HAIPR: A High-Throughput Affinity Prediction Framework

Jannis de Riz, Tom U. Schlegel, Jens Meiler, Torben Schiffner

Keywords: binding-affinity prediction, deep mutational scanning, high-throughput screening

TL;DR: We provide a unified framework for end-to-end in-silico protein binding affinity maturation based on deep mutational scanning data.

Abstract: Accurate prediction of protein binding affinity is key for drug discovery and protein engineering, but commonly used evaluation protocols like Random Cross-Validation (RandomCV) can misrepresent true model generalization. We present HAIPR, a unified, open-source framework that streamlines the full machine learning pipeline for affinity prediction from training and optimization to inference, with curated benchmark datasets and robust, biologically meaningful evaluation protocols. By extending the BindingGYM benchmark and introducing realistic data splits, HAIPR reveals that RandomCV substantially overestimates model performance on out-of-distribution tasks. We systematically compare Support Vector Regression (SVR) using protein language model (pLM) embeddings to parameter-efficient fine-tuning (PEFT) of pLMs. SVR shows competitive results and increased stability in data-scarce scenarios, while PEFT excels as datasets grow larger and tasks become more complex. Analysis of model input setups shows that incorporating structural information does not always improve, and may sometimes hinder, performance for practical affinity prediction. Finally, we determine the lower limits of data required for reliable prediction, finding that even compact models can achieve performance close to the reproducibility limit of state-of-the-art assays, a practical ceiling for computational prediction. Code and pre-computed embeddings are publicly available.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 20560