Use BioPhi Online
Commercially Available BioPhi No-Code Web Server
Try BioPhi
BioPhi: Open-Source AI for Antibody Humanization and Humanness Evaluation
BioPhi is an open-source platform designed to automate and scale antibody humanization while providing granular, interpretable humanness assessments. Developed by researchers at Merck & Co., BioPhi leverages deep learning and the vast diversity of the Observed Antibody Space (OAS)—a database of over 500 million natural human antibody sequences.
By moving beyond traditional manual methods, BioPhi enables therapeutic discovery campaigns to produce highly human-like antibodies at scale, reducing immunogenicity risks without sacrificing binding affinity.
Key Innovations: Repertoire-Scale Intelligence
BioPhi features two novel primary methods that utilize natural immune repertoires for superior antibody engineering:
Sapiens (Deep Learning Humanization): A Transformer-based masked language model trained on millions of OAS sequences. It "recognizes and repairs" non-human residues by predicting the most probable human residues in their place, achieving results comparable to human experts.
OASis (Interpretable Humanness Scoring): A granular evaluation method based on exact 9-mer peptide searches in the OAS database. It calculates the prevalence of every overlapping peptide across the human population, providing a "humanness report" that pinpoints specific high-risk regions.
Attention-Guided Design: Sapiens uses attention mechanisms to capture long-range structural and evolutionary dependencies between antibody loops, ensuring generated sequences maintain conformational stability.
Interactive Designer Interface: A user-friendly web dashboard that allows for manual sequence adjustments guided by real-time Sapiens scores, positional frequencies, and germline identity.
Performance Benchmarks
BioPhi outperforms traditional homology-based metrics and matches the performance of human experts in standard humanization tasks.
Task | Metric | BioPhi Result | Key Finding |
Humanness Classification | ROC AUC | 0.966 | Outperforms T20 (0.896) and Z-score (0.837) |
Immunogenicity Correlation | Explained Variance (R^2) | 0.28 | Stronger correlation with clinical ADA responses than Z-score |
Humanization Quality | Shared Mutations | ~73–77% | High overlap with mutations made by expert human designers |
Processing Throughput | Speed | 1,000 mAbs / 2.3 min | Enables bulk humanization of entire discovery libraries |
Scientific Breakthroughs in Antibody Engineering
Scalable Humanization of Murine Leads
In a benchmark of 177 antibodies, BioPhi's Sapiens produced humanized sequences that achieved the same humanness as experimental clinical sequences while maintaining superior preservation of the parental variable region. This high preservation is critical for retaining the original binding affinity of the model animal source.
Interpretable Risk Mitigation
Unlike "black-box" AI models, OASis provides a granular breakdown of an antibody’s human-likeness. By identifying specific 9-mers that are rare in the human population, researchers can target individual mutations to lower the risk of anti-drug antibody (ADA) responses in patients.
Discovery of Putative Parents
The BioPhi framework can reconstruct putative parental sequences from known humanized therapeutics with over 90% identity. This capability allows researchers to analyze and learn from thousands of successful clinical humanization campaigns.
BioPhi on Tamarind Bio: Professional Antibody Design
Tamarind Bio provides a managed, high-performance environment for the BioPhi suite, abstracting away the complexities of large database management (e.g., the 22GB OASis database) and GPU configuration.
Managed OASis Database: Access the full power of repertoire-based scoring without local infrastructure requirements.
Automated Batch Humanization: Rapidly process hundreds of discovery-stage sequences in minutes through a streamlined web interface.
How to Use BioPhi on Tamarind Bio
Access the Platform: Log in to tamarind.bio and select the BioPhi tool.
Upload Sequences: Provide your non-human (e.g., murine or rabbit) variable region sequences in FASTA format.
Run Humanness Report: Evaluate your initial leads using OASis to identify potential immunogenicity hotspots.
Select Humanization Mode: Choose between Sapiens (deep learning) or traditional CDR Grafting.
Configure CDR & Vernier Zones: Define your preferred CDR numbering scheme (Kabat, IMGT, Chothia, or North) and set back-mutation rules for Vernier positions.
Refine in Designer: Use the interactive interface to manually adjust mutations while monitoring humanness scores and sequence properties in real time.
Export Results: Download high-confidence humanized designs for synthesis and wet-lab validation.