprotein. At 10 µM, Bundibugyo neutralization was 66.6% and VSV-G was 19.4%, with no measurable loss in cell viability. To our knowledge, there is no previous data regarding homochlorcyclizine’s function on the Bundibugyo ebolavirus.
De novo small molecules: the teams that ventured into the unknown #
Rather than repurposing already marketed drugs, the-gummy-bears and cellico-bio designed and submitted molecules that were net new.
the-gummy-bears designed a host-side entry block. At a high level, Bundibugyo ebolavirus entry works something like this:
They targeted human NPC1, the receptor required for the ebolavirus entry after cathepsin cleavage of the GP (Step 4 in Figure 2), aiming for small molecules that occupy the domain-C GP-binding site and block the GPcl-NPC1 protein–protein interaction. They used known Class A entry inhibitors and related chemotypes (adamantane–piperazine tool compound EBOV-IN-1) as seeds for searching synthesis-accessible analogues in onepot CORE. They docked each compound to full-length NPC1 with Boltz2 under rational pocket constraints (PDB 5F1B). Pose ensembles were ranked by Boltz binding confidence (BC) with secondary metrics (ipTM, ipLDDT) and a fixed-pose classical Smina rescore, along with steric clash and strain filters. The final triage combined BC and Smina physics with ADMET/drug-likeness cutoffs (MW, cLogP/cLogD proxy, HBD/HBA, TPSA, rotatable bonds) to finalize their list of candidates.
At 10 µM, their candidate neutralized Bundibugyo GP-mediated entry by 78%, with VSV-G at 5% and no loss in cell viability.
cellico-bio used muni’s harness to target the GP1/GP2 pocket on PDB 5JQ7. They pulled candidates from onepot CORE, dense-docked with Rowan/qvina2, rescored pose quality with GNINA, estimated binder likelihood with Boltz affinity, and finally filtered with Rowan ADMET (MW, cLogP, HBD/HBA, TPSA, rotors) plus hERG vs toremifene.
Stay tuned for a follow-up on how we built the agentic workflows behind this hackathon and where we’re taking them next.
The next frontier is functional #
To our knowledge, this was the first hackathon to bring human- and agent-designed molecules through synthesis and functional testing in cell-based assays.
As models get better at designing binders, there will be a shift in focus to biological function: “Is this compound doing what we want, where we want it to?” At this hackathon, that meant testing whether compounds selectively inhibited viral entry into cells.
It’s clear that this shift will increase the need for more wet lab validation — beyond binding. Functional activity, ADME, toxicity, and pharmacokinetics will each introduce new questions that require experimental validation from different labs. Agents can continue to generate hypotheses at scale, but the time it takes to test them will set the pace of discovery.
This is why we’re building muni Dispatch, which lets you connect agents and harnesses to AI-native chemistry platforms like Onepot to take candidates through synthesis and into assays, as we did for this hackathon. And with onepot’s expansion from 3.4 billion to 30 billion executable compounds, you can now run muni’s workflow with nearly 9× the chemical space.
Our goal at muni is to unlock as much scientific discovery as possible in this age of intelligence, and we’re betting that early and highly iterative discovery will increasingly go to CROs that combine rigorous science with fast turnaround, transparent services and pricing, and data that agents can use directly.
For solo drug hunters or small teams, this means paying for experimental capabilities only as you need them and not taking on the capital expenditure in-house. For AI-native chemistry platforms like onepot, this means making automated synthesis and experimental capabilities directly accessible to a growing ecosystem of discovery agents, while using experimental data to continuously improve chemistry and expand accessible chemical space.
If you’re an AI-native CRO or platform looking to have agents as customers, sign up at [https://muni.bio/dispatch/cro](https://muni.bio/dispatch/cro) or email me at [kat@muni.bio](mailto:kat@muni.bio).