How to Use LMI4Boltz

LMI4Boltz is a low-memory fork of Boltz-2 designed for all-atom biomolecular complex prediction when GPU memory, batch size, or complex size would otherwise be limiting. It keeps the same practical modeling scope as Boltz-2 across proteins, nucleic acids, ligands, cyclic or modified biopolymers, and restraint-guided complexes, while exposing additional chunking and precision controls that make larger systems more tractable on constrained hardware.

On Neurosnap, researchers can use Input Sequences, Input Molecules, optional Custom MSA, and the same Pocket Restraints, Covalent Restraints, and Contact Restraints workflow as in Boltz-2, but also tune low-memory settings such as Chunk Size Transition Z, Chunk Size Transition MSA, Chunk Size Triangle Attention, Chunk Size Outer Product, Chunk Size Threshold, Triangle Multiplication Gate Chunks, and Use Bfloat16. That makes the service useful when the biological question fits Boltz-2, but the system size or hardware budget requires more explicit memory management.

Scientifically, LMI4Boltz is best understood as a systems-oriented extension of the Boltz-2 architecture rather than a new prediction paradigm. The underlying modeling target remains AlphaFold3-class all-atom structure prediction with an affinity head for ligand-containing systems, but the implementation introduces aggressive chunking, reduced-precision execution, and memory-offloading strategies to lower VRAM pressure during inference. In practical terms, that means the user can explore larger multimolecular assemblies or more difficult jobs without giving up the familiar Boltz-2-style confidence outputs and interaction-analysis workflow.

On Neurosnap, the standard Boltz controls still define the core experiment. MSA Mode and Custom MSA determine how evolutionary context is provided. Recycling and sampling parameters govern refinement depth and structural diversity. Use Inference Time Potentials and the restraint inputs remain the main way to inject mechanistic priors such as hypothesized binding pockets, covalent chemistry, or atom-level contacts. The LMI-specific controls should be viewed as computational tuning knobs rather than biological assumptions: smaller chunk sizes and optional Use Bfloat16 primarily trade numerical behavior and runtime for lower memory overhead.

Researchers should therefore interpret LMI4Boltz in the same way they would interpret Boltz-2 scientifically. pLDDT, PAE, PDE, pTM, ipTM, ipSAE, LIS, pDockQ, and pDockQ2 remain the key outputs for deciding whether chain placement, interface geometry, and local residue environments are trustworthy. For ligand-containing systems, affinity predictions are still most useful comparatively across candidate ligands, poses, or analog series. The main distinction is that LMI4Boltz broadens the class of systems that can be attempted under finite-memory constraints.

• Boltz-2-compatible science, lower memory cost: LMI4Boltz preserves the multimodal biomolecular modeling workflow of Boltz-2 while adding explicit controls for memory-intensive parts of inference.
• Larger-complex practicality: Additional chunk-size controls and optional Use Bfloat16 make it more feasible to run difficult protein-protein, protein-nucleic-acid, and protein-ligand systems that might otherwise exceed available VRAM.
• Constraint-aware complex modeling: Pocket Restraints, Contact Restraints, Covalent Restraints, steering potentials, and custom MSAs remain available, so reducing memory pressure does not require giving up biologically informed setup.
• Decision-ready interpretation on Neurosnap: Researchers still receive the same confidence and interface-analysis outputs used for downstream prioritization in docking follow-up, medicinal chemistry, mutagenesis, and wet-lab validation.