How to read a Provotics report.

Input format

Provotics takes a single tumor RNA-seq expression profile, gene-level counts or TPM. Genes should be identified by HGNC symbol or Ensembl gene ID; the pipeline maps them to a common gene space. A profile covers roughly 18,000 protein-coding genes; the pipeline then maps it to the fixed 3,882-gene panel the model reads. No identifiable patient data should ever be included, expression values only.

The pipeline

• QC. The profile is screened for quality and obvious technical problems before any prediction is made.
• Harmonize. Genes are mapped to the model's reference space and the profile is normalized so it is comparable to the training distribution.
• Model. An ensemble scores the profile across the prediction heads.
• Calibrate. Raw scores are turned into calibrated probabilities, and out-of-distribution profiles are flagged.

Outputs

• Site of origin across 25 anatomical sites, with a calibrated probability and the top alternatives.
• Molecular subtype where applicable (for example PAM50 on breast).
• Candidate druggable targets, surfaced as research hypotheses.
• Immune profile, from immune-cold to inflamed.
• Driver genes behind each call, so the prediction is reviewable.

Reading the confidence

Every call carries a calibrated probability: a stated 80% means roughly 80% in practice. When a profile is low-confidence or looks unlike anything in training, Provotics abstains and flags it as novel rather than forcing a label. Treat an abstention as a signal, not a failure, it means the honest answer is "not enough evidence."

Access

The model is invite-only during the research phase and granted under a confidentiality agreement. See Pricing, Safety, and the access terms, then apply for access.