CICERO

CICERO (CRISPR Cas9 PAM predictor model) predicts the protospacer-adjacent motif (PAM) preference of a Cas9 protein directly from its amino acid sequence. Every Cas9 ortholog will only cut where a short, ortholog-specific PAM sits next to the target site, so PAM requirements are the single biggest constraint on where in a genome CRISPR editing is possible. Discovering orthologs with alternative or relaxed PAMs is therefore a central route to broader editing coverage, but determining a new nuclease's PAM has traditionally required either wet-lab depletion assays or alignment-based metagenomic inference that only works when enough matching protospacers exist in phage and plasmid databases.

CICERO was developed at the University of Zurich by the Krauthammer and Schwank groups, with collaborators at ETH Zurich, and posted as a bioRxiv preprint in August 2025, then published in Nature Communications. It accompanies CRISPR-PAMdb, a curated database built by mining more than 3.8 million bacterial and archaeal genomes for Cas9 proteins and more than 7.4 million phage and plasmid sequences for protospacers. Spacer–protospacer alignment yielded consensus PAMs for 8,003 of the 62,542 Cas9 clusters identified — leaving the large majority of the family unannotated, which is precisely the gap CICERO fills.

It answers a different question from crisprSFM, which scores guide RNA off-target specificity against genomic DNA: CICERO characterizes the nuclease itself, asking what target-site grammar an uncharacterized Cas9 recognizes.

#Key Features

• PAM profiles from sequence alone: Takes a Cas9 amino acid sequence and outputs a 10x4 matrix of nucleotide probabilities across ten PAM positions, requiring no protospacer alignments or experimental assays.
• Calibrated confidence head: A second training phase adds a confidence network that regresses the model's own expected accuracy, correlating with true accuracy at r = 0.8 on external data and enabling reliable thresholding.
• Zero-shot expansion of PAM annotation: Applied to Cas9 clusters where alignment-based inference was impossible, generating PAM profiles for 50,308 additional proteins, of which 17,453 are high-confidence.
• Companion database: CRISPR-PAMdb pairs the model with alignment-inferred PAMs for 8,003 Cas9 clusters, providing both the training signal and an independent annotation source.
• Modular to other effectors: The framework is not specific to Cas9 and can be retrained for other CRISPR effector families.

#Technical Details

CICERO builds on the ESM-2 protein language model. The released model, CICERO-650M, encodes a Cas9 sequence with the 650M-parameter ESM-2 checkpoint, takes the [CLS] token as a whole-protein representation, and passes it to a two-layer MLP head (hidden dimension 1,280, ReLU, dropout 0.2) that emits the PAM logits. The language model backbone is fine-tuned during this first training phase. Training combines cross-entropy on the PAM logits with an augmented cosine-similarity loss that weights high-information positions and models low-information positions with a fictitious "N" base; that same augmented cosine similarity is the accuracy metric reported throughout. In phase two, both the backbone and the PAM head are frozen and only the confidence head is trained, with an L2 loss against per-sequence accuracy.

Training used the 8,003 alignment-annotated Cas9 clusters, reduced to 7,571 sequence–PAM pairs after removing proteins shorter than 200 residues or longer than 1,538, under five-fold stratified cross-validation for 15 epochs at a learning rate of 1e-4. CICERO-650M reached a median accuracy of 0.69 +/- 0.03 on held-out test splits and 0.75 on an external benchmark of 79 experimentally characterized Cas9 orthologs. Restricting to predictions with confidence above 0.9 raised accuracy to 0.86 on database test data and a median of 0.95 on the external benchmark. Backbone sizes from 8M to 3B parameters were compared; performance rose through 650M and fell at 3B, reflecting the limited training set. Accuracy is lower for short Cas9 sequences, which are sparsely represented in the database.

#Applications

CICERO is aimed at genome-editing tool discovery. Given the sequence of a newly mined or engineered Cas9, it returns a PAM profile and a confidence score, letting researchers prioritize which orthologs to test when they need a nuclease that can reach a site unreachable by SpCas9's NGG PAM — a common obstacle in base and prime editing of specific disease alleles. The 17,453 high-confidence predictions form a searchable pool of candidate nucleases indexed by predicted target-site grammar.

#Impact

By combining a large metagenomic mining effort with a protein language model, CICERO expands computationally annotated Cas9 PAM coverage roughly sevenfold beyond what spacer–protospacer alignment alone supports. The work was subsequently published in Nature Communications. Code and pipeline are MIT-licensed, checkpoints for all five cross-validation folds download without registration, and CRISPR-PAMdb ships with the paper. Its limitations are stated plainly by the authors: accuracy is moderate in absolute terms, training data covers only Cas9 rather than Cas12 or Cas13, experimentally validated PAMs remain scarce for phylogenetically distant clades, and the predictions are computational hypotheses that still require in vitro or in vivo confirmation.