LlamaAffinity: AI-Powered Leap in Predicting Antibody–Antigen Binding
- Zhandos Sembay
- May 28
- 2 min read
A Llama 3-Based Deep Learning Model Outperforms State-of-the-Art Antibody Binding Predictors
Birmingham, AL – In a major leap forward for immunoinformatics and therapeutic antibody design, researchers at the University of Alabama at Birmingham have developed LlamaAffinity, a novel antibody–antigen binding prediction model powered by the Llama 3 large language model (LLM). The model outperforms all current AI methods in the field—while requiring half the training time.
Published in the Proceedings of CIBB 2025 and posted on bioRxiv, this open-access study offers a new benchmark for therapeutic antibody screening, especially in oncology, infectious disease, and autoimmune research.
📖 Read the full preprint: https://doi.org/10.1101/2025.05.28.653051
Why LlamaAffinity Matters
Therapeutic antibodies are the cornerstone of modern biologics, yet predicting which will bind effectively to target antigens has been a costly and time-intensive process.
LlamaAffinity uses antibody sequences (heavy and light chains) from the Observed Antibody Space (OAS) and feeds them into a Llama 3-based model with 4 transformer layers and global pooling to predict binding affinity.
📈 Key results:
Accuracy: 96.40%
F1-Score: 96.43%
ROC AUC: 0.9936
Training Time: 0.46 hrs (faster than AntiFormer, AntiBERTa, and AntiBERTy)
Benchmark Showdown
In direct comparison, LlamaAffinity significantly outperformed other state-of-the-art tools in both speed and accuracy:
What’s Next?
With AI-driven antibody discovery attracting hundreds of millions in pharma investments, LlamaAffinity stands out as a generalizable, efficient, and open-source tool to speed up biologics pipelines.
📘 Citation: Hossain D, Saghapour E, Song K, Chen JY. LlamaAffinity: A Predictive Antibody–Antigen Binding Model Integrating Antibody Sequences with Llama3 Backbone Architecture. bioRxiv. https://doi.org/10.1101/2025.05.28.653051
Comments