ALSEBO

Active Learning Sequence Exploration via Bayesian Optimization

ALSEBO is a Python framework for navigating protein sequence space using a closed-loop active learning strategy. It combines a Variational Autoencoder (VAE) that generates a continuous latent landscape of sequences with Bayesian Optimisation (BO) to iteratively propose the most promising candidates for experimental testing.

Pipeline overview

ALSEBO closed-loop active learning pipeline

Closed-loop active learning pipeline: a VAE-derived latent landscape feeds sequence selection, evaluation, and Bayesian optimisation, which in turn proposes the next batch of sequences.

MSA
 │
 ▼
VAE  ──────────────────────────────────────────────────────────────┐
 │  generates a latent landscape & samples diverse sequences        │
 ▼                                                                  │
Sequence Space                                                      │
 │  featurized via DCA · ESM · latent coordinates                   │
 ▼                                                                  │
Initial Training Set                                                │
 │  diverse subset selected by t-SNE/PCA + k-means clustering       │
 ▼                                                                  │
Wet-lab / in silico evaluation                                      │
 │  measure objective(s): fitness, stability, activity …            │
 ▼                                                                  │
Gaussian Process Regression                                         │
 │  fits a surrogate model per objective                            │
 ▼                                                                  │
Acquisition Function (UCB)                                         │
 │  scores the unexplored sequence space                            │
 ▼                                                                  │
Next Batch  ────────────────────────────────────────────────────────┘
 top-k sequences recommended for the next experiment round

Key features

  • Multi-objective BO — weighted scalarisation of multiple fitness objectives with configurable direction (maximise / minimise).

  • Pluggable featurisation — swap between DCA (Direct Coupling Analysis), ESM protein language model embeddings, or raw VAE latent coordinates.

  • Diversity-aware initialisation — t-SNE or PCA projection followed by k-means ensures the first experimental batch covers the full sequence landscape.

  • Append-friendly data model — each experimental round appends to a single CSV, making it easy to resume or inspect the optimisation history.

Installation

git clone https://github.com/dulithaprasanna/ALSEBO.git
cd ALSEBO
pip install -e .

See the installation guide for full details including the DCA dependency.