ThermoMPNN-D¶

Structure-based prediction of protein stability changes (ddG) for single and double mutations, with epistatic interaction modeling via transfer learning from ProteinMPNN.

License: MIT · Molecules: protein · Tasks: stability_prediction

Actions: predict · Variants: 1

At a glance¶

Use it when

You need to evaluate pairs of simultaneous mutations for synergistic stability effects in a combinatorial protein engineering campaign
You are performing directed evolution and want to predict which two-mutation combinations to test experimentally, prioritizing pairs with positive epistasis
You want to compare additive vs epistatic double-mutation predictions to identify residue pairs with significant non-additive interactions
You have a 3D structure and need to understand how mutations at spatially proximal positions interact to affect overall stability
You are engineering thermostable enzymes and need to combine multiple stabilizing mutations while checking for antagonistic epistatic effects
You want single-mutation ddG predictions with the same model that also supports double-mutation analysis for a consistent prediction framework

Strengths

Uniquely supports double (paired) mutation ddG prediction with an explicit epistatic interaction mode, going beyond simple additive approximations to capture synergistic and antagonistic mutation combinations
Three prediction modes in a single model: single-mutation ddG, additive double-mutation ddG, and epistatic double-mutation ddG -- enabling comparison between additive and non-additive effects
Structure-based GNN architecture (ProteinMPNN backbone) directly models the 3D spatial context of mutation pairs, detecting interactions between spatially proximal residues
Configurable CA-CA distance threshold (default 5.0 Angstroms) for double-mutation scanning focuses predictions on physically interacting residue pairs where epistasis is most likely
MIT license with no commercial restrictions -- suitable for pharmaceutical protein engineering pipelines
Builds on the well-validated ThermoMPNN framework with a dedicated double-mutation extension, maintaining consistency with single-mutation predictions

Limitations

Higher memory requirement than ThermoMPNN (12 GB vs 8 GB RAM) because it loads both single-mutation and epistatic model checkpoints simultaneously
Published benchmarks show that current stability models, including ThermoMPNN-D, struggle to capture epistatic interactions -- additive models perform comparably except on stabilizing double mutations
Requires a 3D structure (PDB format) as input -- cannot operate from sequence alone
Limited to single-chain proteins up to 1024 residues -- no multi-chain complex support
Double-mutation scanning scales quadratically O(N^2) with sequence length, making full scans computationally expensive for longer proteins
Membrane proteins and intrinsically disordered proteins are not well-supported due to training data bias and the requirement for a folded reference structure
No uncertainty quantification on individual predictions -- reported epistatic effects may be within noise margins

Reach for something else when

You only need single-mutation ddG predictions -- use the simpler ThermoMPNN model, which has lower memory requirements and equivalent single-mutation accuracy
You do not have a 3D structure available -- use TemBERTure for sequence-only thermostability assessment
You need to evaluate combinations of three or more simultaneous mutations -- ThermoMPNN-D is limited to pairs; consider SPURS for multi-mutation ddG
You need global melting temperature (Tm) rather than per-mutation ddG -- use TemBERTure
You are working with very long proteins where quadratic double-mutation scanning is impractical -- consider prioritizing candidate pairs using single-mutation scans first
You need evolution-based variant effect scores from multiple sequence alignments rather than structure-based stability predictions

Alternatives

Model	Better when	Worse when
thermompnn	Simpler and lighter (8 GB vs 12 GB RAM) with equivalent single-mutation ddG accuracy on the same PDB structure; preferred when only single-point ddG is needed.	Single-point mutations only -- no double-mutation or epistatic interaction modeling.
spurs	Accepts a structure and handles arbitrary multi-mutation combinations (not just pairs) and full DMS matrices via ESM2+GNN; broader than pairwise scope.	Higher resource use, requires an accompanying sequence alongside the structure, and lacks ThermoMPNN-D's purpose-built pairwise epistatic mode benchmarked for double mutations.

API & schema¶

Call an action with POST /api/v1/{slug}/{action} — the request envelope is {"items": [...], "params": {...}} and a success returns {"results": [...]}. See the HTTP API page for the base URL, error shape, and full contract.

`predict`¶

Call it

curl -X POST http://127.0.0.1:8000/api/v1/thermompnn-d/predict \
  -H "Content-Type: application/json" \
  -d '{
  "items": [
    {
      "pdb": "<contents of a .pdb file>"
    }
  ]
}'

Request — ThermoMPNNDPredictRequest

Field	Type	Required	Constraints	Description
`params`	ThermoMPNNDPredictParams	no		Optional parameters controlling this action (defaults are used when omitted).
`items`	list[ThermoMPNNDPredictRequestItem]	yes	items 1–1	Batch of inputs to process in a single request. Up to 1 PDB per request.

Nested types

ThermoMPNNDMode

Allowed values: single, additive, epistatic

ThermoMPNNDPredictParams

Field	Type	Required	Constraints	Description
`mode`	ThermoMPNNDMode	no	default `single`	Prediction mode: 'single' for individual mutation ddG, 'additive' for summed single-mutation effects, 'epistatic' for full pairwise interaction modeling.
`chain`	string \| null	no		Chain ID to use for prediction. If not specified, uses first chain in PDB.
`distance`	number	no	≥0.0; default `5.0`	Distance threshold (Angstroms) for filtering double mutations. Only mutations within this distance are considered.
`threshold`	number	no	default `-0.5`	ddG threshold (kcal/mol) for filtering results. Only mutations with ddG <= threshold are returned. Set to a high value (e.g., 100) to return all mutations.

ThermoMPNNDPredictRequestItem

Field	Type	Required	Constraints	Description
`pdb`	string	yes	len 1–2500000	Input structure in PDB format.
`mutations`	list[string] \| null	no		List of mutations. For single mode: format 'WT{position}MUT' (e.g., 'A100V'). For double modes: format 'WT1{pos1}MUT1:WT2{pos2}MUT2' (e.g., 'A100V:B200L'). If not provided, a site-saturation mutagenesis (SSM) scan will be performed.

Raw JSON Schema

{
  "$defs": {
    "ThermoMPNNDMode": {
      "enum": [
        "single",
        "additive",
        "epistatic"
      ],
      "title": "ThermoMPNNDMode",
      "type": "string"
    },
    "ThermoMPNNDPredictParams": {
      "additionalProperties": false,
      "description": "Parameters for ThermoMPNN-D prediction",
      "properties": {
        "mode": {
          "$ref": "#/$defs/ThermoMPNNDMode",
          "default": "single",
          "description": "Prediction mode: 'single' for individual mutation ddG, 'additive' for summed single-mutation effects, 'epistatic' for full pairwise interaction modeling."
        },
        "chain": {
          "anyOf": [
            {
              "type": "string"
            },
            {
              "type": "null"
            }
          ],
          "default": null,
          "description": "Chain ID to use for prediction. If not specified, uses first chain in PDB.",
          "title": "Chain"
        },
        "distance": {
          "default": 5.0,
          "description": "Distance threshold (Angstroms) for filtering double mutations. Only mutations within this distance are considered.",
          "minimum": 0.0,
          "title": "Distance",
          "type": "number"
        },
        "threshold": {
          "default": -0.5,
          "description": "ddG threshold (kcal/mol) for filtering results. Only mutations with ddG <= threshold are returned. Set to a high value (e.g., 100) to return all mutations.",
          "title": "Threshold",
          "type": "number"
        }
      },
      "title": "ThermoMPNNDPredictParams",
      "type": "object"
    },
    "ThermoMPNNDPredictRequestItem": {
      "additionalProperties": false,
      "properties": {
        "pdb": {
          "description": "Input structure in PDB format.",
          "maxLength": 2500000,
          "minLength": 1,
          "title": "Pdb",
          "type": "string"
        },
        "mutations": {
          "anyOf": [
            {
              "items": {
                "type": "string"
              },
              "type": "array"
            },
            {
              "type": "null"
            }
          ],
          "default": null,
          "description": "List of mutations. For single mode: format 'WT{position}MUT' (e.g., 'A100V'). For double modes: format 'WT1{pos1}MUT1:WT2{pos2}MUT2' (e.g., 'A100V:B200L'). If not provided, a site-saturation mutagenesis (SSM) scan will be performed.",
          "title": "Mutations"
        }
      },
      "required": [
        "pdb"
      ],
      "title": "ThermoMPNNDPredictRequestItem",
      "type": "object"
    }
  },
  "additionalProperties": false,
  "properties": {
    "params": {
      "$ref": "#/$defs/ThermoMPNNDPredictParams",
      "description": "Optional parameters controlling this action (defaults are used when omitted)."
    },
    "items": {
      "description": "Batch of inputs to process in a single request. Up to 1 PDB per request.",
      "items": {
        "$ref": "#/$defs/ThermoMPNNDPredictRequestItem"
      },
      "maxItems": 1,
      "minItems": 1,
      "title": "Items",
      "type": "array"
    }
  },
  "required": [
    "items"
  ],
  "title": "ThermoMPNNDPredictRequest",
  "type": "object"
}

Response — ThermoMPNNDPredictResponse

Field	Type	Required	Constraints	Description
`results`	list[ThermoMPNNDPredictResponseItem]	yes		Predicted ddG results, one entry per evaluated mutation (single or double).

Nested types

ThermoMPNNDPredictResponseItem

Field	Type	Required	Description
`mutation`	string	yes	Mutation string in WT{pos}MUT format (single) or WT1{pos1}MUT1:WT2{pos2}MUT2 format (double).
`position`	integer \| null	no	Residue position for single mutations (1-indexed).
`position1`	integer \| null	no	First residue position for double mutations (1-indexed).
`position2`	integer \| null	no	Second residue position for double mutations (1-indexed).
`wildtype`	string \| null	no	Wildtype amino acid for single mutations.
`wildtype1`	string \| null	no	First wildtype amino acid for double mutations.
`wildtype2`	string \| null	no	Second wildtype amino acid for double mutations.
`mutation_aa`	string \| null	no	Mutant amino acid for single mutations.
`mutation_aa1`	string \| null	no	First mutant amino acid for double mutations.
`mutation_aa2`	string \| null	no	Second mutant amino acid for double mutations.
`ddg`	number	yes	Predicted change in free energy (ddG) in kcal/mol.
`distance`	number \| null	no	CA-CA distance (Angstroms) between the two mutation sites for double mutations.

Raw JSON Schema

{
  "$defs": {
    "ThermoMPNNDPredictResponseItem": {
      "description": "Response item for a mutation prediction (single or double)",
      "properties": {
        "mutation": {
          "description": "Mutation string in WT{pos}MUT format (single) or WT1{pos1}MUT1:WT2{pos2}MUT2 format (double).",
          "title": "Mutation",
          "type": "string"
        },
        "position": {
          "anyOf": [
            {
              "type": "integer"
            },
            {
              "type": "null"
            }
          ],
          "default": null,
          "description": "Residue position for single mutations (1-indexed).",
          "title": "Position"
        },
        "position1": {
          "anyOf": [
            {
              "type": "integer"
            },
            {
              "type": "null"
            }
          ],
          "default": null,
          "description": "First residue position for double mutations (1-indexed).",
          "title": "Position1"
        },
        "position2": {
          "anyOf": [
            {
              "type": "integer"
            },
            {
              "type": "null"
            }
          ],
          "default": null,
          "description": "Second residue position for double mutations (1-indexed).",
          "title": "Position2"
        },
        "wildtype": {
          "anyOf": [
            {
              "type": "string"
            },
            {
              "type": "null"
            }
          ],
          "default": null,
          "description": "Wildtype amino acid for single mutations.",
          "title": "Wildtype"
        },
        "wildtype1": {
          "anyOf": [
            {
              "type": "string"
            },
            {
              "type": "null"
            }
          ],
          "default": null,
          "description": "First wildtype amino acid for double mutations.",
          "title": "Wildtype1"
        },
        "wildtype2": {
          "anyOf": [
            {
              "type": "string"
            },
            {
              "type": "null"
            }
          ],
          "default": null,
          "description": "Second wildtype amino acid for double mutations.",
          "title": "Wildtype2"
        },
        "mutation_aa": {
          "anyOf": [
            {
              "type": "string"
            },
            {
              "type": "null"
            }
          ],
          "default": null,
          "description": "Mutant amino acid for single mutations.",
          "title": "Mutation Aa"
        },
        "mutation_aa1": {
          "anyOf": [
            {
              "type": "string"
            },
            {
              "type": "null"
            }
          ],
          "default": null,
          "description": "First mutant amino acid for double mutations.",
          "title": "Mutation Aa1"
        },
        "mutation_aa2": {
          "anyOf": [
            {
              "type": "string"
            },
            {
              "type": "null"
            }
          ],
          "default": null,
          "description": "Second mutant amino acid for double mutations.",
          "title": "Mutation Aa2"
        },
        "ddg": {
          "description": "Predicted change in free energy (ddG) in kcal/mol.",
          "title": "Ddg",
          "type": "number"
        },
        "distance": {
          "anyOf": [
            {
              "type": "number"
            },
            {
              "type": "null"
            }
          ],
          "default": null,
          "description": "CA-CA distance (Angstroms) between the two mutation sites for double mutations.",
          "title": "Distance"
        }
      },
      "required": [
        "mutation",
        "ddg"
      ],
      "title": "ThermoMPNNDPredictResponseItem",
      "type": "object"
    }
  },
  "properties": {
    "results": {
      "description": "Predicted ddG results, one entry per evaluated mutation (single or double).",
      "items": {
        "$ref": "#/$defs/ThermoMPNNDPredictResponseItem"
      },
      "title": "Results",
      "type": "array"
    }
  },
  "required": [
    "results"
  ],
  "title": "ThermoMPNNDPredictResponse",
  "type": "object"
}

Usage¶

One-line summary: Structure-based prediction of protein stability changes (ddG) for single and double mutations, with epistatic interaction modeling via transfer learning from ProteinMPNN.

Overview¶

ThermoMPNN-D is a graph neural network developed by Dieckhaus and Kuhlman (2024) that predicts changes in protein thermostability (ddG in kcal/mol) for both single and double mutations. It extends ThermoMPNN by adding explicit modeling of epistatic (non-additive) interactions between paired mutation sites. The model supports three modes: single mutation prediction, additive double mutation estimation, and full epistatic double mutation prediction with distance-based filtering.

Architecture¶

Property	Value
Architecture	Message-passing neural network (GNN) with epistatic module
Base model	ProteinMPNN (v_48_020)
Models loaded	2 (single + epistatic)
Input	PDB structure + mutations
Output	ddG in kcal/mol (+ CA-CA distance for doubles)
Max sequence length	1024 residues
Batch size	1 PDB per request

Capabilities & Limitations¶

CAN be used for: - Single mutation ddG prediction given a PDB structure - Additive double mutation ddG estimation (sum of individual effects) - Epistatic double mutation ddG prediction (non-additive interactions) - Full SSM scans in all three modes - Distance-based filtering of double mutation pairs - Threshold-based filtering of results by ddG value

CANNOT be used for: - Sequence-only prediction (requires PDB structure input) - Triple or higher-order mutations - Proteins longer than 1024 residues - Membrane protein stability in lipid bilayer context

Other considerations: - Single mutations: format WT{position}MUT (e.g., A100V) - Double mutations: format WT1{pos1}MUT1:WT2{pos2}MUT2 (e.g., A100V:B200L) - When mutations is null, a full SSM scan is performed in the selected mode - Distance threshold (default 5.0 A) filters double mutations by CA-CA distance - ddG threshold (default -0.5 kcal/mol) filters results; set to high value (100) to return all

Usage Examples¶

from models.thermompnn_d.schema import (
    ThermoMPNNDMode,
    ThermoMPNNDPredictParams,
    ThermoMPNNDPredictRequest,
    ThermoMPNNDPredictRequestItem,
)

# Single mutation prediction
single_request = ThermoMPNNDPredictRequest(
    params=ThermoMPNNDPredictParams(
        mode=ThermoMPNNDMode.SINGLE,
        chain="A",
    ),
    items=[
        ThermoMPNNDPredictRequestItem(
            pdb=pdb_string,
            mutations=["M1V", "V2A"],
        )
    ],
)

# Epistatic double mutation prediction
epistatic_request = ThermoMPNNDPredictRequest(
    params=ThermoMPNNDPredictParams(
        mode=ThermoMPNNDMode.EPISTATIC,
        chain="A",
        distance=5.0,
        threshold=-0.5,
    ),
    items=[
        ThermoMPNNDPredictRequestItem(
            pdb=pdb_string,
            mutations=["M1V:V2A"],
        )
    ],
)

# Full SSM scan in epistatic mode
ssm_request = ThermoMPNNDPredictRequest(
    params=ThermoMPNNDPredictParams(
        mode=ThermoMPNNDMode.EPISTATIC,
        chain="A",
        distance=5.0,
        threshold=100.0,  # Return all mutations
    ),
    items=[
        ThermoMPNNDPredictRequestItem(
            pdb=pdb_string,
            mutations=None,
        )
    ],
)

Architecture & training¶

Architecture¶

Model Type & Innovation¶

ThermoMPNN-D extends ThermoMPNN to predict stability changes (ddG) for both single and double (paired) mutations. The key innovation is the ability to model epistatic interactions between two simultaneous mutations -- something that additive models fundamentally cannot capture. The model uses two separate architectures:

Single/Additive model: Predicts ddG for individual mutations. In additive mode, double mutation effects are estimated by summing two single-mutation predictions.
Epistatic model: A specialized architecture that explicitly models pairwise interactions between two mutation sites, capturing non-additive (epistatic) effects.

Parameters & Layers¶

Component	Details
Architecture	Message-passing neural network (GNN) with epistatic interaction module
Base model	ProteinMPNN (v_48_020)
Single checkpoint	ThermoMPNN-ens1.ckpt
Epistatic checkpoint	ThermoMPNN-D-ens1.ckpt
Input	PDB structure + mutations
Output	ddG in kcal/mol
Distance filtering	CA-CA distance threshold (default 5.0 Angstroms)

Training Data¶

The model was trained on experimental stability measurements including both single and double mutation data, with the epistatic model specifically trained on paired mutation datasets. Full training dataset details are described in Dieckhaus & Kuhlman (2024).

Loss Function & Objective¶

Regression loss for predicting ddG (change in Gibbs free energy of unfolding) in kcal/mol. The epistatic model additionally learns pairwise interaction terms between mutation sites.

Tokenization / Input Processing¶

Input processing follows the same pattern as ThermoMPNN:

PDB parsing: Structure parsed using ThermoMPNN-D's load_pdb utility
Chain selection: Target chain extracted
Distance matrix: CA-CA distance matrix computed for filtering double mutations
Mutation encoding: Single format WT{pos}MUT (e.g., A100V); double format WT1{pos1}MUT1:WT2{pos2}MUT2 (e.g., A100V:B200L)

Performance & Benchmarks¶

Published Benchmarks¶

Benchmark results are described in Dieckhaus & Kuhlman (2024) (bioRxiv 2024.10.10.617658).

BioLM Verification Results¶

Integration tests verify response format across all three modes (single, additive, epistatic), checking for mode-appropriate fields and numeric ddG values. Six test cases cover all modes with both targeted mutations and SSM scans.

Comparison to Alternatives¶

Model	Task	Modes	Advantage
ThermoMPNN-D	Single + double ddG	Single, additive, epistatic	Handles epistatic double mutations
ThermoMPNN	Single ddG only	Single	Simpler, lower resource usage
TemBERTure	Global Tm	Classification, regression	Sequence-only, no structure needed

Strengths & Limitations¶

Pros¶

Three prediction modes: single, additive, and epistatic
Captures non-additive (epistatic) interactions between paired mutations
Distance-based filtering reduces computational cost for double mutations
Threshold-based filtering returns only mutations meeting ddG criteria
Full SSM scan support for all three modes

Cons¶

Requires PDB structure input
Loads two models (single + epistatic), requiring more memory (12 GB)
Batch size limited to 1 PDB per request
Epistatic SSM scans can be computationally expensive for large proteins
Maximum sequence length of 1024 residues

Known Failure Modes¶

Very large proteins with many close-contact residue pairs can produce very large output sets in epistatic SSM mode
Distance threshold filtering is based on CA-CA distance, which may not capture all relevant interactions
Missing residues in PDB structures may affect predictions at nearby positions

Implementation Details¶

Inference Pipeline¶

Request --> Validate PDB + mutations + mode
  --> Write PDB to temp file
  --> Parse PDB (load_pdb)
  --> Select chain
  --> Route by mode:
      Single:    run_single_ssm --> format_output_single
      Additive:  run_single_ssm --> format_output_double (additive)
      Epistatic: run_epistatic_ssm (batched, distance-filtered)
  --> Apply threshold filter
  --> Format response with distances
  --> Cleanup temp files

Memory & Compute Profile¶

Resource	Value
GPU	T4
Memory	12 GB (loads 2 models)
CPU	2 cores

Determinism & Reproducibility¶

Torch manual seed: Yes (42)
CUDA manual seed: Yes (42)
Both models set to eval mode: Yes
Inference under torch.no_grad(): Yes
PyTorch Lightning load_from_checkpoint patched to default to CPU for Modal snapshot compatibility

Caching Behavior¶

Response caching is not performed by the model container; callers may add a caching layer at the infrastructure level.

Versions & Changelog¶

Version	Date	Changes
v1	2024	Initial implementation with single, additive, and epistatic modes

Biology¶

Molecule Coverage¶

Primary Molecule Type(s)¶

ThermoMPNN-D is designed for globular proteins with known 3D structures (PDB format). It predicts the thermodynamic effect (ddG) of single and double amino acid substitutions on protein stability. The model handles single-chain proteins up to 1024 residues and requires backbone atom coordinates.

Cross-Applicability¶

Molecule Type	Applicability	Evidence	Caveats
Globular proteins	High	Primary training target	Best for well-folded soluble proteins
Enzymes	High	Well-represented in stability datasets	Useful for thermostable enzyme engineering
Antibodies	Moderate	Single-chain structures may work	Multi-chain requires specifying chain
Membrane proteins	Low	Under-represented in training data	Membrane stability context not modeled
Intrinsically disordered	Not applicable	Requires folded structure	No stable baseline state

Biological Problems Addressed¶

Epistatic Stability Prediction¶

In protein engineering, mutations rarely occur in isolation. When two mutations are introduced simultaneously, their combined effect may differ from the sum of their individual effects -- a phenomenon called epistasis. Epistasis can be:

Positive epistasis: The double mutant is more stable than predicted by summing individual effects
Negative epistasis: The double mutant is less stable than predicted
Sign epistasis: One mutation is stabilizing alone but destabilizing in combination with another

Understanding epistasis is critical for: - Combinatorial protein engineering: Identifying synergistic stabilizing mutation pairs - Directed evolution: Predicting which mutation combinations to test experimentally - Understanding evolutionary constraints: Why certain mutation combinations are favored or forbidden in nature

ThermoMPNN-D's epistatic mode directly models these pairwise interactions, going beyond simple additive approximations.

Single Mutation Stability Prediction¶

Like its predecessor ThermoMPNN, the model supports single mutation ddG prediction and full SSM scans. The single mode provides a baseline for comparison with the additive and epistatic double-mutation predictions.

Additive Double Mutation Prediction¶

The additive mode estimates double mutation ddG by summing the individual single-mutation effects. While this approximation misses epistatic interactions, it is computationally cheaper and provides a useful baseline. Comparing additive vs epistatic predictions reveals which mutation pairs exhibit significant non-additive effects.

Applied Use Cases¶

ThermoMPNN-D has been evaluated alongside related stability models in several benchmark studies. See sources.yaml for the catalogued applied literature entries, which include comparisons on double-mutation datasets (M1261) and large-scale site-saturation mutagenesis benchmarks.

Predecessor Models¶

ThermoMPNN-D extends ThermoMPNN (Dieckhaus et al., 2023), which itself builds on ProteinMPNN (Dauparas et al., 2022). The progression is:

ProteinMPNN: Protein sequence design from structure
ThermoMPNN: Single mutation stability prediction via transfer learning
ThermoMPNN-D: Single + double mutation stability with epistasis modeling

Complementary Models¶

ThermoMPNN: For single mutations only (simpler, faster). Use when double mutations are not needed.
ESM2: Sequence-based representations that can complement structure-based stability predictions.
ESMFold / Chai-1: Structure prediction models to generate input PDB structures when experimental structures are unavailable.

Alternative Models¶

Alternative	Advantage over ThermoMPNN-D	Disadvantage
ThermoMPNN	Simpler, lower memory (8 GB vs 12 GB)	No double mutation or epistasis support
TemBERTure	Sequence-only, no structure needed	No per-mutation ddG predictions

Biological Background¶

Protein stability engineering often requires introducing multiple mutations to achieve desired stability improvements. While individual stabilizing mutations can be identified through computational screening, their combined effects are not always additive.

Epistasis in protein stability arises from the complex network of intramolecular interactions. Two mutations may interact through:

Direct contact: Mutations at spatially proximal residues that physically interact
Allosteric coupling: Mutations at distant sites that communicate through the protein structure
Compensatory effects: One mutation disrupts a local interaction while another restores it in a different way

The CA-CA distance threshold used by ThermoMPNN-D (default 5.0 Angstroms) focuses predictions on spatially proximal mutation pairs, where direct epistatic interactions are most likely. This distance-based filtering also reduces the computational burden of scanning all possible double mutations, which scales as O(N^2) with sequence length.

Understanding and predicting epistasis is particularly important in: - Thermostable enzyme engineering: Where multiple mutations may be needed to reach process-temperature stability - Therapeutic protein development: Where stability, activity, and immunogenicity must be simultaneously optimized - Ancestral sequence reconstruction: Where epistatic constraints shape evolutionary trajectories

Sources & license¶

License: MIT (text)

Papers

Predicting the effect of single and multiple mutations on protein stability — bioRxiv preprint, 2024 · DOI

Source repositories

github: https://github.com/Kuhlman-Lab/ThermoMPNN-D

Cite

@article{dieckhaus2024thermompnnd,
  title={Predicting the effect of single and multiple mutations on protein stability},
  author={Dieckhaus, Henry and Kuhlman, Brian},
  journal={bioRxiv},
  year={2024},
  doi={10.1101/2024.10.10.617658}
}

ThermoMPNN-D¶

At a glance¶

API & schema¶

predict¶

Usage¶

Overview¶

Architecture¶

Capabilities & Limitations¶

Usage Examples¶

Architecture & training¶

Architecture¶

Model Type & Innovation¶

Parameters & Layers¶

Training Data¶

Loss Function & Objective¶

Tokenization / Input Processing¶

Performance & Benchmarks¶

Published Benchmarks¶

BioLM Verification Results¶

Comparison to Alternatives¶

Strengths & Limitations¶

Pros¶

Cons¶

Known Failure Modes¶

Implementation Details¶

Inference Pipeline¶

Memory & Compute Profile¶

Determinism & Reproducibility¶

Caching Behavior¶

Versions & Changelog¶

Biology¶

Molecule Coverage¶

Primary Molecule Type(s)¶

Cross-Applicability¶

Biological Problems Addressed¶

Epistatic Stability Prediction¶

Single Mutation Stability Prediction¶

Additive Double Mutation Prediction¶

Applied Use Cases¶

Related Models¶

Predecessor Models¶

Complementary Models¶

Alternative Models¶

Biological Background¶

Sources & license¶

`predict`¶