Humatch outputs and interpretation
Humanness scores, flagged positions, and closest germline from Humatch.
Humatch returns, per sequence, an overall humanness score, a pass/fail against your threshold, per-V-gene and per-position scores, the closest human germline V gene, and a list of the positions that pull the score down. Read the flagged positions alongside the overall score: the score tells you whether to humanize, the flagged positions tell you where.
Per-sequence result
Each scored sequence carries these fields:
| Field | Type | What it tells you |
|---|---|---|
sequence_name | string | The name you gave the sequence. |
humanness_score | float | Overall humanness (0–1). |
is_human | bool | Whether the score clears your threshold. |
chain_type | string | The chain type scored (H or L). |
v_gene_scores | map gene→float | Per-V-gene humanness scores. |
top_human_germline | string | null | Closest human germline V gene. |
positional_scores | map position→float | Per-position humanness, on the numbering scheme used. |
flagged_positions | string[] | Positions with low humanness — the humanization targets. |
Paired result
A request that used chain_pairs also returns paired_results, one entry per pair. For a single-chain request the list is empty.
| Field | Type | What it tells you |
|---|---|---|
pair_name | string | The name you gave the pair. |
paired_score | float | null | The paired classifier's score for this heavy/light combination (0–1). null when the pair failed, with error saying why. |
paired_score_class | string | null | Which class paired_score is the probability of. |
heavy | object | null | The full per-chain result for the heavy chain, in the shape above. |
light | object | null | The same for the light chain. |
error | string | null | Why this pair has no paired score. |
Reading the paired score
paired_score is the probability Humatch's paired CNN assigns to its true class — the classifier's two classes are fake and true. The returned paired_score_class names which one the number refers to, so read it rather than assuming.
⚠ Humatch publishes no interpretation bands for the paired score, so none are applied. There is no pass/fail on it, and no threshold: score_threshold governs the per-chain humanness classification only. The number is reported as measured, with no verdict attached, and inventing a cutoff for it means inventing a standard the model's authors did not set.
Job-level summary
| Field | Type | Meaning |
|---|---|---|
failed_sequences | string[] | Sequences Humatch could not score. |
success_count | int | Sequences scored successfully. |
fail_count | int | Sequences that failed. |
paired_success_count | int | null | Pairs that produced a paired score. null — not 0 — when no pair was requested. |
paired_fail_count | int | null | Pairs that did not. null when no pair was requested. |
partial | boolean | true when at least one sequence or pair failed. Check it before treating the run as complete. |
scheme_used | string | Numbering scheme applied. |
requested_scheme | string | null | The scheme you asked for. Recorded, not applied. |
threshold_used | float | Classification threshold used. |
paired_score_method | string | null | How paired_score was computed. null when no pair was requested. |
Humatch always aligns to IMGT
numbering_scheme reports the numbering that was actually used, and it is always IMGT. A chothia or kabat request is recorded in requested_scheme and not applied, so position numbers in the result are IMGT whatever you asked for. Compare the two fields before reading a position number across tools.
Reading the result card
The Humatch result renders on the canvas as an antibody-structure card, surfacing the humanness score and per-chain read so you can judge the candidate at a glance.
What to do next
- Take the
flagged_positionsto BioPhi and humanize them — Humatch scores, BioPhi rewrites. - Cross-check the flagged positions against the paratope so you humanize away from binding residues.
- Run antibody developability before committing a candidate.
Read next
Was this page helpful?
Humatch inputs and parameters
Up to 50 sequences, the chain_type that picks the heavy or light model — mislabel it and the score is meaningless; the threshold moves only the human call.
Humatch quickstart
Paste a sequence, set the chain to H or L, and read back a humanness score, the closest human germline V gene, and the low-humanness positions to target.