Pro-1 8B GRPO is the 8B-parameter GRPO-refined variant of Pro-1 (4-bit quantized LoRA on Llama 3.1), further trained with reinforcement learning using Rosetta REF2015-derived rewards plus creativity and specificity rewards to propose stabilizing sequence modifications. The API accepts a single amino acid sequence (10–2000 AA, best in the 50–500 AA globular range) plus structured enzyme context and prior mutagenesis data, and returns chain-of-thought reasoning, parsed mutation proposals, and an optional modified sequence. Relative to the SFT base variant (pro1-8b), it produces more creative, diverse designs (e.g. larger insertions/deletions) and attains a higher success rate on the author’s 40-enzyme stability benchmark; recommended as the default for aggressive thermostability optimization.

Generate

Run the generate action for Pro-1 8B GRPO.

python 
from biolmai import BioLM
response = BioLM(
    entity="pro1-8b-grpo",
    action="generate",
    params={'max_iterations': 2,
     'max_new_tokens': 2048,
     'temperature': 0.9,
     'top_p': 0.9,
     'seed': 42},
    items=[{'sequence': 'MVTGKQAKDLGVRLAVASGASGKTALA',
      'name': 'Thermostable kinase variant',
      'ec_number': '2.7.1.1',
      'reaction': [{'substrates': ['ATP', 'D-glucose'],
                    'products': ['ADP', 'D-glucose-6-phosphate']}],
      'general_information': 'Bacterial cytosolic kinase used in a '
                             'high-temperature bioreactor. Goal: improve '
                             'stability at 60-70°C without compromising '
                             'catalytic lysine residues. Homologs tolerate '
                             'hydrophobic core packing and helix capping '
                             'mutations.',
      'metal_ions': ['Mg2+'],
      'active_site_residues': ['K15', 'D19', 'E96'],
      'known_mutations': [{'mutation': 'A24V',
                           'effect': 'slightly increases Tm by ~2°C but '
                                     'reduces activity by 10%'},
                          {'mutation': 'G10P',
                           'effect': 'introduces local rigidity, increases '
                                     'aggregation above 55°C'}]}]
)
print(response)
bash 
curl -X POST https://biolm.ai/api/v3/pro1-8b-grpo/generate/ \
  -H "Authorization: Token YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
  "params": {
    "max_iterations": 2,
    "max_new_tokens": 2048,
    "temperature": 0.9,
    "top_p": 0.9,
    "seed": 42
  },
  "items": [
    {
      "sequence": "MVTGKQAKDLGVRLAVASGASGKTALA",
      "name": "Thermostable kinase variant",
      "ec_number": "2.7.1.1",
      "reaction": [
        {
          "substrates": [
            "ATP",
            "D-glucose"
          ],
          "products": [
            "ADP",
            "D-glucose-6-phosphate"
          ]
        }
      ],
      "general_information": "Bacterial cytosolic kinase used in a high-temperature bioreactor. Goal: improve stability at 60-70\u00b0C without compromising catalytic lysine residues. Homologs tolerate hydrophobic core packing and helix capping mutations.",
      "metal_ions": [
        "Mg2+"
      ],
      "active_site_residues": [
        "K15",
        "D19",
        "E96"
      ],
      "known_mutations": [
        {
          "mutation": "A24V",
          "effect": "slightly increases Tm by ~2\u00b0C but reduces activity by 10%"
        },
        {
          "mutation": "G10P",
          "effect": "introduces local rigidity, increases aggregation above 55\u00b0C"
        }
      ]
    }
  ]
}'
python 
import requests

url = "https://biolm.ai/api/v3/pro1-8b-grpo/generate/"
headers = {
    "Authorization": "Token YOUR_API_KEY",
    "Content-Type": "application/json"
}
payload = {
      "params": {
        "max_iterations": 2,
        "max_new_tokens": 2048,
        "temperature": 0.9,
        "top_p": 0.9,
        "seed": 42
      },
      "items": [
        {
          "sequence": "MVTGKQAKDLGVRLAVASGASGKTALA",
          "name": "Thermostable kinase variant",
          "ec_number": "2.7.1.1",
          "reaction": [
            {
              "substrates": [
                "ATP",
                "D-glucose"
              ],
              "products": [
                "ADP",
                "D-glucose-6-phosphate"
              ]
            }
          ],
          "general_information": "Bacterial cytosolic kinase used in a high-temperature bioreactor. Goal: improve stability at 60-70\u00b0C without compromising catalytic lysine residues. Homologs tolerate hydrophobic core packing and helix capping mutations.",
          "metal_ions": [
            "Mg2+"
          ],
          "active_site_residues": [
            "K15",
            "D19",
            "E96"
          ],
          "known_mutations": [
            {
              "mutation": "A24V",
              "effect": "slightly increases Tm by ~2\u00b0C but reduces activity by 10%"
            },
            {
              "mutation": "G10P",
              "effect": "introduces local rigidity, increases aggregation above 55\u00b0C"
            }
          ]
        }
      ]
    }

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

url <- "https://biolm.ai/api/v3/pro1-8b-grpo/generate/"
headers <- c("Authorization" = "Token YOUR_API_KEY", "Content-Type" = "application/json")
body <- list(
  params = list(
    max_iterations = 2,
    max_new_tokens = 2048,
    temperature = 0.9,
    top_p = 0.9,
    seed = 42
  ),
  items = list(
    list(
      sequence = "MVTGKQAKDLGVRLAVASGASGKTALA",
      name = "Thermostable kinase variant",
      ec_number = "2.7.1.1",
      reaction = list(
        list(
          substrates = list(
            "ATP",
            "D-glucose"
          ),
          products = list(
            "ADP",
            "D-glucose-6-phosphate"
          )
        )
      ),
      general_information = "Bacterial cytosolic kinase used in a high-temperature bioreactor. Goal: improve stability at 60-70°C without compromising catalytic lysine residues. Homologs tolerate hydrophobic core packing and helix capping mutations.",
      metal_ions = list(
        "Mg2+"
      ),
      active_site_residues = list(
        "K15",
        "D19",
        "E96"
      ),
      known_mutations = list(
        list(
          mutation = "A24V",
          effect = "slightly increases Tm by ~2°C but reduces activity by 10%"
        ),
        list(
          mutation = "G10P",
          effect = "introduces local rigidity, increases aggregation above 55°C"
        )
      )
    )
  )
)

res <- POST(url, add_headers(.headers = headers), body = body, encode = "json")
print(content(res))
POST /api/v3/pro1-8b-grpo/generate/

Generate endpoint for Pro-1 8B GRPO.

Request Headers:

Request

  • params (object, required) — Generation parameters:

    • max_iterations (int, range: 1-10, default: 3, required) — Maximum number of generation iterations

    • max_new_tokens (int, range: 128-16384, default: 8192, required) — Maximum number of new tokens to generate per iteration

    • temperature (float, range: >0.0-2.0, default: 0.95, required) — Sampling temperature

    • top_p (float, range: >0.0-1.0, default: 0.95, required) — Top-p (nucleus) sampling cutoff

    • seed (int, optional) — Random seed for reproducibility

  • items (array of objects, min: 1, max: 1, required) — Single protein to engineer:

    • sequence (string, min length: 10, max length: 2000, required) — Amino acid sequence using standard 1-letter codes

    • name (string, max length: 200, default: “”) — Protein name used in prompt context

    • ec_number (string, max length: 32, default: “”) — EC number for enzymes

    • reaction (array of objects, max: 8, default: []) — List of reactions with substrates and products

      • substrates (array of strings, default: []) — Reaction substrates

      • products (array of strings, default: []) — Reaction products

    • general_information (string, max length: 4000, default: “”) — Freeform biological context

    • metal_ions (array of strings, max: 20, default: []) — Metal ions or cofactors present

    • active_site_residues (array of strings, max: 200, default: []) — Active site residues that must not be modified

    • known_mutations (array of objects, max: 50, default: []) — Prior mutagenesis results

      • mutation (string, max length: 16, required) — Mutation in standard notation

      • effect (string, max length: 500, required) — Observed effect of the mutation

Example request:

http 
POST /api/v3/pro1-8b-grpo/generate/ HTTP/1.1
Host: biolm.ai
Authorization: Token YOUR_API_KEY
Content-Type: application/json

      {
  "params": {
    "max_iterations": 2,
    "max_new_tokens": 2048,
    "temperature": 0.9,
    "top_p": 0.9,
    "seed": 42
  },
  "items": [
    {
      "sequence": "MVTGKQAKDLGVRLAVASGASGKTALA",
      "name": "Thermostable kinase variant",
      "ec_number": "2.7.1.1",
      "reaction": [
        {
          "substrates": [
            "ATP",
            "D-glucose"
          ],
          "products": [
            "ADP",
            "D-glucose-6-phosphate"
          ]
        }
      ],
      "general_information": "Bacterial cytosolic kinase used in a high-temperature bioreactor. Goal: improve stability at 60-70\u00b0C without compromising catalytic lysine residues. Homologs tolerate hydrophobic core packing and helix capping mutations.",
      "metal_ions": [
        "Mg2+"
      ],
      "active_site_residues": [
        "K15",
        "D19",
        "E96"
      ],
      "known_mutations": [
        {
          "mutation": "A24V",
          "effect": "slightly increases Tm by ~2\u00b0C but reduces activity by 10%"
        },
        {
          "mutation": "G10P",
          "effect": "introduces local rigidity, increases aggregation above 55\u00b0C"
        }
      ]
    }
  ]
}
Status Codes:

Response

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

    • reasoning (string) — Full chain-of-thought reasoning trace for the generated proposal, including any model-internal tags

    • mutations (array of objects) — Parsed mutation proposals for the generated proposal

      • mutation (string) — Single mutation in standard notation (e.g. “K116E”), combining original residue, 1-based position, and new residue

      • rationale (string) — Text rationale describing the basis for the proposed mutation

    • modified_sequence (string, optional) — Amino acid sequence of the designed variant with the parsed mutations applied, 10-2000 characters (unambiguous 1-letter codes)

Example response:

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

      {
  "results": [
    {
      "reasoning": "\n\n<think>\n\nConsidering the thermostable kinase variant's need for improved stability at 60-70\u00b0C without compromising catalytic lysine residues, we can explore mutations that enhance the protein's ther... (truncated for documentation)",
      "mutations": [
        {
          "mutation": "A24V",
          "rationale": "tions, we'll build upon the information provided. First, let's analyze the mutations provided: 1. A24V: This m ... (truncated for documentation)"
        },
        {
          "mutation": "G10P",
          "rationale": "explore alternative mutations that could potentially improve both thermostability and activity. 2. G10P: This ... (truncated for documentation)"
        },
        "... (truncated for documentation)"
      ],
      "modified_sequence": "MVTGKQAKDLGVRLAVASGASGKTALA"
    },
    {
      "reasoning": "\n\n<think>\nTo optimize the stability of the thermostable kinase variant at 60-70\u00b0C, we need to consider mutations that improve hydrophobic core packing, helix capping, and surface exposure of hydrophob... (truncated for documentation)",
      "mutations": [
        {
          "mutation": "G10P",
          "rationale": "o counteract this, we can try a mutation that maintains rigidity but avoids the negative effects of G10P. A mu ... (truncated for documentation)"
        },
        {
          "mutation": "A24V",
          "rationale": "0 could provide a more stable conformation with improved flexibility. 2. Position 24: The mutation A24V has sl ... (truncated for documentation)"
        },
        "... (truncated for documentation)"
      ],
      "modified_sequence": "MVTGKQAKDLGVRLAVASGASGKTALA"
    }
  ]
}

Performance

  • Model class and deployment: - Decoder-only reasoning LLM (Llama 3.1–8B with LoRA adapters, 4-bit quantized) specialized for protein thermostability engineering - Hosted on NVIDIA data-center GPUs with ≥8–12 GB effective memory per replica, giving latency similar to mid-sized generative protein models (e.g., ProGen2 Medium), and substantially lower cost/latency than structure-prediction pipelines that invoke ESMFold or AlphaFold2

  • Relative speed and resource use within BioLM: - Much faster per design iteration than structure predictors (ESMFold, AlphaFold2, ProstT5 AA2Fold), because Pro-1 operates purely at the sequence/text level and does not perform 3D folding or relaxation at inference - Modestly slower per token than generative sequence models (ProGen2 family, Evo 1.5 / Evo 2) due to longer reasoning traces, but practical per-protein design time is comparable since Pro-1 often needs only a small number of high-quality iterations - Significantly slower than lightweight scoring / effect-prediction models (ESM2StabP, Pro4S Regression, GEMME), which run single-pass scoring instead of full autoregressive reasoning; use Pro-1 when interpretability and proposal quality are more important than raw throughput

  • Predictive performance for stability optimization: - On the 40-enzyme benchmark described by the original author, the 8B creativity- and specificity-tuned variant attains a 47% success rate (improved Rosetta REF2015 energy), matching or slightly exceeding ProteinMPNN and EvoProtGrad under comparable evaluation despite using less curated biological training data - In a wet-lab FGF-1 case study, Pro-1 identified a single mutation (K116E) with a +24°C melting-temperature improvement while preserving receptor binding, demonstrating real-world thermostability performance comparable to physics-heavy design pipelines at substantially lower inference compute

  • Scaling, variability, and pipeline integration: - Total latency and GPU usage grow approximately linearly with max_iterations (each iteration runs a full reasoning and generation pass); in practice, 2–3 iterations often balance diversity and turnaround time - As a stochastic generative model, outputs vary across runs; the seed parameter improves reproducibility and can be used with different temperatures to explore alternative local optima in sequence space - For high-value single targets with limited wet-lab capacity, Pro-1 often yields a higher fraction of experimentally useful variants per synthesized sequence than brute-force generative approaches; for very large libraries, it works well as a seeding step before expansion with higher-throughput models such as ProteinMPNN or ThermoMPNN and downstream ranking with ESM2StabP or physics-based scores

Applications

  • Thermostability optimization for industrial enzymes and biocatalysts, using Pro-1’s GRPO-tuned reasoning over single-chain amino acid sequences to propose stabilizing point mutations and small insertions/deletions that are predicted (via REF2015 during training) to lower folding energy, enabling companies in carbon capture, detergents, food processing, and fine chemicals to raise enzyme Tm into process-relevant ranges while keeping active-site residues specified in active_site_residues intact; most effective for soluble globular enzymes without metals or transmembrane domains and should be followed by independent structure prediction/Rosetta-style rescoring and experimental validation to avoid REF2015 “hacks”

  • Iterative enzyme variant design in lab-in-the-loop campaigns, where process development or protein engineering teams feed prior mutagenesis results into the known_mutations field and use the API’s single-protein generator endpoint to obtain the next generation of variants with explicit biochemical rationale (e.g., stronger helix capping, added surface salt bridges), reducing the number of design–build–test cycles needed to converge on robust, manufacturable enzyme leads for biomanufacturing or formulation at scale

  • Stabilization of non-enzymatic globular proteins (e.g., small cytokine-like scaffolds, soluble receptors, binding domains) when only sequence and limited functional context are available, by leveraging Pro-1’s language-model grounding in protein stability mechanisms to suggest targeted mutations and compact insertions that are likely to improve folding stability while avoiding user-specified functional sites, enabling biotech teams to harden sequence-based reagents for storage, shipping, and subcutaneous administration; less suitable for heavily glycosylated, membrane-anchored, or multi-chain constructs where training data and ESMFold- plus REF2015-derived rewards are weak

  • Early-stage protein scaffold engineering and developability-oriented screening, where platform teams use Pro-1 to explore stability-focused redesigns of single-chain scaffolds in the 50–500 amino acid range before committing to large combinatorial libraries or costly directed evolution, taking advantage of the model’s interpretable, literature-style reasoning traces to highlight core packing, loop rigidity, and surface charge patterns that are likely to generalize across payloads or fusions, while relying on downstream tools and wet-lab assays to address properties outside thermal stability such as aggregation, solubility, or activity changes

  • Automated proposal generation for in silico stability screening pipelines, integrating Pro-1 as a front-end sequence proposal engine that uses its GRPO-trained policy to sample diverse, higher-stability candidates around a production enzyme or protein of interest (within the 10–2000 amino acid API limits, with best performance in the 50–500 range), which can then be batch-scored with structure-based or embedding-based models and filtered for synthesis; useful for high-throughput computational design workflows, with the caveat that performance degrades for metalloproteins, membrane proteins, and very short or very long sequences outside the model’s demonstrated operating regime

Limitations

  • Batch size & throughput: Pro-1 is designed for single-protein, interactive use. Pro1GenerateRequest.items must contain exactly one Pro1ProteinData entry (min_length=1, max_length=1, effective batch_size=1). High-throughput library design (thousands–millions of variants) should use faster, batched sequence models (e.g. ESM-style scorers or ProteinMPNN) and reserve Pro-1 for later-stage, low-volume reasoning and refinement.

  • Sequence length & protein type: Pro1ProteinData.sequence must be between 10 and 2000 amino acids (min_length=10, max_length=2000). Performance is best for globular proteins in the ~50–500 AA range; outputs degrade for very short peptides (<20 AA), long multi-domain or multi-chain constructs (>500 AA), membrane proteins, intrinsically disordered regions, and many metalloproteins (even if metals are listed in metal_ions). For antibody CDR design, membrane protein optimization, or peptide therapeutics, consider models specialized for those molecule classes.

  • Stochastic reasoning & partial parsing: Pro-1 is a generative reasoning model, not a deterministic mutagenesis engine. Each iteration (controlled by params.max_iterations and sampling parameters temperature and top_p) can return different reasoning traces and mutation sets. mutations and modified_sequence are parsed from free-form text and may be empty or inconsistent (e.g. mutation mentioned in reasoning but not correctly applied to modified_sequence); always verify that parsed mutations match the returned modified_sequence before downstream use.

  • Physics-based reward limitations: Training is aligned to Rosetta REF2015 scores on ESMFold-predicted structures, which only approximate real stability. Pro-1 can propose variants that look favorable in silico but are unstable, misfolded, or non-functional experimentally, and it may implicitly “hack” this energy function. general_information, active_site_residues and known_mutations help steer designs but do not guarantee preservation of activity, binding, solubility, or expression; external stability/activity predictors and wet-lab assays are still required.

  • Hallucinated literature & soft biological context: The reasoning field often includes mechanistic explanations and literature-style references. These citations are not validated and are frequently hallucinated; treat them only as hypotheses. Context fields such as ec_number, reaction, metal_ions and general_information are used as soft prompt guidance, not hard constraints, so Pro-1 may still suggest changes that conflict with known mechanisms or cofactors. Domain expert review is necessary for critical applications.

  • Non-optimal use cases: Pro-1 targets thermostability-oriented sequence changes with interpretable reasoning. It is not optimal when rapid scoring of very large libraries, accurate 3D structural modeling, binding affinity optimization, or antibody-specific loop modeling are primary goals. In those settings, use structure-prediction models (e.g. ESMFold/Boltz), inverse-folding or diffusion-based designers, or antibody-focused models for bulk design, and apply Pro-1 only to a small subset of candidates where detailed stability reasoning is valuable.

How We Use It

Pro-1 integrates into stability-focused protein engineering programs as a reasoning layer that turns broad thermostability goals into concrete, testable variant proposals. Via scalable, standardized APIs, teams connect assay history, prior mutagenesis, and structural context with downstream structure prediction and scoring (e.g., ESMFold/Boltz, Rosetta-style energy functions, ESM2 embeddings, ProteinMPNN redesign), enabling iterative design cycles where Pro-1 suggests and rationalizes mutations, other models stress-test candidates, and new assay results feed back into subsequent calls to converge on target Tm and manufacturability profiles with fewer synthesis rounds.

  • Enables stability-first design tracks for enzymes, globular proteins, and early antibody scaffolds that plug directly into existing variant-ranking and synthesis workflows.

  • Accelerates decision-making by turning unstructured experimental notes and domain knowledge into structured, model-driven mutation proposals that can be orchestrated and audited across teams through API-based pipelines.

References