Kang and Wu et al. engineered a VHH antibody against growth hormone for alkaline chromatography using the Pro-PRIME large language model in a two-round strategy: zero-shot scoring of 45 saturated single-point mutants, then fine-tuned prediction of 20 multi-point mutants, all validated experimentally within two months. Success rates improved from 30% to 35% for alkali resistance and from 77.8% to 100% for thermostability between rounds. The best single mutant improved alkali resistance 44.7% (Tm +6.37Β°C); the best multi-point mutant reached 67.7% improvement (Tm +10.02Β°C) .
Multi-point mutants frequently outperformed the sum of their parts. The A57D;P29T double mutant showed strong alkali resistance despite both single mutations being individually negative, and P29T;G134R reduced EC50 to roughly one-third of wild-type even though each single mutation worsened affinity. The authors interpret this as evidence that LLMs capture epistatic interactions beyond rational design's local-optimum trap . Molecular dynamics on A57D;P29T showed a lower RMSD, increased hydrogen bonding at mutation sites, and no major conformational shift, supporting a rigidity-based stabilization mechanism .
Dynamic binding capacity (DBC) in affinity chromatography showed the wild-type losing 74.1% of residual DBC after 6 hr of 0.5 M NaOH and retaining only 15.2% after 24 hr; designed mutants retained 60-90% after 6 hr and over 50% after 24 hr. Yield was sustained through over 140 chromatography cycles versus ~60 for WT . Scale is stated at 5000 L with reported cost savings exceeding $1 million/year.
Conflicts of interest: Five co-authors are employees of Changchun GeneScience Pharmaceuticals, which owns the VHH patent; patents were filed on all mutations in 2023 . Sample size: 65 tested mutants is modest for generalization claims; generalizability to other proteins or scaffolds is untested. Correlation gaps: Spearman correlation between alkali resistance and Tm of single mutants was only -0.29, implying the two properties are weakly coupled and fine-tuned prediction of multi-point mutants relied on very small training sets. Comparison gap: no head-to-head benchmark against rational-design or high-throughput-screening control arms is reported, so relative AI advantage is inferred rather than directly measured . Reproducibility: source data provided for Figures 1, 3, 4; methods are documented in sufficient detail for replication of assays, though the Pro-PRIME fine-tuning hyperparameters are not fully specified.
Note: "Designed mutants" values represent the reported range midpoint (60-90% and >50% bands); error bars reflect the spread of the stated ranges, not measured SD.
Independent replication outside GeneScience facilities, benchmarking against a rational-design control arm, and extension to a second protein scaffold would test whether the efficiency claims generalize or are scaffold-specific. A failed replication of the 140-cycle chromatography durability claim would undermine the central industrial narrative.
Know what changed, what holds up, and what remains uncertain. Every Friday. No ads.