Loading…
Enzyme function prediction that scores whether two sequences catalyze the same reaction, via attention pooling over frozen ESM Cambrian embeddings.
No providers recorded yet. Browse all providers
A BLAST hit tells you two enzymes are evolutionarily related, not that they perform the same chemistry — homologs routinely diverge onto different substrates, and distant relatives converge on the same one. Yet copying the closest hit's annotation onto the query remains how most uncharacterized enzymes get a function assigned. UNKAI addresses the narrower question that follows a homology search: do these two sequences catalyze the same reaction on the same substrate?
UNKAI answers it with a Siamese network over a frozen protein language model. Residue-level embeddings from ESM Cambrian at the 6B scale are collapsed to one fixed-length vector per protein by a learned attention-pooling layer, not by averaging. The element-wise absolute difference between the two pooled vectors feeds a small multilayer perceptron that emits a probability of functional identity. Because the pooling weights are learned against the functional-identity label, they concentrate on the residues that decide the answer — and those turn out to be catalytic and binding sites, making the model's reasoning inspectable.
Kotaro Ukai, Suguru Fujita and Tohru Terada of the University of Tokyo posted UNKAI in May 2026 and published it in Computational and Structural Biotechnology Journal that August. It succeeds the same group's FUJISAN, a LightGBM model over hand-engineered sequence, domain and pocket-similarity features derived from AlphaFold 2 structures — the name is a pun, unkai being the sea of clouds that extends beyond the summit of Fujisan.
ESM-C 6B produces a 2,560-dimensional embedding per residue and is never fine-tuned. Attention logits come from a linear projection of each residue embedding, whose weighted sum yields v ∈ ℝ²⁵⁶⁰. The classifier is four fully connected layers of 2,560 → 1,599 → 781 → 117 → 1 with batch normalization, ReLU and dropout at 0.303, ending in a sigmoid — roughly 5.4 million trainable parameters, all in the head. Inputs cap at 2,048 residues, ESM-C's context limit. A second checkpoint is released alongside it, tuned for the seen-unseen split: the same Siamese scaffold over a narrower classifier of 2,560 → 1,408 → 640 → 512 → 1, with dropout raised to 0.528 and a further 0.163 dropout applied to the attention weights themselves.
Training data came from 232,692 Swiss-Prot sequences carrying Rhea reaction identifiers, with functional identity defined as sharing at least one Rhea ID. An all-against-all BLAST search (E < 10) supplied candidate pairs, from which 100,000 positive and 100,000 negative pairs were sampled and split 70/15/15. Adam at a learning rate of 3.33 × 10⁻⁴ ran binary cross-entropy for 10 epochs at batch size 64, hyperparameters chosen by 50 Optuna trials and the best-F1 validation checkpoint retained — about four days on a single RTX 5080. On the held-out test set UNKAI reaches AUROC 0.9939 and AUPR 0.9931, against 0.9427/0.9413 for FUJISAN. On a low-sequence-similarity set of 680 positive and 680 negative pairs built independently of training, it reaches 0.9673/0.9749. Under MMseqs2 clustering at 40% identity with clusters held disjoint across splits, accuracy is 0.8206 and AUROC 0.9012 in the strict setting, rising to 0.8950 and 0.9608 when one member of each pair comes from a seen cluster.
UNKAI sits downstream of a homology search rather than replacing it: BLAST proposes candidate partners for an uncharacterized enzyme, UNKAI scores each one, and annotations are ranked by predicted functional identity rather than E-value alone. That fits orphan enzymes in newly sequenced genomes and metagenomes, where similarity to characterized proteins is weak, and enzyme discovery campaigns seeking new catalysts of a known reaction. With no structures predicted, the per-pair cost is an embedding lookup plus a small forward pass, and the attention profile nominates residues for mutagenesis.
UNKAI's contribution is methodological: a learned pooling layer over a frozen general-purpose protein language model outperforms the same group's hand-engineered structural features, without manual feature engineering. The authors state plainly that the UNKAI–FUJISAN comparison is not a strict head-to-head benchmark, the two having been evaluated on different datasets and settings. The clearest limitation is extrapolation: under leave-one-EC-subclass-out evaluation the plain E-value baseline outperformed UNKAI on most metrics, and the authors conclude the model still generalizes poorly to unseen functional categories. The release itself is open: the architecture and a prediction script, both trained checkpoints and the labelled pair splits are all published under MIT, though the training loop, the hyperparameter search and the evaluation code are not among them.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.