Unified autoregressive generation of 3D atomic structures — crystals, molecules, proteins, and protein-ligand complexes — with a single Llama decoder that jointly models discrete tokens and continuous 3D coordinates.

UniGenX represents a structure as one sequence that interleaves discrete tokens (element / residue symbols and special tokens) with continuous coordinate slots. At each step the standard LM head predicts the next token, while a diffusion head samples the (x, y, z) coordinate at every position flagged by a coordinate mask. One forward pass therefore produces both the sequence and its geometry.

Pre-training backbones. csp_pretrain and unified_pretrain are fine-tuning starting points, not per-benchmark evaluation checkpoints, and neither reproduces a specific paper table row on its own. unified_pretrain carries saved args and loads through unigenx_infer.py; csp_pretrain has no saved args and cannot be driven through unigenx_infer.py as-is — train / fine-tune from it first.

c - Conditional

mc - Multi-Conditional

b - base

bs - conditional on bulk

ms - conditional on mag

wo - without structure

1_m to 12_m - 1FME,2F4K,2JOF,2WAV,A3D,CLN025,GTT,NTL9,NuG2,PRB,UVF,lambda (leave-one-out)

mat - material

p - protein

mol - molecules

If you use UniGenX, please cite the paper.

Released under the MIT License; see LICENSE.

see UniGenX-Team/UniGenX: About Code for UniGenX: Unified Generation of Sequence and Structure with Autoregressive Diffusion https://arxiv.org/abs/2503.06687