AbLang2 paired
Represent one antibody heavy/light-chain pair as a numerical vector for comparison and downstream analysis.
API/CLI tool ID: ablang2-paired-embedding
Understanding inputs and outputs
An antibody heavy chain and light chain are distinct protein chains. AbLang2 paired reads their amino acid sequences together and represents the pair with 480 numbers, called an embedding. You can compare embeddings produced by the same model or use them as inputs to a separately trained classifier.
Input settings
The only setting is inputFile: select a UTF-8 JSON file containing the object below. It is supplied as the single execution input file input.json. Each run accepts one pair.
The example below uses synthetic sequences to illustrate the format. It is not a real antibody or a validated chain pair.
{"heavy":"ACDEFGHIKLMNPQRSTVWY","light":"YWVTSRQPNMLKIHGFEDCA"}Both heavy and light are required, each 1–160 characters long. Only uppercase ACDEFGHIKLMNPQRSTVWY is accepted. The entire JSON file is limited to 4096 bytes. These bounds are the service input policy, not a claim about maximum model capacity or antibody identity.
Additional or duplicate JSON keys are rejected.
FASTA headers, whitespace or line breaks inside a sequence, lowercase letters, X, *, -, empty chains and single-chain inputs are rejected. No automatic uppercasing, whitespace removal, masking or alignment is performed. Do not include a model selector or a batch array in this input.
Original bytes and reproducibility
inputSha256 is the SHA-256 of the uploaded original bytes, not of reserialized JSON. Changing valid JSON whitespace or line breaks around the object changes this digest even when the sequences are identical. Retain the original file and the result's model, code and weight metadata for later comparisons.
Output file
The result is embedding.json.
embedding is an array of 480 finite numbers and dimension is 480. modelId is ablang2-paired, mode is seqcoding, and pooling is upstream-mean-including-special-tokens. The file also records schemaVersion, modelRecord, codeCommit, inputSha256, heavyLength, lightLength and weightsSha256. It does not duplicate the input sequences.
The official implementation formats the pair as <heavy>|<light> and includes boundary tokens <, > and the chain separator | in its mean. Do not treat this as equivalent to a mean with those tokens removed, concatenated per-chain vectors, or vectors from another model.
Interpreting results and next steps
Individual values have no higher-is-better or lower-is-better interpretation. This tool does not return a similarity score or classification; choose a comparison metric or downstream model separately and validate it on relevant study data. Verify the chain pairing separately as well.
An embedding is not binding affinity (how strongly two molecules bind), efficacy, a success probability or three-dimensional structure coordinates. Matching numerical replay does not establish actual binding, biological function or clinical effect.
Execution availability
This guide does not mean execution is available. Check the available version, resources, and cost in your project execution plan, review the inputs, settings, and terms of use, and then approve.
Licenses and delivery scope
Nadoo Bio runs inference on managed servers and provides results. This configuration does not deliver container images, model weights or executable binaries to customers. The root code license is BSD-3-Clause; the paired weight record identifies BSD-3-Clause-Clear. Their respective notices, conditions and disclaimers are retained.
BSD-3-Clause-Clear does not grant patent rights. That alone is not a commercial-use prohibition, nor is it an assurance of patent clearance. Code, weights and dependencies have distinct terms; no endorsement or warranty by the upstream authors is claimed.