MEVO

Structure-based drug design (SBDD) has not benefited from generative modeling as dramatically as vision or language, and the reason is data. Experimentally determined protein-ligand complexes number in the tens of thousands, and affinity-labelled sets such as PDBbind are smaller still, while the purchasable chemical space has grown past a trillion compounds. Models trained on pocket-conditioned complexes inherit the narrow chemical diversity and affinity biases of those datasets; ligand-based models enjoy effectively unlimited training data but ignore binding site geometry altogether.

MEVO, released by ByteDance Seed in July 2025, addresses that asymmetry by splitting the problem rather than enlarging the dataset. Universal chemical rules are learned once from billion-scale ligand-only libraries through pharmacophore-conditioned generation, and target-specific adaptation happens at inference through an evolutionary loop that requires no additional training. The pharmacophore — a spatial arrangement of chemical features a ligand must present — is the bridge: it can be extracted from abundant ligand-only data during training, and inferred from pocket geometry and from previously generated molecules at design time.

The framework has three parts: a vector-quantized variational autoencoder (VQ-VAE) mapping 3D molecules to discrete latent tokens, a latent discrete diffusion model conditioned jointly on pharmacophore and pocket, and a physics-scored evolutionary strategy that refines the output. This places MEVO alongside pocket-aware generators such as Apo2Mol, FLOWR.root, and UniLingo3DMol, but with target-specific optimization moved out of the network weights and into a search loop.

#Key Features

• Pharmacophore as a data bridge: Conditioning on 3D pharmacophores lets the generator train on billion-scale ligand-only libraries while still producing molecules that satisfy pocket-derived requirements, sidestepping the scarcity of protein-ligand complexes.
• Training-free evolutionary optimization: Seed molecules from each generation are converted back into pharmacophore conditions for the next round, so affinity improves through iteration alone — no fine-tuning or per-target retraining.
• Near-lossless latent representation: The VQ-VAE unifies categorical atom and bond features with continuous coordinates in one discrete latent space, reconstructing molecules at over 99.99% categorical accuracy and under 0.05 Å coordinate RMSD.
• Physics-informed scoring: Candidates are ranked by the product of a binding potential energy change and the fraction of user-specified interactions satisfied, at roughly docking cost but ranking hits better than docking scores.
• Use across design stages: One model handles de novo hit generation, fragment growing, linker design, and substituent enumeration for lead optimization.

#Technical Details

The VQ-VAE is trained on 9.6 billion synthetically feasible compounds from Enamine REAL plus over 750 million commercially available ZINC20 molecules, resampled to balance heavy-atom counts across the 20–50 range. Hydrogens are represented implicitly through hydrogen counts and formal charges, preserving protonation states that govern hydrogen bonding. The generator is a Discrete Denoising Diffusion Probabilistic Model (D3PM) over those latent tokens, with a transformer denoising network that attends to stacked pharmacophore and pocket condition embeddings; pocket-conditioned training uses protein-ligand complexes from HelixDock. Generated conformations are relaxed with molecular mechanics force fields — amber-ff14SB for protein, ByteFF for ligand — shifting coordinates by an average 0.7 Å RMSD.

On five targets from a public binding free energy benchmark, absolute free energy perturbation (ABFEP) calculations on the top-10 generated molecules per target gave affinities comparable to or better than known reference binders for four of the five, reached after 10 to 50 evolution rounds, with low Tanimoto similarity to the reference ligands and high scaffold diversity. Across 250 generated molecules, drug-likeness metrics tracked 2,094 FDA-approved drugs closely on logP and ring-size distribution, with slightly lower QED and higher molecular weight. In a KRAS G12D case study, MEVO grew a fragment of MRTX-1133 into a new switch-II pocket binder whose absolute binding free energy was computed at -22.31 kcal/mol with Schrödinger FEP+, comparable to MRTX-1133 itself.

#Applications

MEVO targets the early stages of small-molecule campaigns, where it can substitute for a virtual screen: given only a protein structure and the residues defining a pocket, it proposes binders without fragments, scaffolds, or reference ligands. Because the evolutionary loop is decoupled from training, medicinal chemists can specify required interactions, constrain heavy atom counts, or swap in a different scoring function — including objectives such as solubility or permeability — without retraining anything. The KRAS G12D study illustrates the workflow, chaining hit generation, linker design, and substituent enumeration through the same model.

#Impact

MEVO's contribution is methodological as much as practical: it demonstrates that scarce structural supervision can be traded for abundant ligand-only data by choosing an intermediate representation both sources share, a strategy the authors argue extends to other structure-based biomolecular tasks starved of training data. The evaluation is entirely computational — FEP-predicted affinities, with no synthesis or experimental assays reported — so the designed KRAS G12D leads remain unvalidated in the laboratory, and comparisons against other pocket-aware generative baselines are not included. The work is an arXiv preprint that has not been peer reviewed, and neither code nor model weights have been released.