Protein2PAM

Every CRISPR-Cas enzyme can only cut next to a short protospacer-adjacent motif, or PAM. That requirement is the largest constraint on where an editor can be aimed: SpCas9's NGG PAM leaves long stretches untargetable, and for a therapeutic directed at one pathogenic variant, the absence of a usable PAM can end the program. Broadening PAM recognition has historically meant directed evolution or structure-guided mutagenesis on a single enzyme — slow work repeated for the next scaffold.

Protein2PAM, from Profluent with collaborators at the University of Washington, Massachusetts General Hospital, and Harvard, inverts the problem. It learns the mapping from Cas protein sequence to PAM specificity from evolutionary data alone: PAMs inferred bioinformatically from CRISPR arrays across a vast microbial genome survey, paired with the effectors that read them. A protein language model then predicts, for any Cas sequence, the nucleotide probability matrix describing which PAMs it recognizes — without structures, alignments, or experimental measurements.

Because that prediction is differentiable and fast, the model can also be run in reverse as a design objective, computationally evolving Nme1Cas9 — a compact Cas9 favored for delivery but restricted by a long, demanding PAM — into variants with broadened and redirected recognition, confirmed experimentally. Posted to bioRxiv in January 2025 and published in Nature Biotechnology, it joins Profluent's CRISPR work alongside OpenCRISPR-1.

#Key Features

• Sequence-to-PAM prediction: A Cas protein sequence goes in and a position-weight matrix over PAM nucleotides comes out, with no structure or alignment required.
• Coverage across CRISPR types: Separate checkpoints handle Type I (Cas8/10d subunits of the Cascade complex), Type II (Cas9), and Type V (Cas12), and a companion confidence model flags unreliable predictions for filtering.
• Structure-free interpretability: In-silico deep mutational scanning over the fixed model recovers the residues contacting PAM bases, matching known PAM-interacting positions in solved Cas9 structures without ever seeing one.
• Generative enzyme engineering: Markov chain Monte Carlo sampling steers a scaffold toward a chosen PAM while a protein language model keeps the sequence plausible.

#Technical Details

Each Protein2PAM model is a pretrained 650-million-parameter ESM-2 transformer encoder, fine-tuned end-to-end, with a two-layer MLP head over the final-layer [CLS] token emitting PAM nucleotide probabilities. Training data comes from mining CRISPR-Cas operons for spacers with identifiable protospacer matches, yielding 45,816 distinct protein-PAM pairs covering 71.6% of surveyed operons — a 2.8-fold increase over the largest prior bioinformatic Cas9 PAM dataset. Input choice was tuned per system: the PAM-interacting domain alone for Cas9, Cas8/10d for Type I, the full sequence for Type V.

Accuracy, measured as cosine similarity to held-out PAMs, reached 0.949 for Type I, 0.868 for Type II, and 0.955 for Type V. Against experimentally characterized enzymes it recovered consensus PAMs for every active system among 14 Type I CRISPR and CAST systems, including proteins only 25% identical to training data, and reached median accuracy 0.797 across 112 Type II systems, rising to 0.883 after confidence filtering removed 52% of predictions. Type V results were mixed: 0.772 for Cas12b and Cas12f (n = 14) but 0.460 for Cas12a and other subtypes (n = 31), where the model over-predicts TTTN.

For design, Gibbs-with-Gradients MCMC optimized Nme1Cas9 toward six target PAMs under a ProGen2 likelihood constraint, generating 30,000 variants from which 22 were characterized by HT-PAMDA. The paper's tables report larger effects than the abstract's "up to 50-fold" summary: the single-point N1029A variant cleaved N4GNAT PAMs 270-fold faster than wild type, and the 13-mutation NmeN4G.1 variant cleaved 42 N4G PAMs versus 7, its top ten PAMs cut 56.4-fold faster than wild type and 21.5-fold faster than eNme2-C, a broad-PAM variant from multiple rounds of phage-assisted evolution.

#Applications

The immediate use is scoping: given a genomic target, screen natural and engineered Cas orthologs for one whose PAM lands where needed, with a confidence estimate attached. The hosted webserver makes this accessible without installation, while the PyTorch package and Hugging Face checkpoints support batch screening of ortholog collections and metagenomic discovery; the generative pipeline covers the harder case where no natural enzyme fits. Code and weights are released for non-commercial research only — code under the Polyform Noncommercial License 1.0.0, models and data under CC BY-NC 4.0.

#Impact

Protein2PAM demonstrates that a property mediated by specific protein-DNA contacts can be learned from evolutionary sequence pairs alone and inverted into a design objective, with wet-lab confirmation that the designs behave as predicted. That it recovers PAM-contacting residues without structural input argues that evolutionary-scale representations encode molecular recognition detail. Its limits sit where the data thins: Type V performance drops sharply for underrepresented Cas12 subtypes, several designed PAM targets yielded few active enzymes (1 of 10 for N7A, N6TT, and N6TA), and training uses bioinformatic inferences rather than direct measurements. Noncommercial licensing also means industrial groups cannot deploy the released checkpoints without a separate arrangement.