IgT5 Unpaired

Encode¶

Generate embeddings for input sequences

from biolmai import BioLM
response = BioLM(
    entity="igt5-unpaired",
    action="encode",
    params={
      "include": [
        "mean"
      ]
    },
    items=[
      {
        "heavy": null,
        "light": null,
        "sequence": "EVQLVESGGGLVQ"
      }
    ]
)
print(response)

curl -X POST https://biolm.ai/api/v3/igt5-unpaired/encode/ \
  -H "Authorization: Token YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
  "params": {
    "include": [
      "mean"
    ]
  },
  "items": [
    {
      "heavy": null,
      "light": null,
      "sequence": "EVQLVESGGGLVQ"
    }
  ]
}'

import requests

url = "https://biolm.ai/api/v3/igt5-unpaired/encode/"
headers = {
    "Authorization": "Token YOUR_API_KEY",
    "Content-Type": "application/json"
}
payload = {
      "params": {
        "include": [
          "mean"
        ]
      },
      "items": [
        {
          "heavy": null,
          "light": null,
          "sequence": "EVQLVESGGGLVQ"
        }
      ]
    }

response = requests.post(url, headers=headers, json=payload)
print(response.json())

library(httr)

url <- "https://biolm.ai/api/v3/igt5-unpaired/encode/"
headers <- c("Authorization" = "Token YOUR_API_KEY", "Content-Type" = "application/json")
body <- list(
  params = list(
    include = list(
      "mean"
    )
  ),
  items = list(
    list(
      heavy = None,
      light = None,
      sequence = "EVQLVESGGGLVQ"
    )
  )
)

res <- POST(url, add_headers(.headers = headers), body = body, encode = "json")
print(content(res))

POST /api/v3/igt5-unpaired/encode/¶

Encode endpoint for IgT5 Unpaired.

Request Headers:

• Content-Type – application/json

• Authorization – Token YOUR_API_KEY

Request

• params (object, optional) — Configuration parameters:

  • include (array of strings, default: [“mean”]) — Embedding types to include in the response; allowed values: “mean”, “residue”

• items (array of objects, min: 1, max: 8) — Input sequences:

  • heavy (string, optional, min length: 1, max length: 256) — Heavy chain amino acid sequence using extended amino acid codes

  • light (string, optional, min length: 1, max length: 256) — Light chain amino acid sequence using extended amino acid codes

  • sequence (string, optional, min length: 1, max length: 512) — Unpaired amino acid sequence using extended amino acid codes

Example request:

http Copy

POST /api/v3/igt5-unpaired/encode/ HTTP/1.1
Host: biolm.ai
Authorization: Token YOUR_API_KEY
Content-Type: application/json

      {
  "params": {
    "include": [
      "mean"
    ]
  },
  "items": [
    {
      "heavy": null,
      "light": null,
      "sequence": "EVQLVESGGGLVQ"
    }
  ]
}

Status Codes:

• 200 OK – Successful response

• 400 Bad Request – Invalid input

• 500 Internal Server Error – Internal server error

Response

• results (array of objects) — One result per input item, in the order requested:

  • embeddings (array of floats, size: 1024) — Mean embedding vector for the encoded input; numerical values unbounded.

  • residue_embeddings (array of arrays of floats, shape: [sequence_length, 1024]) — Per-residue embedding vectors for the encoded input; numerical values unbounded.

Example response:

http Copy

HTTP/1.1 200 OK
Content-Type: application/json

      {
  "results": [
    {
      "embeddings": [
        -0.014581063762307167,
        -0.05809000879526138,
        "... (truncated for documentation)"
      ]
    }
  ]
}

Performance¶

• IgT5 Unpaired runs on NVIDIA T4 GPUs with 4 vCPUs and 16 GB RAM allocated per worker, enabling GPU-accelerated embedding generation for large antibody datasets.

• Compared to general protein language models fine-tuned on antibodies (e.g., ProtT5, ProtBert), IgT5 Unpaired achieves substantially higher antibody sequence recovery, especially in hypervariable CDRs (e.g., CDR-H3 recovery ~0.60 vs. ProtT5 ~0.34), while maintaining high framework-region accuracy.

• Within BioLM’s antibody-specific encoders, IgT5 Unpaired matches or slightly exceeds IgBert Unpaired on overall heavy-chain recovery (≈0.912 vs. 0.910) and light-chain recovery (≈0.951 vs. 0.949–0.956), offering improved CDR performance with a larger T5 architecture.

• For downstream regression on antibody-focused tasks, embeddings from IgT5 Unpaired improve binding-affinity prediction over ProtT5 on representative benchmarks (R² ≈ 0.299 vs. 0.290), but general protein models like ProtT5 remain preferable for broad protein properties such as expression (R² ≈ 0.697 vs. 0.567).

Applications¶

• Antibody affinity maturation and optimization by prioritizing beneficial mutations in complementarity-determining regions (CDRs) using IgT5 embeddings as features in downstream models, enabling rapid in silico ranking of variants for improved antigen binding; valuable for accelerating therapeutic antibody lead optimization, but not sufficient alone to predict broader expression, aggregation, or stability properties.

• Identification and scoring of putative paratope residues from IgT5 residue-level embeddings, allowing efficient prioritization of positions most likely to impact antigen binding; useful for rational engineering tasks such as specificity tuning or epitope focusing, although not a substitute for full structural modeling or epitope mapping.

• High-throughput antibody sequence filtering and ranking by training regression or classification models on IgT5 mean embeddings to approximate antigen binding affinity, enabling rapid down-selection of large antibody libraries in discovery campaigns; effective for sequence-based enrichment but limited in capturing developability attributes such as solubility, viscosity, or immunogenicity.

• Clustering and diversity analysis of antibody repertoires using IgT5 mean or residue-level embeddings, facilitating selection of diverse panels of candidates while maintaining coverage of sequence space; supports library design and repertoire mining, but embeddings alone do not provide detailed functional annotation and still require experimental follow-up.

• Analysis of heavy-light chain pairing signals by comparing IgT5 embeddings of paired versus unpaired sequences, supporting development of custom pairing-compatibility predictors for repertoire and single-cell sequencing data; useful for reconstructing likely native pairs in human-like repertoires, though performance may degrade on highly non-human or unusual antibody formats.

Limitations¶

• Maximum Sequence Length: For paired inputs (heavy and light chains), each chain must be ≤ 256 amino acids. For unpaired inputs (sequence), the maximum length is 512 amino acids.

• Batch Size: API requests are limited to a maximum of 8 sequences per items list. Larger workloads must be split across multiple requests.

• IgT5 embeddings are designed for antibody variable regions; using arbitrary or non-immunoglobulin proteins for general property prediction (for example, expression across diverse protein families) may underperform compared to general protein language models such as ProtT5.

• Training data is predominantly human antibody sequences from the Observed Antibody Space (OAS); performance may degrade for sequences unlike typical human antibodies (for example, non-human species, synthetic frameworks, or highly engineered CDRs).

• The API exposes IgT5 as an encoder only: it returns sequence embeddings via embeddings (mean over residues) and optionally residue_embeddings. It is not intended for direct generative use (for example, de novo sequence design) or for antibody–antigen 3D structure prediction.

• Cross-chain context learned from paired heavy–light data is only available when you provide both heavy and light; using unpaired sequence inputs will not capture pairing-specific features and may reduce accuracy for tasks that depend on native chain pairing.

How We Use It¶

IgT5 Unpaired is used as a backbone encoder in antibody discovery and optimization programs, where its embeddings of unpaired variable regions drive rapid screening, property prediction, and design-space exploration before committing to wet-lab cycles. The model supports scalable, standardized encoding of large repertoires, and its representations feed into downstream models for binding, expression, and developability that we integrate with structure-based tools and lab data.

• IgT5 Unpaired embeddings enable prioritization and clustering of candidates in early antibody campaigns, reducing the number of variants that need to be synthesized and tested.

• Combined with paired IgT5 and other BioLM encoders, IgT5 Unpaired supports end-to-end workflows from repertoire mining through sequence optimization and multi-round affinity maturation.

Related¶

• IgT5 Paired – Paired variant of IgT5 trained on heavy–light chain complexes, useful when embeddings must capture cross-chain context beyond what unpaired IgT5 provides.

• IgBert Unpaired – Antibody-specific BERT encoder trained on unpaired sequences, offering an alternative embedding space for the same types of inputs as IgT5 Unpaired.

• ABodyBuilder3 Language – Structure-prediction model that consumes sequence-based embeddings, commonly used with IgT5 Unpaired for sequence-to-structure antibody modeling.

• AntiFold – Inverse folding model for sequence design under structural constraints, complementary to IgT5 Unpaired embeddings in sequence–structure antibody design workflows.

References¶

• Kenlay, H., Dreyer, F. A., Kovaltsuk, A., Miketa, D., Pires, D., & Deane, C. M. (2024). Large scale paired antibody language models. PLoS Computational Biology, 20(12), e1012646. https://doi.org/10.1371/journal.pcbi.1012646