EpHod

An industrial enzyme run two pH units from its catalytic optimum can lose most of its activity, so where on the pH scale a candidate actually works is a routine screening question in biocatalysis. The usual computational shortcuts answer a different question: a protein's isoelectric point correlates poorly with the pH at which it catalyses, and a PROPKA-style calculation locates the pH where the fold is most stable rather than where the chemistry happens. EpHod predicts the catalytic optimum (pHopt) directly from sequence, learning it from measured values instead of from a charge heuristic.

The model emerged from a systematic method sweep run at the National Renewable Energy Laboratory with collaborators at Harvard Medical School, the Broad Institute and Dana-Farber Cancer Institute. The authors trained 11,550 model instances, crossing sequence representations — one-hot encodings, amino acid composition, engineered iFeature descriptors, and embeddings from five protein language models — with classical regressors and four neural architectures. Language model embeddings dominated: the best method without them reached a sample-weighted R² of 0.415 on held-out sequences below 20% identity to training, while a support vector regression on averaged ESM-1v embeddings reached 0.644. Embeddings from the convolutional CARP model performed comparably to those from transformers.

The binding constraint is data, not architecture: BRENDA yields only 9,855 enzymes with a usable measured pHopt. EpHod works around that with a labelled proxy corpus — 1.9 million secreted bacterial proteins tagged with their source organism's optimal growth pH — trained on first, then fine-tuned on the catalytic labels.

#Key Features

• Two heads on one frozen backbone: predictions average a residual light-attention network over per-residue embeddings with a support vector regression over the same embeddings mean-pooled. The language model is frozen throughout, so only the heads are trained.
• Proxy-label transfer learning: environmental pH labels from 1.9 million secreted bacterial proteins stand in for the scarce catalytic measurements during the first training stage.
• Loss reweighting for the tails: sample weights inversely proportional to label density push the model onto the acidic and alkaline enzymes biotechnology cares about, improving F1 by 26% and 250% in those bins.
• Per-residue attention weights: the light-attention layer exposes which residues drove a prediction; those weights track solvent accessibility and proximity to the catalytic centre.
• Reusable embeddings: inference optionally writes a 2560-dimensional final-layer representation per sequence, a pH-aware feature set for downstream supervised models.

#Technical Details

EpHod runs ESM-1v frozen and takes 1,280-dimensional per-residue embeddings from its final layer. The neural head applies two parallel width-7 convolutions — one producing values, one producing softmax attention weights — concatenates the attention-weighted sum with a max-pool into a 2560-dimensional vector, then passes it through four residual dense blocks to a scalar output. That head holds roughly 49 million parameters, on top of the 650-million-parameter frozen backbone. Training data came from BRENDA (9,855 enzymes split 7,124 / 760 / 1,971 after MMseqs2 clustering at 20% identity) and from BacDive and NCBI RefSeq (1.9 million secreted proteins, filtered with SignalP 6.0 and declustered against the pHopt test set). EpHod reaches an RMSE of 1.25 pH units on the held-out test set, and a sample-weighted R² of 0.662 on the 999 test sequences below 20% identity to training — ahead of either component alone, at 0.648 for the neural head and 0.644 for the support vector regression. RMSE rises by less than 0.12 pH units across the identity range. Inputs longer than 1,022 residues are truncated, and inference takes roughly 7 seconds per sequence on CPU or 0.1 seconds on GPU.

#Applications

The intended use is triage before wet-lab work: scanning genome- or metagenome-scale homolog sets for enzymes that already tolerate acid or alkali, rather than engineering that tolerance from scratch. Predictions come from sequence alone, with no structure and no alignment, so the model scales to large candidate libraries — plastic-degrading and biomass-processing enzymes among them, where a process stream's pH is a hard design constraint. Residues carrying high attention weights, and point mutants whose predicted pHopt shifts sharply from the wild type, give a shortlist for site-saturation mutagenesis.

#Impact

Against biophysical baselines EpHod is a clear improvement, separating acidic from alkaline enzymes far better than isoelectric point, surface charge balance, or a PROPKA3 stability calculation, on both AUC and correlation with measured values. The honest limits sit in the same numbers: an RMSE of 1.25 pH units is coarse relative to the differences that matter in process design, and leave-one-group-out cross-validation by enzyme commission class shows error climbing to 2.48 pH units for activity classes absent from training — a regime where the simple support vector regression generalises better than the neural head. OphPred, a separately published method, achieves lower unweighted mean absolute error, which the authors attribute to the abundance of near-neutral enzymes in the evaluation data. Code is MIT-licensed, and the weights and both training datasets are archived on Zenodo under CC BY 4.0.