PLAME

Modern structure prediction is only as good as the evolutionary signal it is given. AlphaFold 2 reads a multiple sequence alignment (MSA) of homologs and infers residue contacts from their covariation, which works beautifully for well-populated families and degrades sharply for orphan proteins and shallow alignments — precisely the designed sequences, viral proteins, and novel folds that researchers most often care about. PLAME, developed at the Chinese University of Hong Kong with collaborators at MIT, UC San Diego, the University of Amsterdam, and the University of Pennsylvania, attacks the problem at its source: rather than searching harder for homologs that may not exist, it designs an alignment.

The key move is where the evolutionary prior comes from. Earlier generative MSA methods learn a distribution over homologs from existing alignments, which leaves them helpless when the input alignment is empty or nearly so. PLAME instead conditions on the per-residue embeddings of a pretrained protein language model — ESM-2 — which encodes evolutionary constraints learned across the whole of UniRef and is available for any single query sequence. A generator then emits synthetic homologs from that latent, and the resulting sequences are concatenated with whatever real MSA exists before being handed to the folding model.

Because PLAME operates purely on the input side, it slots into an existing pipeline without touching the predictor. The same fixed checkpoint is applied to proteins it has never seen, and its generated alignments improve AlphaFold 2 across all three of its standard folding configurations, carry over to AlphaFold 3, and let ESMFold approach AlphaFold 2 accuracy while retaining ESMFold's MSA-free inference speed. The work first appeared in June 2025.

#Key Features

• Language-model evolutionary prior: Generation is conditioned on ESM-2 sequence embeddings rather than on the query's existing alignment, so the model still has an informative prior when no homologs can be found.
• Conservation–diversity loss: A combined objective balances agreement at conserved positions against coverage of plausible variation, avoiding the failure mode where high-diversity generators inject noise that degrades folding.
• HiFiAD alignment selection: A selection strategy filters generated candidates down to a high-quality subset, which ablations show is responsible for a meaningful share of the folding gains.
• Sequence-quality metric: The authors define an alignment-quality score complementary to depth-based measures and show it is predictive of downstream folding improvement, giving a way to triage generated MSAs before running a predictor.
• Drop-in adapter: Nothing about the folding model is retrained or modified — PLAME augments the MSA and hands it off, so it composes with AlphaFold 2, AlphaFold 3, and ESMFold alike.

#Technical Details

PLAME uses a T5-style encoder–decoder with 12 encoder and 12 decoder layers, hidden size 1024, 12 attention heads, and a 2048-dimensional gated-GELU feedforward block. Following the MSA Transformer design, each block adds row-wise and column-wise attention to capture co-evolutionary structure across the alignment. Training data comes from the PDB and UniClust30 subsets of OpenProteinSet together with OpenFold's pre-searched MSAs, keeping only entries with at least 64 aligned sequences and removing anything above 90% similarity to the test set, which yields 293,979 samples split 90:10 into training and validation. Optimization uses AdamW at a 5e-5 learning rate for up to 200,000 steps on four NVIDIA A40 GPUs. Evaluation uses 200 proteins drawn from CASP14, CASP15, CAMEO, and the PDB. In the zero-shot setting — queries with no existing MSA — PLAME reaches 71.50 pLDDT under AlphaFold 2's full template-enabled configuration, ahead of the search-derived AlphaFold 2 MSAs at 70.31, MSAGPT at 68.39, and EvoDiff at 64.39; under AlphaFold 3 it reaches 70.23 pLDDT against 66.34 for standard MSAs.

#Applications

PLAME targets the cases where conventional homology search returns little or nothing: de novo designed proteins, rapidly evolving viral antigens, metagenomic sequences from underexplored clades, and singleton families. For groups running structure prediction at scale, the sequence-quality metric offers a cheap triage step, flagging which targets are likely to benefit from augmentation before committing GPU time. The ESMFold pairing is the most practical result — it gives a route to near-AlphaFold 2 accuracy at MSA-free speed, which matters for proteome-scale screens.

#Impact

PLAME reframes MSA generation as a design problem informed by a protein language model rather than an extrapolation from existing alignments, and its ablations make a useful negative point: maximizing generated diversity actively hurts folding, whereas matching the statistical profile of natural alignments helps. The main limitation for practitioners is availability — the preprint reports no code or weights release, so the results cannot yet be reproduced or the model applied outside the authors' group. The work is a preprint and has not been peer reviewed, and evaluation is entirely computational, using predicted-structure metrics against experimentally determined references rather than new experiments.