Loading…
Long-context co-folding model for protein, nucleic-acid and ligand assemblies, folding systems up to 16,384 residues on a single GPU.
No providers recorded yet. Browse all providers
A spliceosome, a proteasome or a ribosome is one object that biology assembles and regulates as a whole, but a co-folding model with a quadratic pair representation cannot hold one in memory. The workaround is to cut the machine apart — truncate it, decompose it into subcomplexes, model one interaction at a time — which discards exactly the long-range context that determines assembly, cooperativity, allostery and function. LightFold is Flagship Pioneering's answer: a proprietary co-folding model from the firm's Pioneering Intelligence group, announced in August 2026 to make long-context structural modeling cheap enough to run routinely.
The bottleneck it targets is architectural. AlphaFold-style models carry one pair-representation entry for every pair of tokens, so memory grows as L², and the triangle operations that update that grid reason over triplets, so arithmetic grows as L³. Other efforts push the ceiling outward with hardware — NVIDIA's Fold-CP distributes inference across 64–144 server-grade GPUs and still needs an hour for a single prediction above 10,000 residues. LightFold instead restructures the dependency between the pair representation and the triangular operations, removing the scaling term rather than sharding it, building on the linear-scaling reformulation of invariant point attention (FlashIPA) that the same group presented at NeurIPS 2025. The claim is efficiency and reach, not accuracy leadership: AlphaFold 2, AlphaFold-Multimer and AlphaFold 3 successively widened what co-folding covers, and LightFold widens what it can afford to cover, with per-chain accuracy sitting slightly below the open models it is measured against.
Flagship evaluated LightFold on 894 complexes of 1,000–5,500 residues from a validation set with low sequence homology to its training sequences, at five seeds and five samples per seed. That run consumed 17 GPU-hours against 107 GPU-hours for OpenFold3 — 20 diffusion steps, five seeds, one sample per seed, all memory optimizations — despite drawing five times as many samples. Protein, RNA and DNA per-chain accuracy remains slightly below but closely tracks OpenFold3 over the shared range, and interface accuracy is challenging for both. At 2,048 residues LightFold returns a prediction in 2.5 seconds against 110 seconds for Boltz-2, 135 for Protenix-v2 and 210 for OpenFold3; at 4,000 residues, 6.5 seconds against roughly nine minutes, past which OpenFold3 was impractical on the test hardware.
The headline demonstration is the E. coli 70S ribosome (PDB 4YBB) — 10,730 residues and 144,172 atoms across three ribosomal RNAs and roughly 50 proteins — co-folded in 36 seconds on one RTX PRO 6000, a reported ~6,400× wall-clock speedup and 48× memory reduction against Fold-CP-enabled models. Accuracy there is uneven, and the developers present it that way: the rRNA scaffold is recovered at 4.11 Å median per-chain RMSD and protein chains at 5.52 Å, while protein–RNA interfaces are largely missed at 0.15 lDDT. Accuracy also degrades as complex size grows. Every figure comes from the August 2026 announcement, the only public source; no preprint, parameter count or training-corpus description accompanies it.
LightFold targets drug discovery on systems that only make sense in context: proteasomes where binding has effects hundreds of ångströms away, spliceosomes where protein–RNA interplay is central, the large E3 ligases used in targeted protein degradation, and respiratory supercomplexes. It extends to molecular engineering — viral capsids and nanocages evaluated with their cargo, ribonucleoprotein systems designed in full context — and the per-prediction cost makes high-throughput screening routine. Being proprietary, its beneficiaries are Flagship's portfolio companies rather than the wider field.
LightFold's clearest contribution is to separate two problems the field had coupled: making long context computationally accessible, and learning to predict it accurately. It addresses the first — interface accuracy at scale is where its own numbers are weakest — but cheap long-context training and sampling make the second far easier to study. The experimental record is moving the same way: structures above 2,048 residues rose from 14% of PDB depositions in 2019 to 25% in 2025. The caveat is access: no code, weights, API or paper accompanies the announcement, so none of the reported figures can be independently reproduced.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.