A computational antibody engineering pipeline coupling a pretrained nanobody language model with structure-based binding free energy prediction and all-atom molecular dynamics improves the discovery of higher-affinity single-domain antibodies against viral antigens without requiring initial target-specific model training. Described by researchers in Genome Biology, the open-access workflow, designated LmPro-Nano, addresses long-standing bottlenecks in nanobody affinity maturation, where experimental screening remains resource-intensive and conventional computational pipelines often obscure the distinct contributions of individual design stages. Nanobodies, which represent the single variable heavy-chain domains derived from camelid heavy-chain-only antibodies, offer distinctive advantages over conventional immunoglobulins due to their low molecular weight, extended third complementarity-determining regions, distinctive framework architectures, and convex paratopes capable of accessing recessed epitopes. By evaluating each computational module against independent empirical benchmarks, the study demonstrates that sequence-driven localization of antigen-binding residues and geometric free-energy scoring can enrich candidate libraries for tighter binders across clinically relevant coronavirus targets, including the receptor-binding domain of severe acute respiratory syndrome coronavirus 2.
The structural motivation for specialized single-domain computational design stems from the distinctive molecular characteristics of camelid heavy-chain fragments. However, the rapid emergence of viral escape variants, such as the SARS-CoV-2 Omicron lineage JN.1 that predominated during the investigation, creates an urgent operational demand for design workflows that can be rapidly redeployed directly on updated antigen sequences without redesigning therapeutic scaffolds from scratch. While diverse computational methodologies have emerged—spanning iterative graph neural networks, equivariant sequence-structure generators, and inverse folding protocols—they frequently evaluate candidates purely by sequence recovery against native backbones, which conflates sequence memorization with physical binding improvement and leaves the isolated impact of individual pipeline components unresolved.
To overcome this limitation, the authors structured LmPro-Nano around an architectural core called Lm-Nano, a foundational nanobody language model whose encoder is frozen after dual-phase training. The initial pretraining phase exposed the model to unpaired nanobody sequences obtained from the Integrated Nanobody Database for Immunoinformatics, masking amino acid residues exclusively across the three CDR loops while preserving framework positions. To preserve length transferability, the transformer encoder incorporates multi-head attention equipped with rotary position embeddings, which calculate attention weights based on relative residue separation rather than fixed coordinate indices. Following masked-language pretraining, Lm-Nano underwent structural fine-tuning on protein complexes cataloged in the nanobody subset of the Structural Antibody Database, termed SAbDab-Nano, to impart antigen-conditioned representations. The resulting frozen encoder subsequently powers two distinct functional heads: a generative decoder built from three residual multilayer perceptron layers designed to propose CDR3 sequences alongside coarse-to-fine loop conformations, and a paratope classification head trained to assign per-residue binding probabilities across the entire domain using nanobody and antigen sequences alone.
Downstream candidate selection in LmPro-Nano relies on ProFE-Net, a specialized neural network formulated to predict the change in Gibbs free energy of binding resulting from single amino acid substitutions. A residue-level fusion module combines these representations, processing wild-type and mutant sequences over shared backbone coordinates in both forward and inverted orientations to enforce mathematical anti-symmetry. Benchmark evaluations on 4,947 single-point mutations derived from the RDE-PPI complex dataset revealed that combining ProFE-Net with the geometric baseline RDE-Net at an uncalibrated equal weighting yielded superior rank correlation and reduced prediction error compared to either model operating independently, providing a robust metric for prioritizing candidate variants before conducting explicit-solvent simulations.
The generative loop redesign capacity of the workflow was evaluated on 3-2A2-4, a therapeutic nanobody known to neutralize diverse sarbecoviruses and Omicron subvariants whose crystallographic structure in complex with the ancestral SARS-CoV-2 receptor-binding domain had been resolved under Protein Data Bank accession code 7X2L. When tasked with redesigning the CDR3 loop across an iterated best-of-N resampling cycle filtered by strict biophysical and naturalness thresholds, Lm-Nano proposed an optimized variant carrying three coordinated substitutions: E95N, N96Y, and F100G, converting the parental loop sequence LENGGFFYY into LNYGGFGYY. Notably, the chosen candidate exhibited the joint-highest generator perplexity in the elite sampling pool, demonstrating that candidate selection was governed by downstream energetic screening rather than sequence likelihood memorization. The structural stability and energetic affinity of this redesigned loop were subsequently interrogated across multiple evolutionary backgrounds, providing a rigorous test of the model’s generalizability against divergent viral targets.
To assess the physical consequences of the redesigned 3-2A2-4 loop, the authors performed all-atom explicit-solvent molecular dynamics simulations of both parental and mutated complexes in duplicate 500-nanosecond trajectories across three distinct receptor-binding domains: the ancestral strain, BA.2.86, and JN.1. Energetic analysis using molecular mechanics/Poisson–Boltzmann surface area and generalized Born surface area methods indicated consistent thermodynamic stabilization for the optimized variant over the most settled 50-nanosecond segment of each trajectory, recording favorable binding energy increments of 3.60 kilojoules per mole under MM/PBSA and 1.88 kilojoules per mole under MM/GBSA. Component decomposition revealed that this gain was driven primarily by favorable desolvation rather than increased direct contact interactions; the engineered variant sacrificed 184 kilojoules per mole in electrostatic interaction energy and 16 kilojoules per mole in van der Waals contacts, but compensated by paying 199 kilojoules per mole less in polar solvation penalty. This thermodynamic shift coincided with the removal of an apex carboxylate group via the E95N substitution, allowing an adjacent lysine residue to form stable interface salt bridges, shifting contact frequency from 0.008 to 0.212.
The localization capability of LmPro-Nano was independently tested on VHH72, a nanobody directed against the receptor-binding domain for which comprehensive deep mutational scanning data were available. By leveraging AlphaSeq empirical binding affinities encompassing 1,579 single substitutions across 89 positions, the research team could benchmark computational predictions against quantitative experimental values without prior training contamination. Given only the primary sequence of VHH72 and the receptor-binding domain, the frozen Lm-Nano paratope head scored all 125 positions of the nanobody without restricting its focus to canonical CDR loops. The top 17 predicted positions captured 13 of the 17 structurally verified interface residues within an 8-angstrom C-beta distance threshold, achieving a per-position area under the receiver operating characteristic curve of 0.9771. Framework residues contributed substantially to the localized binding interface, reflecting the functional enlargement of framework region 2 characteristic of single-domain antibodies in the absence of light-chain pairing.
Following interface localization, the team exhaustively enumerated all 18 non-parental amino acid substitutions across the 17 identified positions, yielding 306 candidates, of which 305 possessed empirical binding measurements. Ranking these candidates with the combined ProFE-Net and RDE-Net scoring ensemble identified a top-ten candidate panel in which seven variants exhibited higher experimental binding affinities than the parental clone, representing a 2.57-fold enrichment over the baseline rate of the candidate pool. The highest-ranked candidate, a framework substitution designated T61Y, yielded the largest empirical affinity gain in the dataset, enhancing binding by 2.92 kilojoules per mole and conferring an affinity 3.2 times greater than parental VHH72. Crucially, when the same scoring ensemble was applied across all 1,579 mutational records without the sequence-only paratope gate, only two of the top ten candidates bound more tightly than parental wild-type, demonstrating that sequence-based paratope localization serves as a vital gating filter to exclude non-interface artifacts that energy scoring functions alone fail to penalize.
The ten prioritized VHH72 variants, alongside parental VHH72 and a neutral control substitution designated G42I, were subjected to rigorous molecular dynamics characterization involving 36 independent 500-nanosecond trajectories modeled using the CHARMM36 force field and explicit TIP3P solvent. Trajectory analysis over the most stable 50-nanosecond windows demonstrated that eight of the ten designs achieved more favorable binding enthalpies under MM/PBSA than the parental complex, with a paired mean shift of 13.7 kilojoules per mole in the favorable direction. Electrostatic and polar solvation terms exhibited large compensatory shifts of up to 260 and 226 kilojoules per mole, particularly among mutations altering formal charge. MM/GBSA calculations corroborated these findings across compatible dielectric parameters, yielding an average favorable shift of 15.7 kilojoules per mole. The simulations confirmed that physical modeling in explicit solvent provides effective within-position discrimination, refining candidate prioritization where heuristic energy scores primarily distinguish between residue positions rather than subtle amino acid variations.
To evaluate broader applicability beyond coronavirus antigens, the investigators tested ProFE-Net and the ensemble scorer on external surface plasmon resonance binding datasets covering green fluorescent protein, tumor necrosis factor alpha, and interleukin-23 complexes. On the green fluorescent protein variant panel, the equal-weight ensemble achieved the highest rank correlation among tested models, while standalone ProFE-Net proved superior on cytokine complexes, capturing the tightest-binding variant in one cytokine panel where comparative baselines failed. The authors note, however, that current implementations rely on surrogate energy scoring and simulated structural ensembles, necessitating future wet-lab experimental campaigns to validate fully prospective designs across broader non-viral targets, including G protein-coupled receptors. Overall, LmPro-Nano establishes a modular, verifiable paradigm for antibody engineering, showing that coupling sequence-only language models with physical energy scoring offers a practical route toward rapid, targeted nanobody affinity optimization.
Subject of Research: Computational nanobody optimization using pretrained language models and binding free energy prediction
Article Title: LmPro-Nano: a method for nanobody optimization using pretrained language and energy prediction models
Article References: Sun, B., Cao, S., & Gao, F. (2026). LmPro-Nano: a method for nanobody optimization using pretrained language and energy prediction models. Genome Biology. https://doi.org/10.1186/s13059-026-04355-w
Image Credits: AI Generated
DOI: 10.1186/s13059-026-04355-w
Keywords: LmPro-Nano, nanobodies, single-domain antibodies, SARS-CoV-2, protein language models, binding free energy, ProFE-Net, molecular dynamics, MM/PBSA, VHH72, 3-2A2-4, paratope localization, deep mutational scanning, receptor-binding domain, structural biology, computational antibody design, rotary position embeddings
News Source: Drew Townsend. (November 21, 2026). Computational Workflow Optimizes Nanobody Binding Using Sequence Models and Molecular Simulations. Scienmag.



