PLM-ICE

Organisms that survive in the cryosphere depend on ice-binding proteins (IBPs), which fall into two functionally opposed classes: antifreeze proteins (AFPs), which bind nascent ice crystals and suppress their growth, and ice-nucleation proteins (INPs), which template ordered ice formation at comparatively mild sub-zero temperatures. Both classes matter well beyond cold-adapted microbiology — they underpin cryopreservation protocols, frozen-food texture, anti-icing coatings, and atmospheric ice nucleation. Identifying them has traditionally meant chromatography, mass spectrometry, and enzymatic assays, and the sequence-based tools that replaced those assays (AFP-Pred, AFP-PSSM, TargetFreeze, CryoProtect, RAFP-Pred, AFP-LSE, AFP-CMBPred) all target AFPs only.

PLM-ICE is a pair of sequence classifiers that address both classes with one shared design: a frozen ESM-2 encoder feeding a small multilayer perceptron head. It was released in August 2025 by a team spanning Michigan Technological University, The Ohio State University, and the Rochester Institute of Technology. PLM-ICE-INP is, to the authors' knowledge, the first dedicated INP predictor; before it, no computational tool existed for that class at all.

The design choice that distinguishes PLM-ICE from earlier protein-language-model work in this niche is scale of the negative class. Rather than a small curated non-AFP set, the authors sampled negatives from roughly 12 million RefSeq bacterial reference proteins, then screened them against in-house AFP and INP hidden Markov models and BLAST databases to prevent leakage. That matters directly for the intended use case — genome-wide screening — where false positives dominate the error budget.

#Key Features

• Two classifiers, one architecture: Separate AFP and INP models share an identical frozen-encoder-plus-MLP design, so the same pipeline covers both functional classes of ice-binding protein.
• Frozen protein language model backbone: Keeping the pretrained encoder fixed avoids catastrophic forgetting and lets the whole system train on a few thousand labelled sequences.
• Realistically hard negatives: Negatives drawn from a 12-million-sequence bacterial RefSeq pool, filtered by HMM and BLAST screens, make the independent test sets heavily imbalanced and closer to genome-scale conditions.
• Genome-scale validation: Screening complete proteomes of two Pseudomonas species shows the model returns a biologically plausible number of AFP calls where the leading prior method does not.
• First dedicated INP predictor: Fills a gap the AFP-focused literature had left entirely open.

#Technical Details

Sequences are encoded with the smallest ESM-2 checkpoint, esm2_t6_8M_UR50D — an encoder-only transformer with six layers, eight attention heads per layer, and 8 million parameters, pretrained by masked language modelling on UniRef50. Per-residue embeddings from the final hidden layer are average-pooled into a single protein-level vector, which feeds an MLP with two hidden layers of 64 and 32 units (ReLU) and a single sigmoid output. Encoder weights stay frozen; only the head is trained, using binary cross-entropy, the Adam optimizer at a learning rate of 8 × 10⁻⁵, and 25 epochs. Five-fold cross-validation compared the MLP head against SVM, random forest, AdaBoost, and XGBoost alternatives; the MLP won on both tasks, reaching mean MCC 0.984 for AFPs and 0.927 for INPs.

The AFP benchmark comprises 2,061 CD-HIT-filtered positives (50% identity cutoff) drawn from UniProt, five relevant Pfam families, and a structure-similarity search seeded on PDB 3WP9; the INP set has 347 positives. On the published VotePLMs-AFP benchmark, PLM-ICE-AFP reaches 0.956 accuracy against 0.942 for VotePLMs-AFP. On PLM-ICE's own larger, imbalanced independent set the gap widens sharply: 0.991 accuracy and 0.987 balanced accuracy versus 0.036 and 0.401 for a retrained VotePLMs-AFP. For INPs, PLM-ICE reaches 0.987 accuracy and 0.942 sensitivity against 0.290 sensitivity for a VotePLMs architecture adapted to the task.

#Applications

The intended workflow is proteome- or metagenome-scale triage: score every predicted protein in an assembly and shortlist ice-binding candidates for expression and thermal-hysteresis assays. The authors demonstrate this on the reference proteomes of Pseudomonas syringae (5,211 proteins) and Pseudomonas antarctica (6,069), where PLM-ICE flags 87 and 126 AFPs respectively while VotePLMs-AFP flags 2,035 and 2,841. Downstream users include cryobiologists sourcing new antifreeze agents, food scientists, atmospheric researchers studying biological ice nuclei, and microbial ecologists annotating cold-environment genomes.

#Impact

PLM-ICE is a compact demonstration that a frozen small protein language model plus a shallow head can outperform hand-engineered composition and PSSM features on a data-scarce functional classification task — and that the harder problem is often the negative set, not the architecture. Its INP classifier opens a class of proteins that had no predictor at all. Two caveats bound the claims: the work is a preprint and has not been peer reviewed, and the authors note that even at these accuracies the Pseudomonas screens likely still contain false positives, since both genomes are annotated with a single ice-nucleation protein and no antifreeze proteins. Training and evaluation code plus the curated FASTA datasets are released on GitHub under the MIT license; trained checkpoints are not distributed.