ProtMonomer

A protein's multiple sequence alignment is the channel through which an AlphaFold2-style predictor learns about structure, and how deep it runs largely decides the outcome: sequences with thousands of detectable homologs fold accurately, novel folds and orphans do not. Single-sequence predictors built on protein language models show the mirror-image behavior — weaker than AlphaFold 2 when homologs are plentiful, stronger on orphans. ProtMonomer starts from the observation that a lone query sequence is simply an alignment of depth one, and asks whether that divergence reflects the architectures or the MSA depths each was trained on.

Tested directly, the answer holds. Otherwise-identical models trained on alignments downsampled to fixed depths of 32, 128 and 512 showed the complementary pattern, each ceasing to improve once inference depth passed its training depth. A model trained on an equally weighted mixture of those depths then matched or beat a retrospective oracle picking the better prediction from the corresponding specialists. ProtMonomer is built on that result: trained across a distribution spanning 1 to 2,048 homologous sequences, then refined in two stages weighted toward shallow alignments.

It comes from Yunda Si, Suqi Zhang and Luonan Chen at the University of Chinese Academy of Sciences, with collaborators at the Naval University of Engineering and Shanghai Jiao Tong University. It handles monomers only, and is the protein counterpart to NUMonomer, the same group's nucleic acid predictor, which folds a bare sequence where ProtMonomer requires an alignment.

#Key Features

• Multi-scale MSA-depth training: Alignment depth is sampled from four equally weighted intervals (1–32, 32–128, 128–512, 512–2,048), exposing one checkpoint to the full range of evolutionary information rather than the deep alignments that dominate the PDB.
• No full-network recycling: Predicted structures are never fed back through the encoder, forcing the network to draw its structural signal from the supplied alignment alone and removing the dominant cost of AlphaFold-style inference.
• Alternative conformational states: Paired with MSA-partitioning strategies such as AF-Cluster, SPEACH_AF and CF-Random, it recovers both experimentally observed states of fold-switching proteins several times more often than the predictors it was benchmarked against.
• Confidence that tracks accuracy: Predicted and measured LDDT-Cα correlate at 0.759 across the fold-switching set, against 0.616 for AlphaFold 2 and 0.437 for AlphaFold 3, making pLDDT a usable filter for native-like models in an ensemble.
• High-throughput inference: A 1,214-residue protein folds from a depth-128 alignment in 11.37 seconds on one NVIDIA H100 PCIe GPU.

#Technical Details

The framework has three parts. A 32-block MSA encoder, a modified Evoformer, produces MSA and inter-residue pair representations. Six equivariant structure blocks then predict and iteratively refine all-atom coordinates; each chains a MSA2Atom module, inter-residue and intra-residue attention, and a feed-forward network, starting from a randomly initialized structure. A three-block confidence module predicts per-residue LDDT-Cα. Training ran in three PyTorch stages on H100 and A100 GPUs, the third swapping the all-atom FAPE loss for an all-atom, all-frame variant that cut clashing atom pairs from roughly 5,000 to roughly 200. The data comprised 132,905 experimental monomers deposited in the PDB before 30 December 2021, plus a 905,643-structure distillation set from OpenProteinSet and PSP predictions, de-duplicated at 40% identity.

On the 17 hardest CASP15 targets, ProtMonomer reached a mean TM-score of 0.652 against 0.585 for AlphaFold 2 and 0.603 for AlphaFold 3. On a 176-protein hard subset of 866 recently released PDB structures it scored 0.687 against 0.651 and 0.635, succeeding (TM-score ≥ 0.5) on 86.4% of cases versus 77.2% and 74.4%. On 15 orphan proteins it reached 0.674, ahead of AlphaFold 2 at 0.609 and Boltz-1 at 0.641, level with AlphaFold 3 at 0.680. On 54 CAMEO peptides it averaged 4.63 Å Cα RMSD against 5.55 Å and 5.44 Å, and on 68 fold-switching pairs AF-Cluster sampling recovered both conformations for 54.4% of proteins against 14.7% and 13.2%.

#Applications

The model suits sequences that alignment-hungry predictors handle poorly: designed constructs with no natural relatives, metagenomic proteins from undersampled clades, short peptides and rapidly evolving families. The conformational results extend it to proteins whose function depends on switching folds, where one static model misrepresents the biology. Because inference is roughly two orders of magnitude cheaper than AlphaFold 2, proteome-scale prediction becomes practical on modest GPU allocations.

#Impact

ProtMonomer's contribution is a training-time control knob rather than a new architecture: the generalization profile of an MSA-based predictor can be dialed in by choosing the depth distribution it sees, and a mixture recovers the strengths of specialists trained at individual depths. The authors flag protein complexes and RNA as the places to test whether that carries. Limits are clear: accuracy on peptides and orphan proteins is modest in absolute terms, complexes are out of scope, and the work is a preprint awaiting peer review. It states that source code and pretrained weights are available in a public GitHub repository under an MIT license; that repository presently contains the license file alone.