Research & Analysis Workflows

Deterministic developability/liability gate for GENERATED antibody Fv sequences (e.g. an AlphaFold-Multimer / RFdiffusion+ProteinMPNN designed binder, or an antibody language model) with karyon. Use AFTER designing an antibody/binder and BEFORE expressing or advancing it — answers "is this Fv developable, or does it carry a liability that makes it undruggable?" with a pass/fail verdict and legible per-reason explanations (an unpaired cysteine, an N-glycosylation sequon in a CDR, an extreme CDR-H3 length or isoelectric point — and, disclosed, the deamidation/isomerization/oxidation hotspots, fragmentation sites, and hydrophobicity/charge proxies even approved antibodies carry). Pure stdlib — no GPU, no network, no numpy. Input is the VH (+VL) amino-acid sequence; single-domain VHH/nanobodies are supported.

Audit an ML benchmark for data leakage with karyon — measure how much a reported accuracy is inflated by train/test contamination. Use for leakage detection, data-leakage audit, scaffold leakage, template memorization, random-vs-scaffold split inflation, protein-protein interaction (PPI) node-identity leakage, retrosynthesis / molecular-property / PPI benchmark honesty, or deciding whether a model's headline number is real.

Deterministic physical-validity gate for co-folding model output (protein + ligand predicted in one frame, e.g. Boltz-2, AlphaFold-3, Chai-1) with karyon. Use AFTER predicting a protein-ligand complex and BEFORE trusting the pose — it answers "is this pose physically valid?" with a pass/fail verdict and legible per-reason explanations (ligand-protein clash, volume overlap, out-of-pocket placement, plus the ligand's own bond/angle/ring/strain geometry). No GPU required.

Deterministic interface-validity gate for predicted or designed protein COMPLEXES (two or more protein chains in one frame, e.g. AlphaFold-Multimer models, or RFdiffusion + ProteinMPNN designed binders against a target) with karyon. Use AFTER predicting/designing a complex and BEFORE trusting the interface — it answers "is this interface physically valid?" with a pass/fail verdict and legible per-reason explanations (inter-chain clash, gross interpenetration, chains out of contact). No GPU required.

Deterministic synthesizability/manufacturability gate for GENERATED DNA sequences (e.g. NVIDIA BioNeMo's Evo2, or any genomic sequence generator) with karyon. Use AFTER generating a DNA sequence and BEFORE ordering or cloning it — answers "can this actually be synthesized and cloned, and will it behave?" with a pass/fail verdict and legible per-reason explanations (GC out of the synthesis band, homopolymer / poly-G runs, out-of-range length, strong self-hairpin, restriction-site collisions, and — across a batch — cross-hybridization between sequences). Pure stdlib — no GPU, no network, no numpy.

Deterministic validity / synthesizability / drug-likeness gate over GENERATED molecules with karyon — reject invalid, unsynthesizable, or out-of-range candidates with named reasons before trusting or ordering them. Use for molecule validity, SMILES sanitization, Ertl synthetic-accessibility screening, extreme MW/cLogP rejection, PAINS/Brenk structural-alert and Lipinski/Veber drug-likeness disclosure, or qualifying the output of a generative-chemistry model (e.g. NVIDIA BioNeMo GenMol / MolMIM) before downstream use.

Run a deterministic physical-validity DRC over molecular docking poses with karyon. Use for pose validity, PoseBusters-style geometric checks, bond-length/angle/ring-planarity/steric-clash/strain screening, qualifying or filtering ranked docking poses (e.g. DiffDock output), or deciding whether a predicted binding pose is physically plausible before trusting it.

Functional + synthesizability design-rule check over σ70 bacterial promoter sequences with karyon — qualify whether a designed or generated promoter carries a real −35/−10 architecture before ordering, each rejection named. Use for promoter design rules, −35/−10 box check, inter-box spacer, promoter GC band, regulatory-element QC, or qualifying the output of a DNA/regulatory sequence generator (e.g. Evo2) before synthesis.

Qualify a pooled CRISPR screen's non-hits with karyon — flag genes called "no effect" that the screen was simply under-powered to detect. Use for CRISPR screen QC, under-power detection, silent-failure / false-negative qualification, guide-count reliability, NT/NEG-calibrated screen analysis, or deciding whether a non-hit is trustworthy or just low-powered.

Qualify a single-cell Perturb-seq screen's no-phenotype calls with karyon — flag perturbations called "no effect" where the guide never actually knocked the target down (a silent failure), so the null is an artifact, not biology. Use for Perturb-seq QC, single-cell CRISPR screen silent-failure / false-negative qualification, on-target knockdown reliability, low-cell-count power checks, or deciding whether a "no-phenotype" perturbation is a true negative or just a failed knockdown.