TEMPRO 650M

Predict¶

Predict melting temperature for input nanobody sequences

from biolmai import BioLM
response = BioLM(
    entity="tempro-650m",
    action="predict",
    params={},
    items=[
      {
        "sequence": "QVQLVESGGGLVQPGGSLRLSCAASGFTFSSFNWIHWVRQAPGKGLEYKMIAASSSVGGTPYYADSVEGRFTISRDNAKNTVYLQMNSLRPEDTAVYYCV"
      },
      {
        "sequence": "QVQLQESGGGSVQAGGSLRLSCTASGFNIHKNYLAHWFRQAPGKEREGVAALSPAGGTPYYADSVKGRFTISRDNAKNTVYLQMNSLRPEDTAVYYCARR"
      }
    ]
)
print(response)

curl -X POST https://biolm.ai/api/v3/tempro-650m/predict/ \
  -H "Authorization: Token YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
  "items": [
    {
      "sequence": "QVQLVESGGGLVQPGGSLRLSCAASGFTFSSFNWIHWVRQAPGKGLEYKMIAASSSVGGTPYYADSVEGRFTISRDNAKNTVYLQMNSLRPEDTAVYYCV"
    },
    {
      "sequence": "QVQLQESGGGSVQAGGSLRLSCTASGFNIHKNYLAHWFRQAPGKEREGVAALSPAGGTPYYADSVKGRFTISRDNAKNTVYLQMNSLRPEDTAVYYCARR"
    }
  ]
}'

import requests

url = "https://biolm.ai/api/v3/tempro-650m/predict/"
headers = {
    "Authorization": "Token YOUR_API_KEY",
    "Content-Type": "application/json"
}
payload = {
      "items": [
        {
          "sequence": "QVQLVESGGGLVQPGGSLRLSCAASGFTFSSFNWIHWVRQAPGKGLEYKMIAASSSVGGTPYYADSVEGRFTISRDNAKNTVYLQMNSLRPEDTAVYYCV"
        },
        {
          "sequence": "QVQLQESGGGSVQAGGSLRLSCTASGFNIHKNYLAHWFRQAPGKEREGVAALSPAGGTPYYADSVKGRFTISRDNAKNTVYLQMNSLRPEDTAVYYCARR"
        }
      ]
    }

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

library(httr)

url <- "https://biolm.ai/api/v3/tempro-650m/predict/"
headers <- c("Authorization" = "Token YOUR_API_KEY", "Content-Type" = "application/json")
body <- list(
  items = list(
    list(
      sequence = "QVQLVESGGGLVQPGGSLRLSCAASGFTFSSFNWIHWVRQAPGKGLEYKMIAASSSVGGTPYYADSVEGRFTISRDNAKNTVYLQMNSLRPEDTAVYYCV"
    ),
    list(
      sequence = "QVQLQESGGGSVQAGGSLRLSCTASGFNIHKNYLAHWFRQAPGKEREGVAALSPAGGTPYYADSVKGRFTISRDNAKNTVYLQMNSLRPEDTAVYYCARR"
    )
  )
)

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

POST /api/v3/tempro-650m/predict/¶

Predict endpoint for TEMPRO 650M.

Request Headers:

• Content-Type – application/json

• Authorization – Token YOUR_API_KEY

Request

• params (object, optional) — Additional model parameters:

  • batch_size (integer, default: 8) — Maximum number of items allowed in a single request

  • min_sequence_len (integer, default: 100) — Minimum length of each protein sequence

  • max_sequence_len (integer, default: 160) — Maximum length of each protein sequence

• items (array of objects, min length: 1, max length: 8) — List of protein sequences to predict melting temperatures for:

  • sequence (string, min length: 100, max length: 160, required) — Protein sequence (100-160 amino acids, typical nanobody length with some generalization)

Example request:

http Copy

POST /api/v3/tempro-650m/predict/ HTTP/1.1
Host: biolm.ai
Authorization: Token YOUR_API_KEY
Content-Type: application/json

      {
  "items": [
    {
      "sequence": "QVQLVESGGGLVQPGGSLRLSCAASGFTFSSFNWIHWVRQAPGKGLEYKMIAASSSVGGTPYYADSVEGRFTISRDNAKNTVYLQMNSLRPEDTAVYYCV"
    },
    {
      "sequence": "QVQLQESGGGSVQAGGSLRLSCTASGFNIHKNYLAHWFRQAPGKEREGVAALSPAGGTPYYADSVKGRFTISRDNAKNTVYLQMNSLRPEDTAVYYCARR"
    }
  ]
}

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:

  • tm (float, typical range: 40.0–95.0 °C) — Predicted melting temperature in Celsius

Example response:

http Copy

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

      {
  "results": [
    {
      "tm": 91.82323455810547
    },
    {
      "tm": 73.70567321777344
    }
  ]
}

Performance¶

• Input type: TemproPredictRequest.items (list of 1–8 protein sequences); each sequence is 100–160 amino acids and must contain only the 20 canonical residues

• Output type: TemproPredictResponse.results (list aligned to inputs); each item has tm: float (predicted melting temperature in Celsius)

• Throughput and latency (end-to-end, including ESM-2 650M embedding and regression head) - NVIDIA H100 80GB: 8-sequence batch completes in ~0.30–0.60 s (p95 ~0.75 s); 12–22 sequences/s sustained - NVIDIA A100 40GB: 8-sequence batch completes in ~0.60–1.10 s (p95 ~1.40 s); 8–14 sequences/s sustained - NVIDIA A10G 24GB: 8-sequence batch completes in ~1.20–1.80 s (p95 ~2.0 s); 4–7 sequences/s sustained - Cold start adds ~2–4 s for model weight load on first call to a fresh worker; subsequent calls are within the ranges above

• Memory/compute footprint - Model weights: ~2.4 GB (ESM-2 650M encoder); total VRAM during inference for 8×160 aa: ~4–6 GB - Mixed-precision FP16 kernels and fused attention reduce activation memory by ~30–40% vs FP32 and improve latency by ~20–30% - Padding to 160 aa with attention masking; compute cost scales primarily with batch size × sequence length (quadratic attention within this short-window regime remains negligible on the listed GPUs)

• Accuracy and domain performance (nanobody thermostability) - On held-out nanobody benchmarks derived from NbThermo-like distributions, models using 650M embeddings underperform larger embedding backbones but remain competitive for screening; expect MAE in the mid–single-digit °C range for sdAbs in-distribution - External validation on nanobodies not present in NbThermo shows lower correlation for 650M vs larger variants: R² ≈ 0.25 (650M) vs ≈ 0.58 (3B) and ≈ 0.67 (15B), reflecting the known scaling law of embedding capacity with stability property prediction - Compared to general-purpose protein thermostability predictors (e.g., ProTDet, DeepStabP), TEMPRO family models maintain meaningful correlation on nanobody-only sets where generic tools show near-zero correlation; this advantage is attributable to nanobody-focused training and embedding features

• Relative performance within the TEMPRO family - TEMPRO 650M vs TEMPRO 3B

  • Speed/throughput: 650M is typically 2–3× faster per batch and fits comfortably on 8–12 GB GPUs; 3B requires ~12–16 GB VRAM and incurs ~2–3× higher latency for the same batch/length

  • Predictive accuracy: 3B reduces error and increases correlation on nanobody Tm prediction; on external validation, 3B approximately doubles explained variance vs 650M (R² ~0.58 vs ~0.25), translating to an expected ~0.5–1.5 °C MAE improvement depending on sequence set

  • Practical guidance: choose 650M for high-throughput triage and large-scale scans; choose 3B when per-sequence accuracy is prioritized (lead refinement, down-selection before synthesis)

  • TEMPRO 650M vs TEMPRO 15B (research reference) - 15B offers the strongest accuracy (e.g., MAE ~4.0 °C and RMSE ~5.7 °C on held-out nanobody test sets reported in the literature) but is substantially slower and more resource-intensive; 650M provides the best cost-latency tradeoff for production screening

• BioLM deployment optimizations relevant to users - On-GPU embedding and regression head execution eliminates host-device transfer overhead between stages - Request coalescing across concurrent users increases GPU occupancy for short sdAb sequences, improving throughput at steady state - Deterministic inference: identical inputs yield identical outputs (no sampling); numerics are stable under FP16 due to short sequence length and bounded activation ranges

• Practical considerations for best observed performance - Batch close-to-homogeneous sequence lengths (all ~120–160 aa) to minimize padding overhead; mixing very short and longer inputs in one batch slightly reduces GPU efficiency - For repeated evaluation of the same sequences (e.g., iterative design loops), enable client-side de-duplication to avoid redundant embedding computation and reduce latency by up to ~50–80% on repeated calls within a session - Network overhead dominates below ~200 ms per call; prefer batched requests to achieve the per-batch times noted above

• Notes on comparisons to structure-first pipelines - Structure-confidence features (e.g., AlphaFold2 or antibody-modeler pLDDT by region) exhibit weak correlation with nanobody Tm; even when available, they are slower to compute and add little to predictive power relative to embedding-driven TEMPRO 650M - For nanobody use-cases, TEMPRO 650M achieves higher throughput and better domain accuracy than structure-derived heuristics, making it preferable for early-stage screening and ranking where speed and predictive signal per GPU-second are critical

Applications¶

• High-throughput triage of VHH sequence libraries before expression: use TEMPRO 650M to rapidly score large candidate sets (from phage/yeast display, DNA synthesis, or generative design) and filter out nanobodies with low predicted Tm to reduce downstream assay load; this accelerates hit discovery by focusing expression and biophysics on stability-favored clones (e.g., drop sequences predicted <60–65 °C when room-temperature stability is required); best for coarse, at-scale filtering where speed matters; not optimal for scFv/IgG or VHH-Fc fusions and may underperform on atypical constructs (very long CDR3s, extensive tags), so confirm with DSF/DSC

• Stability gating during affinity maturation and humanization: rank-order proposed VHH variants by predicted Tm to maintain or improve thermostability while exploring affinity-improving substitutions, enabling design cycles that enforce a minimum stability bar (e.g., keep ≥70 °C for subcutaneous delivery profiles); valuable for program teams balancing potency with developability; sequence-level regression only (no residue attribution), so use as a ranking heuristic and validate experimentally; for very tight ranking, consider higher-capacity TEMPRO endpoints

• Developability risk assessment prior to scale-up and formulation: combine predicted Tm with orthogonal in silico metrics to flag nanobody leads at risk of low thermal robustness before investing in upstream cell line development and downstream fill–finish; helps decide whether to prioritize stabilization strategies or alternative leads (e.g., avoid scaling candidates predicted <55–60 °C when RT logistics are expected); predictions reflect sequence-intrinsic stability and do not account for buffer/excipients, pH, or glycosylation, so formulation effects must be confirmed in wet lab

• Change-control and QC for sequence edits across the engineering pipeline: quickly compare predicted Tm between a reference VHH and edited variants (signal peptides removed, purification tags altered, back-translation differences) to catch stability regressions before ordering DNA or scheduling biophysical panels; practical for vendor oversight and tech-transfer checkpoints; model expects single-domain nanobody sequences (12–15 kDa range); multi-domain fusions and long linkers can reduce reliability

• Indication and use-case alignment for target product profiles: use predicted Tm thresholds to align nanobody leads with intended deployment (e.g., prioritize higher Tm for field-deployable diagnostics or RT-stable reagents; accept moderate Tm for cold-chain biotherapeutics with rapid turnaround); enables early portfolio decisions without exhaustive stress studies; not a substitute for real-time and accelerated stability studies, which remain necessary for regulatory filings and final CMC decisions

Limitations¶

• Input/Output contract: Submit a list items of 1–8 entries (see Batch Size below). Each entry must provide a sequence string of amino acids. The response returns results with one object per input, each containing a single float tm (predicted melting temperature in Celsius). No confidence interval or uncertainty is returned.

• Sequence limits: Minimum Sequence Length 100 aa and Maximum Sequence Length 160 aa. Requests with sequence outside 100–160 amino acids are rejected. Only the 20 canonical amino acids are supported; sequences containing non‑standard residues, unknown characters, tags/linkers, or fusion partners should be cleaned/trimmed before submission.

• Batching limits: Batch Size up to 8 sequences per request via items. Larger jobs must be split client‑side; ordering is preserved 1:1 between items and results.

• Scope and generalization: Trained for single‑domain camelid VHH nanobodies (sdAbs). Predictions on full‑length antibodies, scFvs, multi‑domain fusions, shark VNARs, or non‑antibody proteins are out‑of‑distribution and may be inaccurate—even if they fit the 100–160 aa length window.

• Model performance caveats (650M variant): This endpoint uses ESM_2 650M embeddings; it provides lower accuracy than larger ESM variants reported in the publication. External validation of the 650M model showed modest correlation (R²≈0.25) and larger errors for unusual/extreme Tm nanobodies (e.g., very high or very low experimental Tm). Use results primarily for relative ranking, and confirm final designs experimentally.

• Context not modeled: Predictions are sequence‑only and do not account for buffer composition, pH, concentration, assay method, disulfide engineering, formulation, glycosylation/PTMs, oligomerization, or Fc/fusion effects. Use mid/late‑stage in silico triage on cleaned VHH domains; it is not optimal for early ultra‑high‑throughput screening or for constructs where domain context dominates Tm.

How We Use It¶

TEMPRO 650M enables fast, sequence-only nanobody Tm estimation that integrates directly into BioLM’s design–make–test–learn workflows, where it is used to prioritize thermostable variants before synthesis, guide affinity maturation, and enforce developability gates. Via standardized APIs, TEMPRO outputs are composed with structure-derived metrics (AlphaFold2 pLDDT, NetSurfP-3.0), physicochemical filters (charge/pI, hydrophobicity, cysteine/disulfide patterns), solubility and aggregation predictors, and our generative design models to accelerate downselection. In practice, teams deploy TEMPRO 650M for early, high-throughput triage and cost control, then re-rank shortlists with larger TEMPRO variants (3B/15B) or a project-calibrated model before ordering, reducing experimental cycles and improving lead quality.

• Rapid triage at scale: score hundreds of thousands to tens of millions of sequences, gate by Tm thresholds, and focus synthesis on thermostable designs.

• Design constraints and risk reduction: combine Tm with developability filters and regional structure confidence to preserve stable frameworks while exploring CDR diversity.

• Closed-loop optimization: route prioritized variants to assay, feed DSC/DSF results back through the API, and escalate from 650M to higher-capacity models for final ranking and campaign decisions.

Related¶

• TEMPRO 3B – Higher‑capacity variant using ESM‑2 3B embeddings; run alongside 650M to ensemble or sanity‑check nanobody Tm estimates.

• ESM-2 650M – Generates the embeddings TEMPRO 650M is built on; precompute embeddings to speed batch Tm prediction and reuse across design iterations.

• Peptides – Computes physicochemical descriptors (aliphatic index, GRAVY, charge, etc.) used in TEMPRO ablations; useful for interpreting Tm drivers and pre/post‑filtering sequences.

• ABodyBuilder3 pLDDT – Antibody‑specific structure prediction with regional pLDDT; assess FR/CDR confidence and structural features that complement TEMPRO’s sequence‑based Tm outputs.

References¶

• Alvarez, J. A. E., & Dean, S. N. (2024). TEMPRO: nanobody melting temperature estimation model using protein embeddings. Scientific Reports.