The study, published in Frontiers in Chemistry, harnessed ESM3, a generative protein language model from EvolutionaryScale that processes sequences, structures, and functions together. The same model family made headlines when it generated a fluorescent protein so distant from known relatives that it appeared to mimic five hundred million years of evolution. Most AI design pipelines, however, deliberately avoid touching ligand-binding sites, because a single careless mutation can wreck the delicate electrostatics that hold a cofactor in place. They also rarely benchmark their predictions against well-characterized natural mutations. The new work confronts both limitations head-on, using a clinically documented variant as its yardstick.

That yardstick is hemoglobin Fort de France, a naturally occurring human variant first described in 1977, in which histidine 45 in the alpha chain is replaced by arginine. The mutation is known to raise oxygen affinity while preserving structural integrity, a phenotype later confirmed by nuclear magnetic resonance studies. Before generating any models, the researchers mapped the heme-binding pocket and found that histidine 87, the proximal histidine that directly coordinates the heme iron, is untouchable, whereas position 45 is suitable for modification. Crucially, the homologous position in myoglobin, the oxygen-storage protein of muscle, is naturally occupied by arginine, which forms a robust salt bridge with the heme 6-propionate group, stronger than the polar contact mediated by histidine in hemoglobin.

When ESM3 was asked to score the evolutionary fitness of all nineteen possible substitutions at position 45, the wild-type histidine scored zero by definition, while arginine received one of the highest positive scores, 2.6, indicating a change highly compatible with structure and function. The model then generated the engineered variant, dubbed esmHb, using a structure-focused prompting strategy. The team systematically tested how much of the wild-type template structure had to be retained for accurate generation, discovering a nonlinear relationship between retained structural tokens and sequence fidelity, with a retention threshold of roughly 47 contiguous N-terminal tokens needed to halve the local deviation at the mutation site. The optimized protocol achieved something remarkable: a variant carrying exactly one mutation, with one hundred percent sequence identity to the wild-type alpha chain everywhere else, and a local root-mean-square deviation of about one to one and a half angstroms around the engineered residue.

Validation then proceeded through a battery of biophysical simulations. FoldX calculations of folding free energy yielded minus 12.22 kilocalories per mole for the wild-type dimer but minus 18.48 for esmHb, a substantially more favorable value indicating enhanced thermodynamic stability. Monte Carlo refinement with the Rosetta protocol, which explored one thousand trajectories and selected the lowest-energy structures under the Metropolis criterion, consistently assigned lower energy scores to the engineered variant. Molecular docking with AutoDock4 placed the heme-binding affinity of esmHb at minus 11.79 kilocalories per mole, improved over the wild-type value of minus 8.42 and approaching the minus 12.4 of myoglobin itself, the protein whose heme pocket the design was inspired by.

Triplicate molecular dynamics simulations in AMBER, each running five hundred nanoseconds in explicit water at physiological temperature, painted a consistent dynamic picture. The engineered protein relaxed to marginally lower backbone deviations, about half an angstrom, suggesting greater rigidity, while fluctuations in the heme-binding pocket dropped from 4.11 to 3.33 angstroms around position 45. The radius of gyration showed esmHb maintaining a slightly more compact conformation, roughly 19.0 versus 19.5 angstroms, and the heme ligand itself deviated less in the mutant complex. Free energy landscapes derived from principal component analysis revealed perhaps the most striking difference: esmHb samples a single, well-defined energy basin, whereas the wild-type protein wanders across a broader landscape with multiple shallow minima, implying reduced conformational entropy and enhanced kinetic stability for the engineered fold.

Binding free energy calculations from the trajectories reinforced the trend. Using the MM-GBSA and MM-PBSA implicit solvation methods, the researchers found more favorable protein-heme interactions for esmHb, with MM-PBSA reaching statistical significance at p equals 0.03, and stronger protein-protein interactions significant under both methods at p equals 0.04. The authors are careful to interpret the MM-GBSA results as a suggestive trend rather than definitive evidence, a caution that reflects the inherent uncertainty of these estimators. Quantum mechanical calculations added a final layer: density functional theory at the long-range corrected omegaB97XD level, applied to truncated heme-site models extracted from the equilibrated trajectories, showed the formal oxygenation reaction of esmHb to be thermodynamically more favorable by 23.2 kilocalories per mole, with reaction energies of minus 173.1 versus minus 149.9 kilocalories per mole.

The mechanistic story ties the numbers together. The arginine at position 45 forms a salt bridge with the heme 6-propionate, anchoring the cofactor and modulating geminate recombination and oxygen affinity in proportion to its basicity, a dynamic molecular doorstop mechanism first illuminated in myoglobin studies. The computational findings align with experimental data on the natural Fort de France variant, whose increased oxygen affinity and confirmed structural integrity corroborate the DFT results. In effect, ESM3 rediscovered and rationalized a beneficial mutation that nature and clinical hematology had already validated, bridging statistical prediction with empirical observation.

The framework extends well beyond hemoglobin. The authors argue that integrating protein language models with molecular dynamics, docking, and quantum chemistry establishes a generalizable pipeline for engineering therapeutic proteins, applicable to hemoglobinopathies, artificial blood substitutes, enzyme design, and sustainable biomaterials. The limitations are acknowledged with equal candor: ESM3 may inherit biases toward well-represented protein families, its statistical patterns cannot capture subtle quantum effects, and the His45Arg substitution was inspired by myoglobin homology rather than purely de novo generation, highlighting the model’s caution around high-risk changes. Experimental benchmarking of oxygen-binding kinetics and in vivo stability remains essential before any therapeutic claim, since computational metrics do not fully replicate physiological complexities such as allosteric effectors. Still, as a proof of concept, the work marks a milestone in AI-driven protein engineering, demonstrating that a language model trained on evolutionary statistics can safely redesign one of biology’s most critical molecules, one residue at a time.

Subject of Research: Computational protein engineering of hemoglobin using a protein language model

Article Title: Engineering a hemoglobin variant with enhanced stability and oxygen affinity using a protein language model

Article References: Shityakov, S., Novikov, A. S., Skorb, E. V., & Dandekar, T. (2026). Engineering a hemoglobin variant with enhanced stability and oxygen affinity using a protein language model. Frontiers in Chemistry, 14, Article 1898272. https://doi.org/10.3389/fchem.2026.1898272

Image Credits: AI Generated

DOI: 10.3389/fchem.2026.1898272

Keywords: protein language model, ESM3, hemoglobin, protein engineering, molecular dynamics, oxygen affinity, artificial blood substitutes, computational chemistry, Engineering, variant, enhanced, stability