Finding proteins that separate rare earth metals, with Omnii-1

Rare earths frequently occur together in mineral deposits, and their similar chemistry makes separating individual elements difficult. Traditional chemical and mechanical methods often require the use of harsh acids or large amounts of energy²³, making processing costs a major factor in where refining takes place. Lower energy and processing costs, alongside sustained investment, have helped establish China as the dominant center of rare-earth refining.

Bio-based solutions started gaining attention relatively recently. In 2014, researchers identified a bacterium from an Italian volcanic mudpot whose methanol dehydrogenase requires a lanthanide (a rare earth element) instead of calcium⁴. Similar enzymes have since been found in many bacteria. In 2018, researchers discovered lanmodulin, a protein that binds lanthanides selectively⁵. Some engineered variants have shown the ability to separate rare earths in water at room temperature, without organic solvents in the separation step⁶.

Lanmodulin from Hansschlegelia quercus (Hans-LanM) can distinguish neodymium from dysprosium, two rare earth elements used in permanent magnets. The protein responds to small differences in ionic radius through a shift in a coordinating carboxylate group. In a single separation stage, Hans-LanM enriched dysprosium more than sixteenfold, from 5% to 83% purity, while recovering over 98% of it. A single mutation at the dimer interface increased purity to above 98%⁷. Immobilized lanmodulin has also separated light and heavy rare earths from coal fly ash leachate⁸ (Figure 1).

Proteins distinguish metal ions by charge, size, and coordination chemistry: which atoms they bind and how those atoms are arranged. Zinc, for example, often favors four nitrogen or sulfur donors in a tetrahedral geometry, while magnesium typically binds six oxygens in an octahedral arrangement.

Lanthanides are harder to distinguish (Figure 2). Under typical binding conditions, they are trivalent, favor oxygen donors such as carboxylates, and generally coordinate eight or nine oxygen atoms, leaving ionic radius as a major basis for selectivity.

Across the entire lanthanide series, from lanthanum to lutetium, ionic radius decreases by just 0.183 Å for eight-coordinate ions⁹ — less than the roughly 0.23 Å difference between magnesium and calcium. Neodymium and dysprosium differ by only 0.082 Å, and neighboring lanthanides by a mere 0.013 Å, roughly 1% of their radius. Lanmodulin binds these ions through EF-hand motifs, protein motifs in which a loop of amino acids positions oxygen atoms around a metal ion. These motifs typically bind calcium, but in lanmodulin they strongly favor lanthanides (Figure 3).

Strong binding alone does not separate lanthanides. A protein must favor one ion over another despite their nearly identical chemistry. Selectivity depends on differences in binding free energy, not only on how closely an ion fits the pocket (Figure 4).

Those differences reflect several competing effects. Binding displaces water from the ion's hydration shell, rearranges coordinating side chains, and can alter the protein's conformation. The pocket can accommodate different ionic radii by shifting its coordinating groups or, in some proteins, retaining water molecules. Hydrogen bonds beyond the immediate binding site further influence which arrangements are favored (Figure 5).

In Hans-LanM, for example, a small shift in metal coordination alters a hydrogen-bond network involved in protein dimerization, amplifying differences between lanthanides. Such interactions are difficult to predict from structure alone: small changes outside the coordination shell can alter selectivity without substantially changing the metal's immediate environment. Binding affinity and selectivity must therefore be measured, and neither alone establishes how well a protein will perform in a separation process.

In early 2026, experts at Lawrence Livermore made a major leap in laboratory innovation to speed up the exploration of how specific proteins bind to rare earth targets. The lab announced that its scientists developed

a high-throughput platform that screens proteins for their rare-earth element binding preferences at unprecedented scale […] This new technique — called SpyTag-Catcher Immobilization of Lanmodulin for Assaying Metal-Binding Selectivity (SpyCI-LAMBS) and pronounced 'spicy lambs' — drastically cuts the time required for these experiments and enables data collection that can be used to train machine learning models for protein design.¹⁰

In the original assay, each protein was immobilized in a separate well and exposed to all 15 rare earth elements (REEs) simultaneously, the 14 stable lanthanides plus yttrium. The bound metals were then released and quantified by mass spectrometry. Using this method, LLNL researchers screened 621 lanmodulin orthologs and identified eight distinct selectivity profiles, as reported in Nature Chemical Biology¹¹.

The goal of our collaboration is to design proteins for rare-earth separation through iterative cycles of generation, experimental verification, and model improvement. We're starting by building better selectivity verifiers to identify which generated candidates are worth testing, as validation is costly and time consuming and we hypothesize that models better at predicting selectivity are also better at generating selective binders.

For our evaluation, LLNL provided measurements from 1,331 wells across four plates: 198 de novo designs and 1,133 natural sequences. Of the natural proteins, 609 bound enough metal for reliable measurement, spanning 292 sequence clusters. These 609 proteins form our dataset for evaluating binding preferences across the 15 REEs.

Because the assay exposes each protein to all 15 REEs simultaneously, it produces a relative binding profile rather than absolute affinities for individual elements. We therefore frame prediction as a contrastive task: given two elements, such as neodymium and dysprosium, can Omnii-1 rank unseen proteins by how strongly they favor one over the other? We quantify each protein's preference using differences in log-ratio measurements, then compare predicted and measured rankings using Spearman correlation. We repeat this across all 105 possible element pairs and report the median correlation.

We compared three classes of models on their ability to predict how strongly an unseen protein favors one rare earth element over another, using identical evaluation splits. Protein language models included ESM-2¹², ESM-C¹³, ESM-3¹⁴, and ProGen2-XL¹⁵. Structure-based models included AlphaFold2¹⁶, Boltz-2¹⁷, and SaProt¹⁸. We also tested two genome language models: Evo2¹⁹, which uses nucleotide sequences, and Omnii-1²⁰, using its amino-acid representations.

We compared the models against five baselines, including three based on sequence homology. The simplest identifies related proteins using short amino-acid matches. The other two use profile hidden Markov models (HMMs)²¹: one fits a regression to the full matrix of search scores, while the other assigns each held-out protein the measured selectivity profile of its closest match. Amino-acid composition and random noise serve as additional controls. We kept all pretrained encoders, including Omnii-1, frozen and evaluated them using the same regression procedure.

We report performance as a rank correlation. For each pair of elements, we rank the proteins by how strongly a model predicts they favor one element over the other, then compare that ranking with the measured one.

Without any pretraining on rare-earth binding data, Omnii-1 achieved a median rank correlation of 0.340 across all 105 element pairs, compared with 0.254 for the strongest homology baseline, 0.245 and 0.198 for the other two, 0.188 for amino-acid composition and 0.020 for noise. The strongest public model, ESM-C 6B, scored 0.318. Omnii-1 outperformed it across all ten evaluation splits (Figure 6).

Omnii-1 also outperformed all three homology baselines on every element pair. Compared with the two profile HMM baselines, its advantage was greater for the 17 pairs closest in ionic radius (0.117 and 0.134) than across all 105 pairs (0.094 and 0.086). Omnii-1 therefore predicts selectivity more accurately than the homology methods tested, including for the most similar elements.

For the 17 closest REE pairs, separated by 0.02 Å or less, Omnii-1 and ESM-3 performed similarly. Their score difference was 0.006, with a paired error of 0.005, and Omnii-1 led in four of ten evaluation splits. Given the small number of pairs and greater measurement noise, we do not consider this sufficient to rank the models.

Structure-based models generally performed worse than the leading protein language models. AlphaFold2 scored 0.191, below the simplest sequence-similarity baseline of 0.198. SaProt, which reads the same structures through a structure-aware vocabulary, scored 0.217 across all 105 pairs but 0.176 on the 17 closest in ionic radius, so it clears that baseline only where the elements are easy to tell apart. Boltz-2 scored 0.262 when co-folded with individual lanthanides, compared with 0.278 using its protein representations alone. Explicitly modeling the bound ion therefore provided no improvement in this evaluation. Evo2, evaluated on nucleotide sequences, scored 0.133, the lowest of the pretrained models tested.

Model size helped, but gains diminished at larger scales. Across the two families tested at multiple sizes (ESM-2 from 35M to 15B parameters, ESM-C from 300M to 6B), the largest models outperformed the smallest by approximately 0.055, consistently across all ten evaluation splits. Model size alone did not determine performance.

Overall, Omnii-1 predicts rare-earth selectivity more accurately than the other models and homology methods tested, despite having no training on rare-earth binding data. This suggests strong potential for further post-training on experimental measurements to improve its predictions and protein designs.

Designing proteins that fold, bind lanthanides, and distinguish between them is challenging. Our hypothesis is that models better at predicting selectivity will also be better at generating selective binders. A model that captures the sequence features governing metal preference should be better equipped to design proteins with those properties. Here we test the first part of that assumption; experimental validation of our generated candidates through LLNL will test the second.

In the next step of this collaboration, we will use those experimental measurements to further improve Omnii-1's generation quality, as we have done in projects such as antibody design using partner assay data.

Whole branches of society's technology draw from or mimic nature, from enzymatic synthesis and aviation to cancer therapeutics and Velcro. For many of the materials and processes that underpin the modern economy, nature has already evolved solutions that are stronger, more efficient, or more sustainable than those we can engineer today. Better harnessing this vast catalog of evolutionary innovation could transform nearly every aspect of modern life.

The challenge is understanding and engineering these biological systems. At Radical Numerics, we're building general biological intelligence to read and write the sequences and structures of life, with the goal of making nature's innovations accessible to engineering.

Our collaboration with Lawrence Livermore National Laboratory to design proteins for rare-earth extraction is our first industrial application of that technology. Here, Omnii-1 predicts subtle differences in rare-earth selectivity that sequence homology cannot reliably capture. We hypothesize that this same ability to discriminate between biological properties will translate into better generative design.

For the pressing problem of shoring up supplies of rare earth elements and products that depend on their unique characteristics, we aim to continue advancing and applying this line of work with Omnii.

We see this as the beginning of a much broader opportunity: using models and experimental feedback to unlock biological solutions to difficult industrial problems. If you're working on one, particularly with an assay that could guide protein design, we'd like to hear from you.

Partners can bring target materials, protein characterization data, and experimental testing capabilities. We can use Omnii-1 to design and optimize proteins for binding and selectively recovering rare earth elements, advancing more efficient and sustainable biological extraction technologies.