What is LMI4Boltz?
Large Boltz jobs often fail for a simple reason: the model can handle the chemistry, but the GPU runs out of memory before inference finishes. LMI4Boltz solves that bottleneck by running Boltz with low-memory inference changes while keeping the same structure and affinity prediction model family.
The method is designed for biomolecular complexes that are too large or memory-heavy for standard Boltz-2 settings. It predicts structures for proteins, small molecules, DNA, RNA, and mixed complexes, and can use Boltz-2 affinity prediction for protein-ligand binding. LMI4Boltz is most useful when the scientific question calls for Boltz, but the system size pushes ordinary inference close to the VRAM limit.
How to use LMI4Boltz online
Run LMI4Boltz online by combining one or more protein, ligand, DNA, or RNA inputs into a single complex prediction job. ProteinIQ accepts FASTA sequences, RCSB protein fetches, SMILES ligands, PubChem ligands, and CCD ligand codes, then returns predicted CIF or PDB structures plus confidence, pLDDT, PAE, PDE, and affinity files when generated.
LMI4Boltz builds one prediction target from the submitted molecules. Chain IDs are shown in the input interface and become important when interpreting interface confidence, ligand affinity, and multi-chain contacts.
At least one foldable biopolymer is required. Ligand-only jobs are not valid because Boltz predicts complexes, not isolated small-molecule conformer ensembles.
Settings
Most jobs should start with the standard settings: one diffusion sample, Boltz-2, no manual memory tuning, and MSA generation enabled only when evolutionary information is needed for protein chains. The low-memory controls are mainly for large complexes that fail or approach the memory limit.
Standard inference
Advanced inference
Low-memory controls
LMI4Boltz reduces peak memory by chunking expensive pair and MSA operations, moving some tensors to host memory, updating pair representations in place, and optionally using reduced precision. Smaller chunks usually reduce memory demand but can add overhead.
Results
LMI4Boltz returns a structure viewer, downloadable structure files, and data files from the prediction run. File availability depends on selected options and whether the selected model supports affinity prediction.
Understanding LMI4Boltz results
The first check should be structural confidence, not affinity. A strong affinity score on a low-confidence protein-ligand pose is weak evidence because the binding estimate depends on a plausible complex geometry.
Structure confidence
Boltz confidence fields are reported on a 0 to 1 scale, where higher values indicate more model confidence. The most useful distinction is between local confidence and interface confidence.
As a rough guide, confidence values above 0.7 usually justify closer structural interpretation, 0.5 to 0.7 calls for visual inspection and reruns, and values below 0.5 often indicate an underconstrained system, insufficient MSA signal, problematic input chemistry, or genuine conformational ambiguity.
Affinity outputs
Boltz-2 affinity files contain two kinds of predictions. Affinity is requested for form-based jobs when the complex contains one small-molecule ligand and at least one protein target. Jobs with multiple ligands, ligand-only inputs, DNA/RNA-only targets, or Boltz-1 still return structures and confidence files, but skip affinity prediction.
The affinity value is not pIC50. In the Boltz convention, lower is stronger because the value is log(IC50) with IC50 measured in micromolar:
For ligand ranking, compare compounds against the same target, same input preparation, same MSA choice, same sampling settings, and preferably the same random seed policy. Cross-target affinity values are much less meaningful.
How LMI4Boltz works
LMI4Boltz changes how Boltz inference is executed, not the biological question the model is trying to answer. The underlying Boltz model remains a diffusion-based biomolecular structure predictor, with Boltz-2 adding an affinity module trained to estimate protein-ligand binding.
The low-memory changes target the expensive pair-representation operations that scale poorly as complexes grow. Pair tensors represent relationships between tokens, so memory rises quickly with longer proteins, deeper MSAs, and multi-chain assemblies. The LMI4Boltz preprint reports a 66.7% higher Boltz-2 token limit on 24 GB VRAM, and the project README reports a 100% increase for Boltz-1. LMI4Boltz reduces peak VRAM by:
- Updating pair representations in place: Intermediate tensors are reused instead of materializing more copies than needed.
- Offloading less frequently used tensors to host memory: GPU memory is reserved for active computation, while some data can sit in CPU memory.
- Using reduced precision for pair features: Lower precision reduces memory footprint, with small numerical differences expected.
- Chunking memory-heavy operations: Triangle attention, MSA transition, outer product, and related operations can run in smaller pieces.
The LMI4Boltz README reports near-exact reproduction of ordinary Boltz outputs and negligible runtime change in the tested settings. Small numerical differences can still occur because reduced precision and altered execution order change floating-point arithmetic.
When to use LMI4Boltz vs alternatives
LMI4Boltz is not a different scoring model from Boltz. It is the practical choice when Boltz-style prediction is desired and memory is the limiting factor.
For a practical pipeline, LMI4Boltz can generate candidate complex structures for large systems, DockQ can score protein-protein interface models when a reference is available, and GROMACS or gmx_MMPBSA can support downstream simulation-based checks.
Common failure modes
- Incorrect ligand stereochemistry: SMILES must encode the intended enantiomer. Binding predictions can change when stereocenters are missing or wrong.
- Treating peptides as small molecules: Peptide ligands are better entered as protein chains. The SMILES ligand path is capped at 300 heavy atoms and is intended for small molecules.
- Reading affinity without pose quality: Affinity values should be interpreted only after checking ligand interface confidence and visual pose plausibility.
- Disabling MSA for natural proteins: Single-sequence protein predictions can work, but many natural proteins benefit from MSA signal.
- Over-tuning memory chunks first: If a job fails, reduce sample count, MSA depth, or system size before changing every low-memory chunk parameter at once. One change at a time makes reruns easier to interpret.
