GEMME is an MSA-based global epistatic model that infers protein mutational effects from sequence data alone, reconstructing full single-site (20×L) landscapes and optionally scoring user-specified single or multi-residue variants. The API accepts FASTA-style list MSAs or Stockholm MSAs (up to 2,048 residues, 20,000 sequences), runs CPU-based, training-free inference (typically minutes per protein), and returns normalized scores where more negative values indicate more deleterious substitutions for variant ranking and protein engineering or functional analysis.

Predict

Run the predict action for GEMME.

python 
from biolmai import BioLM
response = BioLM(
    entity="gemme",
    action="predict",
    params={},
    items=[{'msa': ['MKTVALIPYF',
              'MKTIALIPYF',
              'MKTLALIPYF',
              'MKSVALIPY-',
              'MRTVALIPY-',
              'MKTVALIPYF',
              'MKTVALIPYF',
              'MKTVALIPY-',
              'MNTVALIPYF',
              'MKTVALIPYF',
              'METVALIPYF',
              'MKTQALIPYF',
              'MKTVALIPYF',
              'MKTVALINPF',
              'MKTVALIPY-',
              'MKTVALIPYF',
              'MKTAALIPY-',
              'MKTVALIPYF',
              'MKTVALIPY-',
              'MKTVALIPYF']}]
)
print(response)
bash 
curl -X POST https://biolm.ai/api/v3/gemme/predict/ \
  -H "Authorization: Token YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
  "params": {},
  "items": [
    {
      "msa": [
        "MKTVALIPYF",
        "MKTIALIPYF",
        "MKTLALIPYF",
        "MKSVALIPY-",
        "MRTVALIPY-",
        "MKTVALIPYF",
        "MKTVALIPYF",
        "MKTVALIPY-",
        "MNTVALIPYF",
        "MKTVALIPYF",
        "METVALIPYF",
        "MKTQALIPYF",
        "MKTVALIPYF",
        "MKTVALINPF",
        "MKTVALIPY-",
        "MKTVALIPYF",
        "MKTAALIPY-",
        "MKTVALIPYF",
        "MKTVALIPY-",
        "MKTVALIPYF"
      ]
    }
  ]
}'
python 
import requests

url = "https://biolm.ai/api/v3/gemme/predict/"
headers = {
    "Authorization": "Token YOUR_API_KEY",
    "Content-Type": "application/json"
}
payload = {
      "params": {},
      "items": [
        {
          "msa": [
            "MKTVALIPYF",
            "MKTIALIPYF",
            "MKTLALIPYF",
            "MKSVALIPY-",
            "MRTVALIPY-",
            "MKTVALIPYF",
            "MKTVALIPYF",
            "MKTVALIPY-",
            "MNTVALIPYF",
            "MKTVALIPYF",
            "METVALIPYF",
            "MKTQALIPYF",
            "MKTVALIPYF",
            "MKTVALINPF",
            "MKTVALIPY-",
            "MKTVALIPYF",
            "MKTAALIPY-",
            "MKTVALIPYF",
            "MKTVALIPY-",
            "MKTVALIPYF"
          ]
        }
      ]
    }

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

url <- "https://biolm.ai/api/v3/gemme/predict/"
headers <- c("Authorization" = "Token YOUR_API_KEY", "Content-Type" = "application/json")
body <- list(
  params = list(),
  items = list(
    list(
      msa = list(
        "MKTVALIPYF",
        "MKTIALIPYF",
        "MKTLALIPYF",
        "MKSVALIPY-",
        "MRTVALIPY-",
        "MKTVALIPYF",
        "MKTVALIPYF",
        "MKTVALIPY-",
        "MNTVALIPYF",
        "MKTVALIPYF",
        "METVALIPYF",
        "MKTQALIPYF",
        "MKTVALIPYF",
        "MKTVALINPF",
        "MKTVALIPY-",
        "MKTVALIPYF",
        "MKTAALIPY-",
        "MKTVALIPYF",
        "MKTVALIPY-",
        "MKTVALIPYF"
      )
    )
  )
)

res <- POST(url, add_headers(.headers = headers), body = body, encode = "json")
print(content(res))
POST /api/v3/gemme/predict/

Predict endpoint for GEMME.

Request Headers:

Request

  • params (object, optional, default: empty object) — Optional parameters for GEMME scoring:

    • mutations (array of strings, optional) — List of mutation identifiers to score (e.g. “A42G”, “L100V”); each string must match ^[A-Z][1-9]\d*[A-Z]$; if omitted, a full mutational landscape matrix is returned

  • items (array of objects, required, min length: 1, max length: 1) — MSA inputs for scoring:

    • msa (array of strings, optional) — Multiple sequence alignment as a list of aligned sequences; must contain at least 2 and at most 20000 sequences; all sequences must have identical length (max length per sequence: 2048); characters limited to the alphabet defined by aa_extended plus "-" and "."

    • stockholm_msa (string, optional) — Multiple sequence alignment in raw Stockholm format; must start with "# STOCKHOLM" and contain a "//" terminator; must include at least one non-comment, non-blank sequence line with at least two whitespace-separated fields

    • (validation) — Exactly one of msa or stockholm_msa must be provided per item

Example request:

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

      {
  "params": {},
  "items": [
    {
      "msa": [
        "MKTVALIPYF",
        "MKTIALIPYF",
        "MKTLALIPYF",
        "MKSVALIPY-",
        "MRTVALIPY-",
        "MKTVALIPYF",
        "MKTVALIPYF",
        "MKTVALIPY-",
        "MNTVALIPYF",
        "MKTVALIPYF",
        "METVALIPYF",
        "MKTQALIPYF",
        "MKTVALIPYF",
        "MKTVALINPF",
        "MKTVALIPY-",
        "MKTVALIPYF",
        "MKTAALIPY-",
        "MKTVALIPYF",
        "MKTVALIPY-",
        "MKTVALIPYF"
      ]
    }
  ]
}
Status Codes:

Response

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

    • query_sequence (string) — Ungapped query (wildtype) amino acid sequence extracted from the input MSA

    • amino_acids (array of strings, length: 20) — Amino acid symbols defining the row order of the mutational landscape matrix

    • landscape (array of arrays of floats, shape: 20 × L, optional) — Full single-site mutational landscape matrix; rows follow amino_acids order, columns index sequence positions; more negative values correspond to more deleterious predicted mutational effects

    • mutation_scores (object, optional) — Dictionary of scores for requested mutations

      • <mutation> (float) — GEMME score for a single mutation key of the form [WT_AA][POSITION][MUT_AA] (e.g. "A42G"); more negative values correspond to more deleterious predicted mutational effects

Example response:

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

      {
  "results": [
    {
      "query_sequence": "MKTVALIPYF",
      "amino_acids": [
        "A",
        "C",
        "... (truncated for documentation)"
      ],
      "landscape": [
        [
          0.0,
          0.0,
          "... (truncated for documentation)"
        ],
        [
          0.0,
          0.0,
          "... (truncated for documentation)"
        ],
        "... (truncated for documentation)"
      ]
    }
  ]
}

Performance

  • Algorithmic runtime characteristics - GEMME has no trainable weights and performs no model training at inference; all computation is per-request, derived directly from the input MSA - For typical protein lengths (100–400 residues) with MSAs in the low thousands of sequences, GEMME completes full single-site mutational landscapes in minutes, dominated by phylogenetic-conservation and evolutionary-distance calculations - Time grows approximately linearly with sequence length and sub-linearly with MSA depth due to internal subsampling in the conservation stage; doubling sequence length roughly doubles wall-clock time, while a 10x increase in MSA depth increases time by less than 10x

  • Predictive accuracy relative to other BioLM models - On the ProteinGym substitution benchmark (217 deep mutational scanning assays), GEMME reaches an average Spearman correlation of ~0.455 between predicted and experimental mutational effects, matching the best-performing deep learning model in that benchmark family (TranceptEVE-L, ~0.456) while using a purely algorithmic, MSA-based approach - Compared with ESM-2 650M (zero-shot, sequence-only model), GEMME typically yields higher correlation to experimental data when a sufficiently deep and diverse MSA is available (ProteinGym: GEMME ~0.455 vs. ESM-2 650M ~0.414); for very shallow MSAs, sequence-only PLMs from the ESM-2 family tend to be more robust - Compared with generative MSA-based models such as DeepSequence and EVE-style models, GEMME achieves similar or better correlations on many proteins while avoiding per-protein neural-network training, which frequently requires hours of GPU time; GEMME remains CPU-only and inference-only

  • Performance relative to structure-based or structure-predictive models - Structure-predictive models (AlphaFold2, ESMFold, ABodyBuilder3, NanobodyBuilder) do not directly output mutational fitness effects; when used indirectly via structural proxies, they are orders of magnitude more computationally intensive than GEMME for exhaustive single-site scans - GEMME reaches mutational-effect accuracy comparable to large protein language models and some structure-aware predictors (e.g., AlphaMissense-style approaches) on many deep mutational scanning assays while using only sequence alignments and no structural input, giving a favorable accuracy-to-compute ratio when 3D information is not strictly required - When accurate structures and population data are integrated in extended frameworks such as ESCOTT/PRESCOTT, GEMME-derived signals can be combined with structural and allele-frequency features to further improve variant classification at the cost of higher runtime

  • Scaling behavior and practical implications - GEMME is well suited as a first-pass, CPU-based filter for large variant libraries or many proteins because no per-protein training is required; users can compute full single-site landscapes across panels of targets before applying more expensive GPU or structure-based models - In workflows where BioLM’s generative sequence models (e.g., ProGen2, Evo 2, DSM) design candidate variants, GEMME can efficiently rescore large design sets to remove variants inconsistent with evolutionary patterns without materially impacting pipeline throughput - GEMME performs best when the target has a reasonably deep and diverse MSA (hundreds to thousands of non-redundant homologs); in regimes with extremely sparse homology (or de novo designed proteins), alignment-free PLMs (ESM-2 family, Evo models) or structure-based stability predictors (ESM2StabP, ThermoMPNN, TEMPRO) are typically preferred

Applications

  • Enzyme mutational landscape mapping for stability and activity engineering: Using GEMME’s fast, MSA-based global epistatic model to score all 20×L single amino-acid substitutions in minutes on CPUs, teams can identify positions that are highly constrained (likely catalytic or structurally core) versus positions that tolerate diverse substitutions. This enables industrial protein engineers to focus directed evolution, semi-rational design, or library construction on permissive sites while deprioritizing positions GEMME predicts as globally sensitive, reducing library size and wet-lab screening costs for thermostable enzymes, process-optimized biocatalysts, or improved variants for harsh industrial conditions.

  • Prioritization of variants in large enzyme design campaigns and ML-guided evolution: In large combinatorial or error‑prone PCR libraries, GEMME can score single and multiple mutations (using its additive multi-mutation model on the same MSA-based landscape) to rank sequences before screening. Because the method is unsupervised and CPU-efficient, it can be integrated upstream of high-throughput assays or downstream predictive models to enrich for variants that are more likely to remain functional, improving hit rates and enabling more efficient machine-learning–guided design loops; this is especially valuable when only limited experimental data are available and GPU resources are constrained.

  • Differentiating functional hotspots from purely structural constraints in enzymes: By comparing GEMME’s evolutionary sensitivity scores with orthogonal stability metrics (e.g., ΔΔG from structure-based tools or experimental thermostability data), companies can separate residues that are essential for catalysis, binding, or regulation from those mainly required for fold stability. This helps in designing mutations that tune substrate specificity, cofactor dependence, or allosteric regulation without collapsing the fold, and in interpreting why certain industrial mutants maintain stability yet lose activity; this analysis is most informative for enzymes with good structural models and sufficiently deep MSAs.

  • Functional motif and interaction-site discovery in poorly structured or partially disordered protein regions: GEMME’s ability to infer constraint patterns directly from MSAs, without training a large neural network, makes it useful for highlighting short, mutation-sensitive segments within otherwise flexible regions of signaling or regulatory proteins used in cellular engineering. These evolutionarily sensitive segments can indicate transient interaction motifs (e.g., partner-binding sites, regulatory patches) that may not appear clearly in structure predictions, guiding mutagenesis to modulate protein–protein interfaces or domain–linker behavior; in extensively disordered regions with very shallow MSAs, signal can be weaker and should be combined with additional structural or biophysical evidence.

  • Pathogenicity-risk assessment for enzyme variants in clinical and companion-diagnostic settings: For therapeutic enzymes, metabolic enzymes, or enzyme targets in precision medicine, GEMME’s variant-effect scores derived purely from evolutionary constraints on an MSA can be used to prioritize missense variants that are likely to disrupt protein function, informing which positions and substitutions warrant experimental characterization or inclusion in diagnostic panels. Because the model predicts magnitude of functional disruption but not gain‑ versus loss‑of‑function and does not incorporate human population allele frequencies by default, its scores are best used as one line of evidence alongside structural modeling, population data, and functional assays, rather than as a standalone clinical classifier.

Limitations

  • Maximum input size and batch behavior: GEMME processes exactly one alignment per request (items length is fixed to 1 and effective batch_size is 1). Each aligned sequence in msa (or extracted from stockholm_msa) must have identical length <= 2048 (see max_sequence_len) and the MSA depth must be 2–20000 sequences (see max_msa_depth). Requests exceeding these limits, with inconsistent sequence lengths, or with characters outside the allowed MSA alphabet (20 amino acids, extended codes, '-', '.') are rejected at validation time.

  • MSA dependency and quality requirements: GEMME cannot operate on a single sequence alone; you must supply either msa (list of aligned sequences with the first sequence as the ungapped query/wildtype and no gaps in that first sequence) or stockholm_msa (a valid Stockholm-format MSA starting with '# STOCKHOLM 1.0' and ending with '//'). Prediction quality depends strongly on both the depth and evolutionary diversity of this alignment; very shallow MSAs (tens of sequences), highly redundant MSAs, or poorly aligned/fragmented regions will produce noisy or misleading scores in landscape and mutation_scores.

  • Scope of predictions and epistasis handling: The API is designed for mutational effect prediction on single-chain proteins and returns a 20×L mutational landscape (rows ordered as in amino_acids) and optional per-mutation values in mutation_scores when params.mutations is provided. GEMME does not explicitly model higher-order epistasis; scores for combinations of mutations are effectively additive/independent, so it is not optimal for designing or ranking densely mutated variants, highly epistatic proteins, or combinatorial variant libraries where interactions between sites dominate.

  • Biological coverage limitations: Because GEMME infers constraints from evolutionary history, it is poorly suited to proteins or regions without informative homologs (orphan proteins, de novo designs, very short peptides) and can be less discriminative in intrinsically disordered regions or highly flexible loops. It is also not antibody-specific and will not capture antibody CDR diversification patterns as well as dedicated antibody models; for multi-chain complexes, each chain must be scored separately and inter-chain effects are not modeled.

  • Interpretation of scores and use in decision-making: GEMME scores are relative effect estimates derived from sequence conservation; they do not indicate direction of functional change (e.g., gain-of-function vs. loss-of-function), do not provide calibrated probabilities or confidence intervals, and are not a substitute for experimental validation. For clinical or regulatory workflows, landscape and mutation_scores should be treated as one line of in silico evidence and combined with structural, functional, and population data; for engineering workflows, they should be used primarily for ranking and filtering rather than as absolute fitness predictions.

  • When another model may be preferable: If you cannot obtain a suitable MSA, or your target is very novel/poorly conserved, single-sequence models (e.g. protein language models) or structure-based predictors may be more appropriate. For tasks focused on 3D structure changes, binding interfaces, or explicit modeling of multi-chain or ligand interactions, structure-based or docking/ΔΔG methods are typically better suited than GEMME’s sequence-only landscape.

How We Use It

BioLM uses GEMME as a fast, alignment-based variant effect engine within protein design and optimization workflows, where it delivers interpretable mutational landscapes that indicate which positions are evolutionarily constrained, which can tolerate exploration, and which specific substitutions are most promising for synthesis and testing. GEMME outputs are combined with language-model scores (e.g. ESM/ESM2-style zero-shot fitness), structure-derived metrics (AlphaFold/ESMFold-based stability and interface features), and downstream generative and filtering stages so engineering programs for enzymes, antibodies, and other proteins can move from broad in silico search to focused, experiment-ready variant sets. GEMME scores are accessed through scalable, standardized APIs so ML engineers and bioinformaticians can inject evolutionary constraint signals directly into automated design loops and multi-round optimization strategies without additional GPU load, while wet-lab teams use these rankings to build smaller, higher-quality libraries and to distinguish mutations likely to impact stability, binding, or function.

  • In directed evolution and lab-in-the-loop campaigns, GEMME accelerates cycle design by ranking local mutational neighborhoods, deprioritizing variants with strong predicted loss of fitness, and enriching libraries for functional and improved sequences before screening.

  • In multi-model scoring pipelines, GEMME complements generative sequence models and structure-based predictors, enabling composite scores that balance evolutionary compatibility with predicted stability and application-specific properties such as activity in non-native conditions, specificity, or developability.

References