7. Protein Engineering Applications

Conformation-Selective FAK Modulators: De novo-designed ~100-residue miniproteins from the Baker lab that selectively activate or inhibit Focal Adhesion Kinase (FAK) by stabilizing distinct conformations of its catalytic domain, rather than competing with ATP at the active site. Designed via RFdiffusion, ProteinMPNN, and AF2 evaluation on six FAK crystal structures, lead inhibitors blocked FAK turnover with low-nanomolar IC50 values (22 and 17 nM) and high kinome selectivity, while lead activators enhanced activity over 2-fold, validated by a 2.04 Å co-crystal structure. In cancer cells, inhibitors impaired 3D-spheroid viability while activators accelerated cell spreading, and partial diffusion re-targeted the scaffolds to suppress Src kinase signaling.
Posted: July 10, 2026
Preprint

NovoTags: Small (13–16 kDa) de novo-designed proteins from the Baker (IPD), Lavis (Janelia), and Mahamid (EMBL) labs that selectively bind three Janelia Fluor dyes (JF494, JF596, JF657) with the brightness and orthogonality of synthetic dyes. Designed via Cα RFdiffusion, LigandMPNN, and Rosetta/AF2 filtering followed by yeast-display screening (~3.8% hit rate across 16,675 designs), binders achieve 1.5–19 nM affinity and >1,000-fold selectivity over the alternative dyes; the NovoTag657 crystal structure matched the design within 0.6 Å RMSD. Demonstrated uses include multiplexed STED imaging, tunable fluorescence lifetime, a covalent binder, a dimerization split system, and a fluorescent proximity sensor.
Published: July 16, 2026
Paper | Code (GitHub)

AI-redesigned starting points and outcomes enhance protein evolution: A study from David Liu's lab combining ProteinMPNN-based computational redesign (Baker lab) with PACE/PANCE directed evolution to test whether redesigned enzymes make better evolutionary starting points than wild type. ProteinMPNN redesigned three botulinum neurotoxin proteases on fixed backbones, yielding stable, active variants up to 2.8-fold more catalytically efficient than wild type. Evolving a redesigned BoNT/E variant alongside wild type across substrates of increasing difficulty, redesigned-derived proteins reached higher activity, revealing an asymmetric fitness landscape (epistasis) where redesign-specific solutions were inaccessible from wild type. Redesigned starting points were evolved to cleave ataxin-2 with ~79-fold greater specificity than the best wild-type comparator.
Published: July 22, 2026
Paper | Code (GitHub)

EVOLVEpro: A few-shot active learning framework for in silico directed protein evolution, combining a protein language model with a top-layer regression model that learns a protein's activity landscape from as few as 10 experimental data points per round. Benchmarked in silico across 12 protein datasets and applied to five real proteins (a monoclonal COVID antibody, a miniature CRISPR nuclease, a Bxb1 integrase, a prime editor, and T7 RNA polymerase), EVOLVEpro achieved up to 100-fold improvement of desired properties within as few as four rounds of evolution, without requiring structural information or prior data.
Posted: July 18, 2024
Preprint | Code (GitHub)

SynTnpBs: A structure- and evolution-guided protein engineering strategy from Jennifer Doudna's lab for redesigning TnpB, a compact CRISPR-Cas12-like RNA-guided nuclease, into highly divergent, catalytically active variants. Unconstrained ESM-IF1 inverse folding reproduced the TnpB fold but altered key RNA/DNA-contacting residues, so the authors masked conserved and coevolving positions before regenerating sequences. The REC and NUC lobes were engineered semi-independently and screened across 1,980 combinatorial pairs; selected variants matched or exceeded wild-type genome-editing activity in human cells and at endogenous loci, and also edited Arabidopsis. Cryo-EM of the most divergent variant captured an unresolved TAM-bound intermediate while preserving the catalytic cycle.
Published: July 16, 2026
Paper | Preprint | Code (GitHub)