Protein Fitness Prediction

Build and compare protein fitness prediction models using multiple backbone architectures.

Table of contents

1. Overview
2. Required MCPs
3. Configuration
   1. Input Data Format
4. Pipeline Steps
   1. Step 0: Setup
   2. Step 1: Generate MSA
   3. Step 2: Build PLMC Model
   4. Step 3: EV+OneHot Model
   5. Step 4: ESM Models
   6. Step 5: ProtTrans Models
   7. Step 6: Compare and Visualize
5. Output Structure
6. Model Performance Reference
7. Run the Workflow

---

Overview

This workflow predicts the fitness landscape of protein variants using five ML approaches:

1. MSA generation via ColabFold/MMseqs2
2. PLMC evolutionary coupling model
3. EV+OneHot combined evolutionary + one-hot encoding
4. ESM protein language model embeddings (ESM2-650M, ESM2-3B)
5. ProtTrans transformer embeddings (ProtT5-XL, ProtAlbert)

All models are evaluated with 5-fold cross-validation and compared in a publication-ready four-panel figure.

Required MCPs

pskill install fitness_modeling

This installs: msa_mcp, plmc_mcp, ev_onehot_mcp, esm_mcp, prottrans_mcp

Configuration

PROTEIN_NAME: "MyProtein"
DATA_DIR: "examples/my_project"
WT_FASTA: "examples/my_project/wt.fasta"       # Wild-type sequence
DATA_CSV: "examples/my_project/data.csv"        # Must have 'seq' and 'log_fitness' columns
RESULTS_DIR: "results/fitness_modeling"

Input Data Format

Your data.csv must contain:

• seq — Full protein sequence
• log_fitness — Log-transformed fitness value (target)

Pipeline Steps

Step 0: Setup

Create results directory and copy input files.

Step 1: Generate MSA

Use msa_mcp to generate a multiple sequence alignment from the wild-type sequence via ColabFold/MMseqs2 server.

Step 2: Build PLMC Model

Convert MSA from A3M to A2M format, then build an evolutionary coupling model using plmc_mcp.

Step 3: EV+OneHot Model

Train a combined evolutionary coupling + one-hot encoding model using ev_onehot_mcp with 5-fold cross-validation.

Step 4: ESM Models

Extract ESM embeddings (ESM2-650M and ESM2-3B) and train regression heads (SVR, XGBoost, KNN) using esm_mcp.

Step 5: ProtTrans Models

Extract ProtTrans embeddings (ProtT5-XL, ProtAlbert) and train regression heads using prottrans_mcp.

Step 6: Compare and Visualize

Aggregate all model results and generate a four-panel figure:

1. Model Performance Comparison — Bar chart of best Spearman rho per backbone
2. Predicted vs Observed — Scatter plot for the best model
3. Head Model Comparison — Table of all backbone + head combinations
4. Execution Timeline — Gantt chart of step timings

Output Structure

RESULTS_DIR/
├── msa.a3m                         # Generated MSA
├── plmc/                           # PLMC model files
├── metrics_summary.csv             # EV+OneHot results
├── esm2_650M_*/                    # ESM model outputs
├── esm2_3B_*/                      # ESM-3B model outputs
├── ProtT5-XL_*/                    # ProtTrans model outputs
├── all_models_comparison.csv       # All model results combined
├── fitness_modeling_summary.png    # Four-panel summary figure
└── fitness_modeling_summary.pdf    # Publication-ready figure

Model Performance Reference

Model	Typical Spearman	Best Use Case
EV+OneHot	0.20–0.35	Baseline, interpretable
ESM2-650M	0.15–0.25	Fast, good balance
ESM2-3B	0.18–0.28	Higher accuracy
ProtT5-XL	0.15–0.25	Alternative to ESM

Run the Workflow

pskill install fitness_modeling
claude
> /fitness-model