Designing Drugs with AI: Rhizome OS-1 Expands Chemical Space

Author: Denis Avetisyan

A new AI-powered platform is automating the design of novel small molecules, offering a promising approach to accelerate drug discovery and enhance chemical diversity.

Rhizome OS-1 employs an iterative design loop-cycling through structural analysis, molecular generation via the r1 engine, filtering, and evaluation-to discover novel compounds, leveraging [latex]1,000-2,000[/latex] molecules per target in a convergence library and scoring affinity with Boltz-2 against calibrated benchmarks, all orchestrated to optimize seed selection and refine generation strategies.

Rhizome OS-1 leverages multi-modal agents and a foundation model for autonomous molecular generation and optimization, demonstrating efficacy in targets like BCL6 and EZH2.

Early-stage drug discovery remains a protracted and often serendipitous process despite advances in computational chemistry. Here, we present ‘Rhizome OS-1: Rhizome’s Semi-Autonomous Operating System for Small Molecule Drug Discovery’, a platform leveraging multi-modal AI agents and a [latex]246M[/latex]-parameter graph neural network to autonomously design and generate novel chemical matter with demonstrated success in oncology targets like BCL6 and EZH2. Notably, the system yields molecules with high scaffold diversity-[latex]91.9\%[/latex] absent from ChEMBL for respective targets-and competitive binding affinity predictions, achieving ROC AUC values of [latex]0.88[/latex] to [latex]0.93[/latex]. Could this semi-autonomous approach redefine the paradigm for rapid, adaptive inverse design and accelerate the identification of promising drug candidates?

The Enduring Challenge of Epigenetic Targeting

Despite decades of research and innovative treatment strategies, a truly effective and universally applicable cure for cancer continues to be a significant challenge, especially when the disease’s origins lie within epigenetic alterations. Unlike genetic mutations which change the DNA sequence itself, epigenetic dysregulation modifies how genes are expressed – essentially controlling which genes are ‘switched on’ or ‘switched off’ – without altering the underlying code. This presents a unique therapeutic hurdle, as these changes are often reversible and can be influenced by environmental factors, making them difficult to target specifically. Moreover, epigenetic mechanisms are crucial for normal cellular function; disrupting these processes systemically can lead to severe side effects, demanding a level of precision that current therapies often struggle to achieve. Consequently, cancers driven by epigenetic errors-a growing area of understanding-often prove resistant to conventional treatments and necessitate the development of novel, highly targeted approaches.

The Polycomb Repressive Complex 2 (PRC2) plays a critical, and often causative, role in numerous cancers by silencing tumor suppressor genes through epigenetic modifications. Central to PRC2’s function is the enzyme Enhancer of Zeste Homolog 2 (EZH2), which catalyzes the addition of methyl groups to histone proteins, effectively compacting DNA and reducing gene expression. While EZH2 inhibition has emerged as a promising therapeutic strategy, its success is hampered by the enzyme’s vital role in normal cellular processes, particularly in maintaining cell identity and preventing developmental abnormalities. Achieving selective inhibition – disrupting PRC2 activity specifically within cancer cells while sparing healthy tissues – remains a significant hurdle; current inhibitors often exhibit limited specificity, leading to potential off-target effects and systemic toxicity. Researchers are actively exploring strategies to overcome this challenge, including the development of more refined inhibitors, targeted delivery systems, and combination therapies that exploit unique vulnerabilities within cancer cells to enhance the therapeutic window.

A significant hurdle in epigenetic cancer therapy lies in the difficulty of selectively targeting disease-specific epigenetic modifications. Current strategies, while showing promise in preclinical models, often struggle to differentiate between aberrant epigenetic patterns in cancerous cells and those essential for normal cellular function. This lack of precision results in widespread disruption of the epigenome, leading to significant toxicity and limiting the therapeutic window. Healthy cells, reliant on precise epigenetic regulation for development and maintenance, are consequently affected, causing adverse side effects that can outweigh potential benefits. Researchers are actively exploring methods to enhance specificity, including the development of targeted delivery systems and the identification of unique epigenetic vulnerabilities present only in cancer cells, with the aim of minimizing off-target effects and maximizing therapeutic efficacy.

The EZH2 SET domain binding pocket exhibits a conserved warhead anchor position for inhibitors, but displays divergent arm geometries as illustrated by the overlay of ligands from PDB structures 4W2R, 5IJ7, 5LS6, and 6B3W.

AI-Driven Molecular Generation: A Paradigm Shift

Rhizome OS-1 constitutes a new operating system designed to accelerate drug discovery through the integration of multiple artificial intelligence agents and a large-scale foundation model. This system departs from traditional methods by employing a software architecture that allows for automated experimental design, data analysis, and iterative molecule generation. The core of Rhizome OS-1 is its ability to coordinate these AI agents-each specialized in tasks such as retrosynthesis, property prediction, and synthetic accessibility scoring-around the [latex]r1[/latex] foundation model. This model serves as a central knowledge base and predictive engine, enabling the system to explore chemical space and propose novel molecular structures with optimized characteristics. The OS aims to streamline the entire drug discovery pipeline, reducing the time and resources required to identify promising drug candidates.

Rhizome OS-1 utilizes a tiered approach to molecule generation, systematically varying the degree of chemical alteration. ‘Tier 1 Modifications’ represent conservative changes, focusing on minimal structural adjustments to existing molecules, while ‘Tier 2 Modifications’ introduce more substantial alterations, exploring a wider range of chemical possibilities. ‘Tier 3 Modifications’ represent the most radical generation strategy, enabling the system to explore significantly different chemical structures and potentially novel molecular scaffolds. This tiered system allows for a controlled exploration of chemical space, balancing the need for generating diverse molecules with the desire to maintain structural plausibility and synthetic accessibility.

The r1 Foundation Model serves as the core engine for molecule generation within the Rhizome OS-1 platform. This model utilizes a graph diffusion approach, enabling it to learn and extrapolate from the complex relationships within molecular structures. Training was conducted on a large-scale dataset comprised of molecular graphs, allowing the model to develop a robust understanding of chemical validity and diversity. Consequently, the r1 Foundation Model has successfully generated over 5,200 novel molecules, initiating these generations from a set of 26 distinct molecular ‘seed’ structures.

Generated molecules, compared to seed structures across seven shape classes for EZH2, demonstrate comparable binding affinity-as measured by Boltz-2 scores-with visual review excluding large T-shape anilines, strained bicyclic caps, and tropane systems.

Orchestrating the AI Workflow: Precision and Refinement

Rhizome OS-1 utilizes a suite of specialized AI agents to compartmentalize the molecular design process. The ‘Structural Analyst Agent’ is responsible for assessing the physical characteristics of proposed molecules, while the ‘Generator Agent’ focuses on creating novel molecular structures based on defined parameters. Crucially, the ‘Evaluator Agent’ independently verifies the quality and viability of generated molecules against pre-set criteria, ensuring that only promising candidates proceed to subsequent stages. This modular approach allows for parallel processing and targeted refinement, enhancing the overall efficiency of the design workflow.

The Optimizer Agent within Rhizome OS-1 functions as a feedback mechanism to enhance molecule generation. It operates by continuously analyzing the results of molecular screening performed by other agents, specifically identifying high-performing compounds designated as ‘Seed Compounds’. These Seed Compounds are not simply retained; they are fed back into the generation process as starting points for subsequent iterations. This cyclical refinement process prioritizes exploration of chemical space around proven structures, effectively biasing the Generator Agent towards producing molecules with characteristics demonstrated to be favorable by the Evaluator Agent and Boltz-2 scoring. The selection criteria for Seed Compounds are dynamically adjusted based on screening data, allowing the Optimizer Agent to adapt the generation strategy and accelerate convergence on desired molecular properties.

Rhizome OS-1 employs an iterative workflow where generated molecules are continuously assessed and refined, leveraging physics-informed scoring via the Boltz-2 algorithm to prioritize compounds with desired characteristics. This process facilitates the rapid identification of promising molecules by computationally screening a substantial library of 4,337 compounds. Boltz-2 integrates physical principles into the scoring function, enhancing the accuracy of predictions and reducing the computational resources required to identify viable candidates compared to purely statistical methods. The iterative nature of the loop, combined with Boltz-2’s efficiency, significantly accelerates the molecular discovery process.

Validation of EZH2 binding affinity using Boltz-2 demonstrates strong correlation with experimental data ([latex]pChEMBL[/latex]) via calibration scatter plots and receiver operating characteristic (ROC) curves, and reveals a distribution of predicted affinities for generated molecules consistent with known active compounds.

Validating the Approach: Targeting Key Epigenetic Regulators

The computational system was utilized to design novel molecules targeting both Enhancer of Zeste Homolog 2 (EZH2) and B-cell lymphoma 6 (BCL6), two proteins critically involved in epigenetic regulation and frequently dysregulated in various cancers. EZH2 functions as a polycomb repressive complex 2 (PRC2) component, contributing to gene silencing, while BCL6 acts as a transcriptional repressor affecting B-cell development and function. Targeting these proteins aims to restore normal gene expression patterns and disrupt cancer progression; the system’s application to both indicates a strategy for potentially impacting multiple epigenetic pathways simultaneously.

Performance benchmarking and assessment of synthetic accessibility were conducted utilizing the ChEMBL database. Analysis revealed that 91.9% of the generated Murcko scaffolds were not present within the ChEMBL database, demonstrating a high degree of chemical novelty in the designed molecules. This indicates the system is capable of generating compounds distinct from previously reported structures, potentially circumventing issues related to prior art and expanding the chemical space for therapeutic development. The use of ChEMBL data ensures generated compounds are evaluated against a large, curated dataset of known compounds, providing a quantitative measure of originality.

The computational system is designed to generate molecules that specifically target the BTB domain of the BCL6 protein and inhibit the enzymatic activity of EZH2. Disruption of the BCL6 BTB domain interferes with its function as a transcriptional repressor, while EZH2 inhibition alters histone methylation patterns. These dual mechanisms of action aim to modulate epigenetic regulation in cancer cells, potentially restoring normal gene expression and hindering tumor progression. This targeted approach represents a strategy for developing novel therapeutic interventions by exploiting key epigenetic vulnerabilities in cancer.

A diverse set of nineteen seed compounds, representing six chemotype families including benzimidazolone, pyrazoloquinazolinone, benzoxazepinone, macrocycles, diverse fragments, and chimeras, were selected to initiate a BCL6 generation campaign.

A Vision for the Future: Precision and Personalized Medicine

Rhizome OS-1 signifies a substantial leap forward in pharmaceutical innovation, showcasing artificial intelligence’s capacity to circumvent longstanding challenges within conventional drug discovery. Traditional methods, often reliant on serendipity and lengthy trial-and-error processes, struggle with the sheer complexity of biological systems and the vastness of chemical space. This platform, however, utilizes sophisticated algorithms to rapidly generate and assess countless molecular candidates, predicting their efficacy and safety with unprecedented speed and accuracy. By automating and optimizing key stages of drug development – from target identification to lead optimization – Rhizome OS-1 drastically reduces both the time and cost associated with bringing novel therapeutics to patients, heralding a new era of computationally-driven pharmaceutical research and offering a blueprint for future AI-powered drug design.

Rhizome OS-1 represents a significant leap towards truly personalized medicine by dramatically shortening the timeline for developing epigenetic therapies. The platform leverages artificial intelligence to swiftly design and assess countless molecular candidates, a process traditionally hampered by lengthy laboratory procedures and limited throughput. This accelerated evaluation allows researchers to identify compounds uniquely suited to an individual’s epigenetic profile – the set of modifications to DNA that influence gene expression – rather than relying on broadly effective, yet often imprecise, treatments. Consequently, therapies can be tailored to address the specific epigenetic dysregulation driving a patient’s disease, potentially maximizing efficacy and minimizing adverse effects. The system’s ability to iterate through molecular designs with unprecedented speed promises to unlock treatments for complex conditions, including cancers where epigenetic alterations play a crucial role in disease progression.

The potential to address diseases at their epigenetic roots represents a paradigm shift in medical treatment. Unlike genetic mutations which alter the DNA sequence itself, epigenetic dysregulation involves changes to how genes are expressed – effectively controlling which genes are ‘switched on’ or ‘off’. This means interventions can potentially reverse disease states without permanently altering a patient’s genome, offering a less invasive and more adaptable therapeutic strategy. Cancers, autoimmune disorders, and neurodegenerative diseases are increasingly recognized as heavily influenced by these epigenetic modifications, creating numerous targets for novel therapies. By focusing on restoring healthy gene expression patterns, personalized epigenetic medicine aims to move beyond broad-spectrum treatments and deliver highly targeted interventions, improving efficacy and minimizing adverse effects – ultimately ushering in an era of precision healthcare.

The pursuit of novel compounds, as demonstrated by Rhizome OS-1, benefits greatly from a reduction in superfluous complexity. The system’s autonomous design process, leveraging multi-modal agents and a foundation model, echoes a principle of elegant efficiency. As Claude Shannon observed, “The most important thing in communication is to convey the meaning with the least amount of information.” Similarly, Rhizome OS-1 distills the vast chemical space into focused molecular generation, prioritizing scaffold diversity and binding affinity prediction for targets like BCL6 and EZH2. The platform’s success lies not in exhaustive exploration, but in a streamlined, purposeful search for meaningful chemical structures.

Where Do We Go From Here?

The proliferation of generative models in molecular design inevitably leads to a question of diminishing returns. Rhizome OS-1, by automating the initial stages of compound creation, addresses a practical bottleneck, but does not resolve the fundamental issue: prediction, however sophisticated, remains a proxy for experimentation. A system that requires elaborate computational scaffolding to suggest molecules has, in a sense, already failed to grasp the inherent simplicity of molecular interaction. The true metric of success will not be the number of compounds generated, but the ratio of predicted affinity to actual binding – a ratio that currently favors the laboratory, not the algorithm.

Future iterations should focus not on expanding the generative capacity, but on refining the predictive core. The emphasis must shift from creating diversity for its own sake to generating molecules with a high probability of success, even if it means narrowing the initial search space. A truly autonomous system would not simply propose candidates; it would prioritize experiments, actively seeking disproof of its own hypotheses.

Ultimately, the value of such platforms lies not in replacing chemists, but in augmenting their intuition. Clarity, in this context, is courtesy – a reduction of noise, a streamlining of possibilities. The ideal is not a machine that designs drugs, but one that efficiently directs resources towards the most promising avenues of investigation, allowing human ingenuity to focus on what it does best: understanding the unexpected.

Original article: https://arxiv.org/pdf/2604.07512.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/