Loading…
Antifreeze protein classifier over frozen ESM-2 embeddings, trained only on sequences whose antifreeze activity was measured in the lab.
No providers recorded yet. Browse all providers
Antifreeze proteins (AFPs) bind the surfaces of nascent ice crystals and depress the freezing point non-colligatively, producing the gap between melting and freezing temperatures known as thermal hysteresis. Organisms from fish to diatoms use them to survive sub-zero conditions, and the same property makes them attractive for organ cryopreservation, frost-tolerant crops and frozen-food texture control.
The obstacle to predicting them computationally has been the labels, not the classifiers. The positive sets AFP predictors train on — the widely used 481-sequence set, and the 8,134-sequence expansion built on it — are assembled by keyword-retrieving UniProt or growing Pfam seeds with BLAST. A sequence enters the positive class because an annotation says "antifreeze," not because anyone measured thermal hysteresis on it, and only a small fraction of either set is backed by an experiment. A model fit to those labels partly learns the annotation pipeline rather than the biology.
AFP-Predictor takes the opposite trade: a far smaller positive class, every member carrying a published measurement. Released in August 2026 by the University of Chinese Academy of Sciences and the Inner Mongolia University of Science and Technology, it is the predictive half of AFP-R, which pairs it with AFP-DB, a database curated by hand from 256 peer-reviewed papers. The model itself is deliberately thin: mean-pooled embeddings from ESM-2, Meta's protein language model, feed a small multilayer perceptron. The pretrained encoder is used off the shelf and never updated, so the only trained component is the classifier head — the frozen-backbone strategy the catalog's PLM-ICE applies independently to ice-binding proteins.
Each input sequence is embedded with esm2_t33_650M_UR50D — the 650-million-parameter ESM-2 checkpoint — and summarized as the mean of its layer-33 hidden states, a single 1,280-dimensional vector per protein. Five multilayer perceptrons consume that vector, each trained on one negative fold plus the full positive set, with activation, solver and learning rate tuned by Optuna over 100 trials against validation AUROC. Training positives are the 153 AFP-DB sequences surviving filters for conflicting reports, short or non-standard sequences and 90% CD-HIT redundancy, split 100 / 25 / 28 across training, validation and internal test. Negatives are 653 sequences sampled from the 333,618 Swiss-Prot proteins left after excluding antifreeze, ice-binding and thermal-hysteresis annotations.
On the balanced internal test set (28 positives, 28 negatives) the ensemble reaches 0.9821 accuracy, 0.9857 sensitivity, 0.9995 AUROC and 0.9651 MCC. On an external set of 164 keyword-retrieved positives and 164 negatives it reaches 0.9756 accuracy, 0.9991 AUROC and 0.9512 MCC, against 0.8994, 0.8445 and 0.8049 accuracy for AFP-LSE, AFP-CKSAAP and AFP-LXGB on the same data. Across the eight variants the classifier mattered far less than the representation: logistic regression, SVM, random forest and MLP land within about a point of each other on the ESM-2 features and all beat their explicit-feature counterparts.
The use is triage before assay work. Cryobiologists prospecting for antifreeze agents, plant scientists screening cold-adapted genomes and food researchers hunting ice-growth inhibitors can score candidates and shortlist them for thermal-hysteresis or ice-recrystallization assays. The AFP-R web server accepts one FASTA sequence at a time and returns a score with an AFP or non-AFP call; the GitHub release runs batches locally from pre-extracted ESM-2 embedding files.
AFP-Predictor demonstrates that in a data-scarce annotation problem, curating the labels can buy more than enlarging the training set. Its coverage is bounded: 153 training positives from 186 validated entries is a narrow slice of AFP sequence space, and recognition of families absent from that curation is untested. The authors are explicit that the external positives are keyword-retrieved "potential AFPs" screened less strictly than the training data: a noisier label set than the model was fit to. The work is a preprint and has not been peer reviewed. Training and prediction code, the labelled sequence CSVs and the trained MLP ensemble are released on GitHub under the MIT license; the ESM-2 backbone is not redistributed and must be obtained from Meta's own release.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.