Improving functional protein generation via foundation model-derived latent space likelihood optimization.

Area of Science:

• Computational biology
• Protein engineering
• Artificial intelligence in life sciences

Background:

• Deep generative models are used for *de novo* protein generation.
• Sequence-based methods are preferred due to data availability and lower complexity.
• Current models focus on exact amino acid sequence matching, which may be overly restrictive.

Purpose of the Study:

• To develop improved functional protein sequence generation models.
• To explore optimizing generative models beyond the amino acid sequence space.
• To leverage pre-trained protein language models (PLMs) as functional validators.

Main Methods:

• Proposed a multi-likelihood optimization strategy for training generative models.
• Simultaneously optimized training data likelihood in amino acid and latent spaces.
• Utilized pre-trained protein language models (PLMs) like ESM2 for latent space encoding.
• Applied the method to train GPT-like autoregressive transformers for antimicrobial peptide (AMP) and malate dehydrogenase (MDH) generation.

Main Results:

• The proposed method outperformed existing deep generative models.
• Demonstrated superior performance compared to standard GPT models and other generative approaches (GAN, VAE).
• Validated the effectiveness of the multi-likelihood optimization strategy for functional protein generation.

Conclusions:

• The multi-likelihood optimization strategy enhances functional protein sequence generation.
• Integrating latent space validation improves generative model performance.
• This approach offers a more effective way to design novel functional proteins.