UniGenX

Most generative models in the molecular sciences split a problem that nature does not: discrete sequence and continuous geometry are optimized by separate machinery, and the function a designer actually cares about is enforced afterwards by filtering rather than during generation. UniGenX, from Microsoft Research AI for Science and academic collaborators in China, is built around the opposite premise — that sequence, structure, and property should be produced together, in one model, under a single objective.

The mechanism is a mixed token stream. Symbolic tokens (amino acids, atom types, element symbols, conditioning tags such as an EC number or a target band gap) and numeric tokens (3D coordinates, lattice parameters) are interleaved into one sequence. A decoder-only Llama-style transformer supplies global context over the whole stream and receives a cross-entropy loss on the discrete positions; at each numeric position it hands off to a lightweight conditional diffusion head that generates the continuous values, steered by the task tokens that precede it. The result is "word-to-word, number-to-number" prediction in which a designed sequence and its coordinates emerge from the same forward pass.

UniGenX is deliberately not a biology-only model. It is trained and evaluated across proteins, small molecules, and crystalline materials, and much of the paper's headline evidence comes from materials science, outside bio.rodeo's scope. The entry here covers the biological and chemical half: protein structure and dynamics, protein-ligand complexes, enzyme design, and molecular conformer generation. The paper was posted in March 2025 and revised that August; it has not been peer reviewed.

#Key Features

• Joint discrete-continuous generation: Sequence tokens and coordinates are produced in the same autoregressive stream, so structure is not a downstream prediction applied to a finished sequence.
• Function-conditioned design: Task tokens prepended to the stream condition generation on the property wanted — an enzyme commission number for catalytic function, a target quantum property for a molecule — rather than screening candidates after the fact.
• Induced-fit and ensemble modeling: The model generates a ligand together with the holo pocket conformation starting from an apo structure, and samples conformational ensembles rather than a single static pose.
• Cross-domain transfer: A single architecture covers proteins, small molecules, and crystals, and ablations report that joint training across domains improves results within each one.
• Lightweight diffusion head: Because diffusion operates only on low-dimensional numeric fields such as 3D coordinates, it adds little cost on top of the backbone.

#Technical Details

Two configurations are reported, with roughly 100M and 400M parameters in the transformer decoder backbone. Protein training used approximately 78 million sequence-structure pairs from the AlphaFold Database plus about 20 million EC-annotated entries from UniProt; molecular training used GEOM-QM9 and GEOM-Drugs, together over 37 million conformations; protein-ligand work used the MISATO set of 20,000 complexes with molecular dynamics trajectories, reduced to 9,744 training samples after preprocessing; materials training combined MP-20, Carbon-24, MPTS-52, and a NOMAD subset. On protein-ligand tasks the model improves success at modeling induced fit (RMSD under 2 Å) by more than 23-fold. Equilibrium distributions for 12 fast-folding proteins from long molecular dynamics trajectories were reproduced with native and unfolded states consistent with the expected free energy landscapes, assessed by time-lagged independent component analysis. On molecular property prediction it reports a mean absolute error of 0.1464 eV for average energy against 0.4324 eV for DMCG, and up to 260% improvement on conditional generation across five quantum property targets. In materials, multi-property conditional generation yielded 436 crystal candidates satisfying three simultaneous constraints, 11 of them with novel compositions.

#Applications

For biology, the practical use cases are enzyme design conditioned on a desired EC class, structure-based ligand design where the binding pocket is expected to reorganize on binding, and conformer ensemble generation for flexible molecules — all cases where generating sequence and geometry together is more faithful than predicting one from the other. Designed enzymes in the paper are evaluated with CLEAN for EC conformance, ESMFold for structure, and BLAST and Foldseek for novelty against known sequences and folds. The same checkpoint family supplies property-conditioned generation for chemistry and materials work.

#Impact

UniGenX argues that the discrete-continuous split running through molecular generative modeling is an artifact of tooling rather than a necessity, and its cross-domain ablations are its strongest evidence. Its artifacts sit in two homes. The MIT-licensed microsoft/UniGenX repository holds the model, generation, and evaluation code but no training entry point; the Hugging Face card points instead to UniGenX-Team/UniGenX, which adds a training script, DeepSpeed trainer, and reproduction guide under MIT. That GitHub account and the Hugging Face account holding the 45 checkpoints are personal accounts created in July 2026, not Microsoft-affiliated organizations, and the latter has no LICENSE file behind its MIT claim. Two caveats: evaluation is entirely computational, with no experimental validation of designed enzymes or crystals, and the paper reports no head-to-head comparison against dedicated structure predictors such as AlphaFold 3 on standard folding benchmarks, so those claims rest on the specialized tasks it defines.