Yes — machine learning can now design molecules that inhibit ice, and it has been demonstrated: in 2023 a team computationally designed novel ice-binding proteins from scratch and confirmed their ice-recrystallization activity in the lab. But "AI designs antifreeze" overstates what the tools do. They generate and rank candidates; a wet lab still synthesises and assays every one, and they cannot yet model the sugar chemistry that makes natural antifreeze glycoproteins so potent.
The gap between the headline and the reality is where the interesting engineering lives. This post walks through what the ML toolkit actually does for ice-control molecule design, the published proof points, the hard limits that keep humans and wet labs in the loop, and how DeepSnow uses these methods to design its R&D-stage DS-100 polypeptides around existing intellectual property.
Key takeaways
- ML tools generate and rank candidate ice-binding sequences; they do not replace synthesis and assay. The wet lab is still the arbiter.
- De novo computational design of ice-binding proteins with confirmed activity has been demonstrated (de Haas et al. 2023).
- Machine learning has also surfaced novel small-molecule ice-recrystallization inhibitors (Warren, Gibson & Sosso 2024).
- Current models do not natively handle glycosylation — the sugar decoration central to natural antifreeze glycoproteins — a real limit on the design space.
- DeepSnow's DS-100 is R&D-stage: AI-proposed Ala/Glu polypeptides, wet-lab synthesised and assayed, engineered around University of Utah patent WO2024258965A2, not into it. Performance figures are lab-scale.
Can AI design antifreeze proteins?
Yes, within limits that matter. Researchers have used de novo protein-design methods to create ice-binding proteins that did not exist in nature and then confirmed, experimentally, that they inhibit ice recrystallization. So the core claim — that computation can propose a working ice-active molecule — is established. What is not true is that the software designs a finished, manufacturable product end to end without wet-lab validation.
The proof point is de Haas et al. 2023 in PNAS, which reported computationally designed ice-binding proteins with confirmed ice-recrystallization-inhibition (IRI) activity. It is the clearest published demonstration that the "design-first" paradigm — propose a structure computationally, then build and test it — works for ice control, not just for the enzymes and binders where de novo design first proved itself.
What does the ML toolkit actually do?
It compresses the search. The design space of possible polymer and polypeptide sequences is astronomically large; the tools narrow it to a shortlist worth synthesising by predicting structure, ranking likely ice-binding activity, and filtering for manufacturability. Each stage is a separate model doing a narrow job, and the output is candidates, not conclusions.
The workflow, in the terms practitioners use:
- Structure prediction and generation — tools such as RFdiffusion generate novel backbone geometries, and structure predictors evaluate whether a sequence will fold into the intended shape (for ice binding, a flat, ordered face that matches the ice lattice).
- Sequence design — ProteinMPNN and related models propose amino-acid sequences that stabilise a target backbone.
- Property prediction — protein language models (the ESM family) and polymer-property models (such as polyBERT) rank candidates by predicted properties before anyone makes them.
- Small-molecule search — for non-polymer IRIs, ML screens chemical space; Warren, Gibson & Sosso 2024 in Nature Communications used machine learning to discover new small-molecule ice-recrystallization inhibitors.
None of these is an "antifreeze designer." Each is a general tool pointed at the ice problem, which is why domain knowledge — what actually binds ice — is what turns the toolkit into results. The mechanism those results are optimised for is explained in what is ice recrystallization inhibition.
What can't the models do yet?
They cannot natively design glycosylation, they cannot reliably predict IRI potency from sequence alone, and they cannot certify that a candidate is synthesisable or stable at scale. These are not minor caveats — they are the reasons the wet lab remains the decision-maker rather than a rubber stamp.
The concrete limits:
| Limit | Why it matters | |---|---| | No native glycosylation modelling | Natural antifreeze glycoproteins owe much of their potency to attached sugars; most protein-design tools model the peptide backbone only. | | Weak sequence-to-IRI prediction | Ice-recrystallization potency depends on subtle surface geometry and dynamics that current predictors capture imperfectly. | | Manufacturability not guaranteed | A high-scoring sequence may be impossible to polymerise cleanly, insoluble, or unstable. | | Assay is ground truth | Predicted activity must be confirmed by splat assay and freeze–thaw testing before it means anything. |
The glycosylation limit is the sharpest. Because the models work on the polypeptide backbone, the synthetic-antifreeze-glycoprotein field has largely moved toward non-glycosylated polypeptides — alanine/glutamate copolymers that capture much of the IRI activity without the sugars. That is a deliberate design-around of a modelling gap, and it happens to be manufacturable, which is why it underpins DS-100. The role of alanine specifically is covered in why alanine is the smallest antifreeze molecule.
How potent is the resulting chemistry?
Very — the published synthetic antifreeze polypeptides reach IRI performance that competes with natural proteins, and they do it with cheap, scalable amino acids. The best measured results come from non-glycosylated Ala/Glu polypeptides, and they scale with chain length, which is exactly the kind of design knob an ML pipeline can explore systematically.
The peer-reviewed numbers to anchor on:
- Deleray, Saini, Wallberg & Kramer 2024 in Chemistry of Materials reported synthetic AFGPs made by NCA polymerization achieving 89% mean-grain-size (MGS) reduction for a 28-mer, 94% for a 57-mer, and 97% for a 170-mer — with alanine identified as the key residue.
- McPartlon et al. 2025 in Advanced Materials described an "ultrapotent, ultraeconomical" antifreeze polypeptide built from inexpensive Ala/Glu copolymers.
- Ampaw et al. 2022 in JPCL showed α-alanine is the smallest known IRI at just 13 atoms — the molecular starting point the larger polypeptides scale up.
That 89–97% range, not a single figure, is the honest way to state synthetic-AFGP performance: potency rises with chain length, and it is measured at the bench. It is also worth separating IRI from thermal hysteresis — two distinct ice-control mechanisms that are often conflated, as we explain in thermal hysteresis versus IRI.
How does DeepSnow use these methods?
DeepSnow runs a closed loop: the discovery engine proposes candidate polypeptide sequences and ranks them by predicted ice-binding, IRI potency, and manufacturability; the wet lab synthesises and assays the shortlist; and the results retrain the models. A fourth constraint is IP-awareness — candidates are screened against existing patents so the pipeline is engineered around prior art, not into it.
That prior art has a specific owner worth naming correctly. The foundational synthetic-AFGP patent, WO2024258965A2, is assigned to the University of Utah Research Foundation (inventors Kramer and Deleray) — not, as press coverage sometimes implies, to a single commercialization startup. DeepSnow's design objective is to explore the Ala/Glu sequence space for variants that deliver comparable IRI outside those claims. The full pipeline framing is in the DeepSnow platform.
The reason this is a platform and not a one-off is that the same loop generalises. A model trained to rank ice-binding candidates for snowmaking additives ranks them for the adjacent cold-chain, food, and cryopreservation problems too — the same physics, different specification.
Why does the wet lab stay in the loop?
Because prediction and reality still diverge, and the assay is where that divergence is caught and fed back. A model can rank a sequence highly on predicted ice-binding and still be wrong about its potency, its solubility, or whether it polymerises cleanly. Every candidate that clears the models is synthesised and measured before it counts, and those measurements are what make the next round of predictions better.
The standard bench test is the splat assay: a thin film of ice micro-crystals is annealed at a fixed sub-zero temperature for a set time, and the growth in mean grain size is measured against an untreated control. A strong IRI holds crystals small; a weak one lets them coarsen. Because grain growth is exactly what damages tissue, texture, and snow structure, the splat assay measures the property that matters commercially, not a proxy for it. That closed loop — predict, synthesise, assay, retrain — is the difference between a model that lists plausible molecules and a program that ships a validated one. It is slower than a pure in-silico pipeline, and that is the point: the slowness is where the errors get removed.
The bottom line
AI can design antifreeze — the demonstrations are real and the resulting chemistry is genuinely potent — but the accurate version of the claim is narrower and more useful than the headline. The tools generate and rank; the wet lab decides; glycosylation remains outside their reach, which is why the field builds non-glycosylated Ala/Glu polypeptides that reach 89–97% MGS reduction and can actually be manufactured. Used honestly, machine learning is a search accelerator over a vast design space, not an autonomous chemist. That is precisely how DeepSnow's DS-100 program uses it.
If you are working on ice control — in snowmaking, cold chain, food, or cryopreservation — and want to compare notes on the discovery approach, get in touch or join the waitlist.
DS-100 is in R&D. Performance figures cited are lab-scale results from the peer-reviewed literature and DeepSnow's own assays; they are not commercial product claims. DeepSnow is the platform brand of SnowLabs Limited (Ireland); DeepSnow Srl (Italy) is in formation.