Minimal-Data Peptide Design and Accessible Protein Language Models for Protein Engineering

Protein engineering is increasingly driven by machine learning, yet two practical barriers still limit broad adoption: (i) the scarcity of experimentally labelled data for new protein targets and (ii) the steep hardware demands of modern protein-language models (pLMs). This dissertation tackles both challenges through two core contributions. First, I introduce Minimal-Data → Maximal-Insight (MDMI), a structure-aware peptide-discovery pipeline that needs only a single round (~100 variants) of experimental screening. MDMI couples AlphaFold-Multimer complex prediction with hybrid statistical/physics scoring (SPServer + PyRosetta) to train a predictor, which then steers a genetic algorithm through sequence space to find novel sequences. As a case study, I applied MDMI to the split-GFP system, where the interaction between a 16-residue GFP11 peptide and its target GFP1-10 fragment reconstitutes fluorescence. MDMI successfully designed GFP11 peptides with over 50% sequence divergence from the wild type while preserving function, demonstrating the MDMI’s capacity to uncover non-obvious, high-diversity variants in data-limited scenarios. By decoupling peptide engineering from large training sets, MDMI offers an accessible strategy for laboratories with limited throughput.Second, I present Quantized Low-Rank Adaptation (QLoRA) for protein language models (pLMs), combining 4-bit weight quantization with low-rank adapters to enable efficient fine-tuning on lower cost and accessible GPUs. Across several pLMs (8 million–3 billion parameters) QLoRA cuts required training memory by an average 46.7%, and up to 90 % for the largest models, while retaining ≥ 90 % of baseline performance on regression tasks (i.e. fluorescence and stability datasets), secondary-structure classification, and de novo protein generation. Together, these methods establish a resource-conscious paradigm: MDMI extracts maximal design power from minimal data, and QLoRA delivers advanced pLMs to laboratories without specialized hardware. By uniting minimal-data modelling with hardware-efficient learning, this work broadens access to next-generation protein engineering and paves the way for rapid, distributed innovation in therapeutics, diagnostics, and biomanufacturing.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.