Rhizome OS-1: Rhizome's Semi-Autonomous Operating System for Small Molecule Drug Discovery

Yiwen Wang; Gregory Sinenka; Xhuliano Brace

cs.AIApr 8, 2026

Rhizome OS-1: Rhizome's Semi-Autonomous Operating System for Small Molecule Drug Discovery

Authors: Yiwen Wang, Gregory Sinenka, Xhuliano Brace

Organizations: Rhizome Research Inc.

Abstract

We present Rhizome OS-1, a semi-autonomous operating system for small molecule drug discovery in which multi-modal AI agents operate as a full multidisciplinary discovery team. These agents function as computational chemists, medicinal chemists, and patent agents: they write and execute analysis code (fingerprint clustering, R-group decomposition, substructure search), visually triage molecular grids using vision capabilities, formulate explicit medicinal chemistry hypotheses across three strategy tiers, assess patent freedom-to-operate, and dynamically adapt generation strategies based on empirical screening feedback. Powered by r1 - a 246M-parameter graph diffusion model trained on 800 million molecular graphs - the system generates novel chemical matter directly on molecular graphs using fragment masking, scaffold decoration, linker design, and graph editing primitives. In two oncology campaigns (BCL6 BTB domain and EZH2 SET domain), the agent team executed 26 seeds and produced 5,231 novel molecules. Across both targets, 91.9% of generated Murcko scaffolds are absent from ChEMBL, with median Tanimoto similarity of 0.56-0.69 to the nearest known active. Boltz-2 binding affinity predictions, calibrated against ChEMBL data, achieved Spearman correlations of -0.53 to -0.64 and ROC AUC values of 0.88-0.93. These results demonstrate that semi-autonomous agent systems, equipped with graph-native generative tools and physics-informed scoring, enable a new paradigm for early-stage drug discovery: scaled, rapid, and adaptive inverse design with embedded medicinal chemistry reasoning.

Figures & tables

Figure 1: Rhizome OS-1 system architecture. PDB structures and ChEMBL activity benchmarks feed a target profiler that extracts pocket geometry, binding models, and seed compounds. An agent-orchestrated design loop cycles through four roles: a structural analyst formulates generation strategies, the r1 engine generates novel molecules on molecular graphs, a filtering pipeline removes invalid or drug-unlike outputs, and an evaluator assesses chemical plausibility and novelty. An orchestrator manages wave scheduling and maintains a convergence library of 1,000–2,000 molecules per target. Boltz-2 affinity screening scores the filtered library against calibrated benchmarks, and an optimizer analyzes strategy $\times$ seed performance to select new seeds for subsequent iterations.

Panel of Figure 2 — Figure 2: BCL6 BTB domain binding pocket. Left: PDB 7LWE (YND, pyrrolopyrimidinone, 1.17 Å resolution) showing the lateral groove with key contacts MET51, GLU115, CYS53, and TYR58. Right: PDB 7OKL (VJ5, N-methyl-2-quinolone) illustrating the alternative cap orientation toward PHE89 and VAL117 extension contacts.

Family	Seeds	ChEMBL analogs	pChEMBL range
Benzimidazolone	4	256	5.0–9.3
Pyrazoloquinazolinone	6	76	5.5–8.5
Benzoxazepinone	4	57	6.0–8.0
Macrocycles	2	$<$ 10	—
Diverse fragments	3	$<$ 10	—
Chimeric designs (cross-family, no dedicated seeds)

Table 1: BCL6 seed families and ChEMBL benchmark coverage.

Metric	Affinity score	Binding probability
Spearman $\rho$	$-0.638$	$0.206$
ROC AUC	$0.875$	$0.665$

Table 2: BCL6 Boltz-2 calibration metrics (518 benchmark compounds, active threshold pChEMBL

\geq 6.5

Figure 3: BCL6: Boltz-2 binding affinity validation. (A) Calibration scatter of Boltz-2 affinity score versus experimental pChEMBL for benchmark compounds. (B) ROC curves for active/inactive discrimination using affinity score and binding probability. (C) Distribution of affinity scores across generated molecules, with benchmark active quartiles shown as dashed red lines.

Figure 4: Seed-versus-generated comparison for BCL6 (9 molecules from 8 chemotype families, including two from the div family). Each row shows the seed structure (left) alongside the generated molecule (right), with Boltz-2 affinity scores. Molecules were selected for structural novelty relative to their seeds; the two div_BM_5 entries derive from the same graph-edited seed but produce distinct scaffolds.

#	ID	Aff.	Sim.	MW	Family
1	SP1_YJJ_oxaazaspiro_fm_0599	$-3.32$	0.71	547	SP
2	TR1_YJJ_piperazine_fm_0648	$-3.28$	0.72	530	TR
3	div_BM_5_sd6_1782	$-2.76$	0.43	427	div
4	ben_BM_1_ge_sd6_1122	$-2.61$	0.64	544	ben
5	LR6_BM5_urea_fm_0306	$-2.57$	0.55	496	LR
6	div_BM_5_sd6_1773	$-2.46$	0.44	435	div

Table 3: Selected BCL6 portfolio (9 molecules from 8 families, visually inspected).

Panel of Figure 5 — Figure 5: EZH2 SET domain binding pocket. Left: PDB 5LS6 showing the SAM-competitive binding site with the pyridone warhead recognition pocket and the solvent-exposed arm extension region. Right: overlay of four co-crystal ligands (PDBs 4W2R, 5IJ7, 5LS6, 6B3W) illustrating the conserved warhead anchor position and divergent arm geometries across inhibitor families.

Seed	Family	pChEMBL	Shape class
1	Indole-2-carboxamide	10.24	Linear rod (flat)
2	Benzodioxole	10.42	Linear rod (sp 3 )
3	Compact diazepanone	9.05	Compact U-shape
4	BCP linker	9.30	Rigid BCP
5	Aniline trisubstituted	9.49	T-shape
6	Pyrrolopyridine	11.0	L-branch

Table 4: EZH2 seed compounds. pChEMBL values refer to the most potent reported assay result for each chemotype.

Metric	Value
Benchmark compounds	1,098
Spearman $\rho$	$-0.529$
ROC AUC (affinity score)	0.927
ROC AUC (binding prob.)	0.812

Table 5: Boltz-2 calibration metrics for EZH2 (5LS6 construct, 695 aa). Active threshold: pChEMBL

\geq 6.5

Figure 6: EZH2: Boltz-2 binding affinity validation. (A) Calibration scatter of Boltz-2 affinity score versus experimental pChEMBL for benchmark compounds. (B) ROC curves for active/inactive discrimination using affinity score and binding probability. (C) Distribution of affinity scores across generated molecules, with benchmark active quartiles shown as dashed red lines.

Figure 7: Seed-versus-generated comparison for EZH2 (7 molecules across 7 shape classes). Each row shows the seed structure (left) alongside the generated molecule (right), with Boltz-2 affinity scores. T-shape anilines ( $>$ 650 Da), strained bicyclic caps, and tropane systems were excluded during visual review.

#	ID	Aff.	Sim.	MW	Shape class
1	1b_4azaindole_…_mol16	$-3.61$	0.54	584	linear rod (flat)
2	chimera_B_bcp_…_mol828	$-3.48$	0.66	567	mixed chimera
3	7d_cyclopentane_…_mol1628	$-3.31$	0.57	586	3D-complex
4	2c_cyclopentane_…_mol1428	$-3.07$	0.59	520	linear rod (sp 3 )
5	6e_pyrrole_…_mol804	$-3.05$	0.55	545	L-branch
6	3c_piperazinone_…_mol353	$-3.04$	0.52	496	compact U-shape

Table 6: Selected EZH2 portfolio (7 molecules: 6 seed shape classes plus mixed chimera, visually inspected).

Target	Generated	Scaffolds	Novel scaffolds	Novelty (%)	Med. Tanimoto
BCL6	2,395	386	360	93.3	0.563
EZH2	2,130	171	152	88.9	0.686
Total	4,525	557	512	91.9	—

Table 7: Structural novelty metrics for generated libraries. Tanimoto similarity is reported as the median maximum similarity to the nearest ChEMBL benchmark compound (lower = more novel). Scaffold novelty is the percentage of generic Murcko scaffolds in the generated set that are absent from ChEMBL for that target.

Target	Seeds	Molecules	Calib. $n$	Spearman $\rho$	ROC AUC	Screened
BCL6	19	2,876	518	$-0.638$	0.875	2,235
EZH2	7	2,355	1,098	$-0.529$	0.927	2,102
Total	26	5,231				4,337

Table 8: Cross-target campaign summary.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 8: BCL6 seed compounds. Nineteen seeds drawn from five chemotype families (benzimidazolone, pyrazoloquinazolinone, benzoxazepinone, macrocycles, diverse fragments); chimeric designs combined motifs across families were used to initiate the BCL6 generation campaign.

Figure 9: EZH2 seed compounds. Seven seeds spanning seven shape classes (linear rod flat, linear rod sp 3 , compact U-shape, rigid BCP, T-shape, L-branch, 3D-complex) were used to initiate the EZH2 generation campaign.

Campaign	Scored	Pass hard	% soft flags	Most common flag	Clean
BCL6	2,235	1,977	49.7	high_mw (265)	994
EZH2	2,102	2,080	82.0	high_mw (269)	375

Appendix

Table 9: Structural alert filtering summary across all campaigns.

Target	ID	Pass	Confidence	Notes
BCL6	SP1_fm_0599	✓	medium	Novel; not in patent databases
BCL6	TR1_fm_0648	✓	medium	Novel; not in patent databases
BCL6	div_BM_5_sd6_1782	✓	medium	Novel; not in patent databases
BCL6	ben_sd6_1122	✓	medium	Novel; not in patent databases
BCL6	LR6_fm_0306	✓	medium	Novel; not in patent databases
BCL6	div_BM_5_sd6_1773	✓	medium	Novel; not in patent databases

Appendix

Table 10: Patentability screening results for selected portfolio molecules. Confidence reflects the automated assessment certainty; formal freedom-to-operate analysis remains a prerequisite for progression.

Target	ID	Canonical SMILES
BCL6 (9 molecules)
	SP1_fm_0599	CCC(C)(C)CNc1cc(NC(=O)Cn2cc(-c3cc(C)c(O)c(C#N)c3)
		c3c(=O)n(C)cnc32)c(Cl)cn1
	TR1_fm_0648	Cc1cc(-c2cn(CC(=O)Nc3cc(C4CCCCC4)ncc3Cl)c3ncn(C)
		c(=O)c23)cc(C#N)c1O
	div_BM_5_1782	Nc1cc(Cl)c(Nc2cc(C[C@@]34CC[C@@H]3C4)c3[nH]c(=O)

Appendix

Table 11: Canonical SMILES for selected portfolio molecules.

Drug Discovery

Drug discovery identifies, designs, and tests molecules that can safely affect disease biology.

See all 164 papers

MADDIE: Multi-Agent Drug Discovery Engine
Oct 12, 2026Ye Liu, Reza Averly, Avery Meyer +4Drug Discovery Multi-Agent LLM Systems
Biomedical Informatics, The Ohio State University. · Computer Science and Engineering, The Ohio State University. · Medicinal Chemistry and Pharmacognosy, The Ohio State University. +1
Mozi: Governed Autonomy for Drug Discovery LLM Agents
Mar 4, 2026He Cao, Siyu Liu, Fan Zhang +7AI Agents for Scientific Discovery LLM Agent Orchestration
M3OS: A Monte Carlo Graph Search-Orchestrated Multi-Agent LLM System for Evidence-Traced Molecular Optimization
Sep 28, 2026Junjie Wang, Yaowei Jin, Ruohui Tang +7Molecular Optimization Multi-Agent LLM Systems
Lingang Laboratory, Shanghai, 200031, China. · School of Information Science and Technology, ShanghaiTech University, Shanghai, 201210, China. · Global Institute of Future Technology, Shanghai Jiao Tong University, Shanghai, 200240, China.

Oct 12, 2026J/K move · Enter open · S save

MADDIE: Multi-Agent Drug Discovery Engine

Ye Liu, Reza Averly, Avery Meyer, Frazier N. Baker+3

Biomedical Informatics, The Ohio State University. · Computer Science and Engineering, The Ohio State University. · Medicinal Chemistry and Pharmacognosy, The Ohio State University.+1

Drug Discovery Multi-Agent LLM Systems LLM Agent Orchestration

Open paper AskPDF arXiv

Drug discovery remains a resource-intensive and time-consuming process, involving billions of dollars and a decade of development to advance a drug to market. While many in silico tools have accelerated this process, operating them still demands substantial time, effort, and expertise from medicinal chemists. Autonomous systems built on Large Language Models (LLMs) have emerged as a promising solution, yet current designs limit their performance. Existing single-agent systems scale poorly to long-horizon, multi-step, and realistic tasks, whereas multi-agent systems typically execute predefined workflows that cannot capture the iterative, non-linear nature of a real drug discovery campaign. Here, we present MADDIE, a multi-agent drug discovery engine. It comprises a central agent, Orchestrator, and four specialized worker agents: HitIdentifier, LeadOptimizer, Synthesizer, and Evaluator, each representing a critical step in a drug discovery campaign. Under Orchestrator's orchestration and coordination, these worker agents are composed into flexible workflows, enabling MADDIE to adapt its strategy and navigate the drug discovery pipeline end-to-end. Across 50 protein targets, MADDIE produces molecules satisfying key property requirements and their synthesis routes, outperforming existing agentic systems, generative models, and general-purpose LLMs by a wide margin. Our analysis reveals how these agents collaborate and that MADDIE can employ different strategies across targets to achieve the campaign's goal, while still being efficient in time and tokens. Together, they demonstrate MADDIE as a low-cost, effective, and efficient autonomous engine for accelerating drug discovery. Code and dataset is available via https://github.com/ninglab/MADDIE.

Abstract

Drug Discovery

Explore similar work

MADDIE: Multi-Agent Drug Discovery Engine

Mozi: Governed Autonomy for Drug Discovery LLM Agents

M3OS: A Monte Carlo Graph Search-Orchestrated Multi-Agent LLM System for Evidence-Traced Molecular Optimization