Glossary

Every specialist term bindsight uses, in plain language. If a word elsewhere in these docs or the app is unfamiliar, it is almost certainly defined here.

In plain terms

In plain terms: bindsight compares RNA-seq counts between conditions to prioritise candidate cell-surface targets. It can then design protein sequences and assess their predicted complexes with structure models, while recording the inputs, settings and outputs. RNA abundance is not surface protein abundance, and model confidence is not proof of binding or clinical safety; those require appropriate experimental validation.

RNA-seq counts → ranked de novo protein binder candidates, with full provenance back to the patient cohort.

RNA-seq : A lab method that measures which genes are switched on in a tissue sample, and how strongly, by sequencing its RNA. It is the starting input to bindsight.

Differential expression (DEG) : Comparing gene activity between two groups — here tumour versus healthy tissue — to find the genes turned notably up or down in disease.

log2 fold-change (log2FC) : How many times higher or lower a gene's activity is between the two groups, on a log2 scale: a value of 4 means about 16× higher in tumour.

padj / FDR : A confidence figure that a difference is real rather than chance, adjusted for testing thousands of genes at once. Smaller is stronger (e.g. 1e-59 is overwhelming evidence).

Antigen : A molecule a drug or the immune system can recognise and target. In bindsight the antigens of interest are proteins sitting on the cell surface.

Surfaceome : The full set of proteins that sit on the outer surface of cells — the ones a binder could physically reach. bindsight keeps only surface proteins as targets.

De novo protein binder : A small protein designed from scratch (rather than borrowed from nature) to stick tightly and specifically to a chosen target protein.

RFdiffusion + ProteinMPNN : Two AI models used together: RFdiffusion invents a plausible 3-D shape for a binder against the target, and ProteinMPNN chooses an amino-acid sequence that would fold into that shape.

Boltz-2 : An AI model that predicts the 3-D structure of the binder and target locked together, and reports how confident it is in that predicted interface. Confidence is not a binding measurement: bindsight's own paired control folded each design beside a shuffle of its own sequence, and the shuffles scored as well. Treat these numbers as a filter to triage wet-lab work, never as evidence that a design binds.

ipTM : A 0–1 score of how confident the structure model is in the predicted interface between two proteins. 0.65 is the threshold the de novo design literature commonly uses, and bindsight reports against it for comparability — but its own scramble control could not separate real designs from shuffles of the same composition at that cutoff, so a value above 0.65 is not on its own evidence of a binder. See the calibration.

PAE (interaction) : Predicted Aligned Error at the binding interface, in ångströms — how uncertain the model is about the contact. Lower means a more trustworthy predicted grip.

Developability : How practical a designed protein would be to actually make and use: is it stable, soluble, and free of trouble-prone residues? Good scores mean fewer surprises in the lab.

Provenance / RO-Crate : A complete, machine-readable record of every input, tool version and step, packaged (as an RO-Crate) so anyone can retrace and reproduce a result — and cite it.

ERBB2 (HER2) : A well-known breast-cancer cell-surface antigen and an approved drug target — the canonical positive control for a discovery pipeline. bindsight does not rediscover it: in the whole unstratified TCGA-BRCA cohort it measures log2fc 0.92, below the fold-change floor, so it does not clear the significance rule. That is a limit of bulk differential expression, not of the ranking; see benchmarks/study/RESULTS.md.