UniGenX

Abstract

Function in natural systems arises from 1D sequences forming 3D structures with specific properties. Current generative models rarely target function directly, optimize discrete sequences and continuous coordinates in isolation, and under-model conformational ensembles. UniGenX is a unified generative foundation model that co-generates sequences and coordinates under direct functional and property objectives across proteins, molecules, and materials.

Key Contributions

• Unified co-generation of sequence + structural coordinates under functional objectives, cross-domain.
• Architecture: decoder-only autoregressive transformer + conditional diffusion head for numeric (coordinate) fields.
• Heterogeneous mixed symbolic/numeric token stream; task-specific token steering for property-conditioned generation.

Results

• Materials: 436 crystal candidates meeting triple constraints; 11 novel compositions. Evidence: strong
• Chemistry: new SOTA on five property targets; improved GEOM conformer ensembles.
• Biology: protein induced-fit modeling 23× improvement (RMSD < 2 Å); improved EC-conditioned enzyme design.

Limitations

Not explicitly enumerated in the fetched abstract/landing content (flag for full-text read).

Metadata

arXiv 2503.06687 (cs.LG); v1 2025-03-09, v2 2025-08-26; CC BY 4.0; 34 authors.

Source: UniGenX (arXiv:2503.06687). Abstract + landing fetched; full PDF not yet read.