Macro-Equi-Diff (MED)

Macro-Equi-Diff (MED) is a deep-learning framework for generating macrocycles—large ring-shaped molecules that occupy a chemical space between small molecules and biologics and are prized for their ability to bind challenging "undruggable" targets. Designing macrocycles is hard: closing a ring imposes strict geometric and conformational constraints, and most generative chemistry models are tuned for acyclic, drug-like small molecules. MED addresses this by directly tackling the problem of turning an open-chain (acyclic) molecule into a valid macrocycle.

The method was posted to bioRxiv in February 2026 by Sai Shobit Kambampati and colleagues at Keshav Memorial Institute of Technology and Drugparadigm Research Lab in Hyderabad, India. MED is a two-stage system: a transformer-based component identifies where and how to cyclize, and an E(3)-equivariant diffusion model generates the macrocyclic structure while respecting 3D geometric symmetry. Equivariance ensures that the model's predictions transform consistently under rotation and translation, an important inductive bias for generating physically realistic 3D molecules.

Trained and evaluated on the ZINC dataset, MED reports strong generative metrics and a case study macrocyclizing known JAK2-targeting drugs. The work is distributed under a CC BY-NC-ND 4.0 license.

#Key Features

• Acyclic-to-macrocycle generation: Converts open-chain molecules into macrocycles, directly targeting a hard-to-access region of chemical space.
• E(3)-equivariant diffusion: Generates 3D structures with an equivariant diffusion model so outputs respect rotational and translational symmetry.
• Transformer-based identification: A transformer component identifies cyclization sites and linkers before the diffusion stage generates the ring.
• Strong generative metrics: Reports 93.92% validity, 99.94% uniqueness, 99.92% macrocyclization, and 82.81% linker novelty on ZINC.
• Target-focused validation: Demonstrated by macrocyclizing four JAK2-targeting drugs, yielding compounds with favorable descriptors and binding properties.

#Technical Details

MED couples a transformer-based identification module with an E(3)-equivariant diffusion model. The transformer stage determines how an acyclic input should be cyclized, and the equivariant diffusion stage generates the macrocyclic product in a way that is consistent under Euclidean symmetries. Trained on the ZINC dataset, the model reports high validity (93.92%), uniqueness (99.94%), successful macrocyclization (99.92%), and linker novelty (82.81%). As a proof of concept, the authors macrocyclized four JAK2-targeting drugs and report that the resulting compounds had favorable molecular descriptors and binding properties. The preprint does not report a released parameter count, and code/weights availability is not specified; results reflect the authors' own evaluation.

#Applications

MED is aimed at macrocycle-focused drug discovery, where converting a known acyclic ligand into a macrocyclic analog can improve binding, selectivity, metabolic stability, or membrane permeability against difficult targets such as protein-protein interactions. Medicinal and computational chemists could use such a tool to generate macrocyclic candidates from existing leads, as illustrated by the JAK2 case study, expanding the design space beyond conventional small molecules.

#Impact

MED contributes to the emerging application of equivariant diffusion models to specialized chemical-design problems, here the underexplored task of programmatic macrocyclization. By combining scaffold identification with symmetry-aware 3D generation, it offers a template for target-conditioned macrocycle design. As a February 2026 bioRxiv preprint without confirmed released code or weights, its practical adoption and the generality of its reported metrics beyond ZINC remain to be validated, and the restrictive CC BY-NC-ND license limits derivative use.