BioPhi

BioPhi webserver overview

BioPhi 1.0.11 humanizes antibody variable domains with Sapiens, evaluates sequence humanness with OASis, performs CDR grafting, and applies explicit Designer mutations. Results include a structured table and the native FASTA, CSV, XLSX, and alignment files produced by the selected mode.

BioPhi is the platform described by Prihoda et al. (2022). It does not predict antigen binding, expression, aggregation, viscosity, specificity, or clinical anti-drug antibody incidence, so its output is a starting point for the checks listed under Validation and follow-up.

Choosing a mode

Both is the usual first run. It returns the Sapiens sequence together with OASis values for the parental and final chains, so every humanized sequence arrives with its reference. OASis mode suits characterization, because it leaves the sequence untouched and returns only the humanness and germline report.

Use Sapiens (Humanization) when only the final sequence is needed. Sapiens positional scores shows the model support for all 20 amino acids at each position, which helps when comparing alternatives at one site, and Sapiens mean score condenses the same evidence into one value per chain. Neither score has a clinical cutoff.

Choose CDR grafting to place the parental CDRs on a human V germline, and Designer mutations when the edits are already decided from a score matrix, a germline comparison, structural review, or experimental data. In a first Sapiens comparison, change one design pressure at a time: keep Humanization iterations at 1 and Humanize CDRs off, then raise one of them in a separate job. Each extra iteration applies Sapiens again to the previous result, so it deepens the humanization instead of producing an independent alternative.

Pricing

BioPhi uses a 10-credit base charge plus 10 credits per FASTA record. A standard paired VH/VL run therefore costs 30 credits:

Input	FASTA records	Credits	Calculation
One VH/VL antibody	2	30	10 + (2 x 10)
One unpaired chain	1	20	10 + (1 x 10)

The exact quote is shown before submission. The examples above apply to FASTA input; review the displayed quote for PDB or RCSB inputs, which BioPhi parses into recognized antibody chains during the run.

Inputs

Input mode	Accepted data	Limits	Notes
Fast Sapiens modes	Separate protein FASTA inputs for Heavy chain (VH) and Light chain (VL)	Up to one sequence in each field; at least one chain is required; 10 MiB per upload	Used by Sapiens, Both, positional scores, and mean score. The Sapiens heavy-chain model scores variable domains up to 142 residues and the light-chain model up to 126. Residues after the numbered domain, such as a constant region, are not scored or returned.
Parser-backed modes	Protein or coding-DNA FASTA (.fasta, .fa, .faa, .fna), PDB, or RCSB PDB fetch	One submitted input or file; up to 10 MiB	Used by OASis, complete Sapiens report, CDR grafting, and Designer. A FASTA can contain one chain or one paired VH/VL antibody. BioPhi parses the source-native input.

Use complete antibody variable domains. The default interface accepts one VH, one VL, or a VH/VL pair. When both are present, ProteinIQ assigns matching BioPhi identifiers automatically, so the submitted FASTA headers do not need to match. Parser-backed modes retain one source-native input field. If that file contains both chains, use the same base identifier with _VH and _VL, or _HC and _LC, suffixes.

Fast protein modes accept the 20 standard amino acids, X, and a terminal *. They reject alignment gaps, DNA-like input, unrecognized variable domains, and sequences beyond the Sapiens model capacity.

Settings

Core setting

Parameter	Type	Default	Description
Processing mode	enum	Both (Humanization + Scoring)	Selects one of the eight workflows listed below.

Mode	Accepted input	Result
Sapiens (Humanization)	Protein FASTA	Final humanized sequence for each recognized chain
OASis (Humanness Scoring)	Protein or coding-DNA FASTA, PDB	Humanness and germline report without sequence modification
Both (Humanization + Scoring)	Protein FASTA	Sapiens sequence plus parental and final OASis metrics
Sapiens positional scores	Protein FASTA	BioPhi score matrix for all 20 amino acids at each position
Sapiens mean score	Protein FASTA	One mean Sapiens score per chain
Sapiens complete report	Protein or coding-DNA FASTA, PDB	Native Sapiens alignment, FASTA, and workbook
CDR grafting	Protein or coding-DNA FASTA, PDB	CDRs grafted onto selected human V germlines
Designer mutations	One protein or coding-DNA FASTA, or one PDB input	Explicit chain-numbered substitutions

Advanced settings

Parameter	Type	Default	Description
Humanization iterations	integer, 1 or greater	1	Number of successive Sapiens passes. Available in Sapiens, Both, and complete-report modes.
Numbering scheme	enum: Kabat, Chothia, IMGT, AHo	Kabat	Antibody numbering scheme passed to BioPhi.
CDR definition	enum: Kabat, Chothia, IMGT, North	Kabat	Defines the CDR boundaries used by BioPhi.
Humanize CDRs	boolean	Off	Allows Sapiens to modify CDR residues. Available in Sapiens, Both, and complete-report modes.
VHH / nanobody input	boolean	Off	Declares a heavy-chain-only single-domain input and adds a VHH-specific interpretation warning. It does not change BioPhi's scientific model. With Kabat or Chothia numbering, Both mode also names changed hallmark positions H37, H44, H45, and H47.
Prevalence threshold	enum: Loose, Relaxed, Medium, Strict, Custom	Relaxed	Requires an OASis 9-mer to occur in at least 1%, 10%, 50%, or 90% of human subjects, or in the percentage set by Custom prevalence. Available in OASis, Both, CDR grafting, and Designer modes.
Custom prevalence (% of subjects)	number, 1 to 90	10	The subject prevalence used when the threshold is Custom. BioPhi accepts any value from 1% to 90%.
Heavy-chain V germline	string	auto	Human heavy V family or gene for CDR grafting, such as IGHV3 or IGHV3-23. auto selects BioPhi's closest germline.
Light-chain V germline	string	auto	Human kappa or lambda V family or gene for CDR grafting, such as IGKV1 or IGLV2. auto selects BioPhi's closest germline.
Backmutate Vernier residues	boolean	On	Preserves parental Vernier residues during CDR grafting.
Run final Sapiens pass	boolean	Off	Applies one Sapiens pass after CDR grafting.
Designer mutations	text	empty, required in Designer mode	Chain-numbered substitutions such as H35:Y, L46:W, or H100A:F, separated by commas or new lines.

Outputs

Output	Format	Modes	Contents
Chain results	Interactive	Both, CDR grafting, Designer	Chain-specific OASis summary, parental and humanized sequences, and BioPhi's residue-level evidence table
Structured data	Spreadsheet	All	Mode-specific chain rows, identifiers, scores, mutations, germline annotations, sequences, and lengths
Humanized sequences	FASTA	Sapiens, Both, complete report, CDR grafting, Designer	Final recognized variable-domain sequences
OASis workbook	XLSX	OASis, Both	Native chain, peptide, humanness, and germline report
Parental OASis workbook	XLSX	Both	OASis report for the submitted sequences before humanization
Positional scores	CSV	Sapiens positional scores	Per-position scores for all 20 amino acids
Mean scores	CSV	Sapiens mean score	One mean Sapiens score per input chain
Humanization alignments	TXT	Complete report, CDR grafting, Designer	Numbered parental and final sequence alignments
Sapiens workbook	XLSX	Complete report	BioPhi's complete native Sapiens report
Humanization workbook	XLSX	CDR grafting, Designer	BioPhi's native CDR-grafting or mutation report
Normalized input	FASTA	All	Protein variable domains used for result interpretation

The Files view also includes results.csv and every native file returned by the selected mode. Files remain downloadable and supported native outputs can be routed into workflows.

Result fields

Field	Meaning
Sequence ID and Chain type	Recognized input identifier and VH or VL class
Parental sequence retained (%)	Percentage of positions unchanged between the parental and final chain
OASis identity (%)	Percentage of evaluated 9-mers that meet the selected subject-prevalence threshold
OASis percentile (%)	Position of the OASis identity within BioPhi's therapeutic-antibody reference distribution
Parental OASis identity (%) and Parental OASis percentile (%)	The corresponding values before humanization
Mutations and Mutation details	Count and substitutions between parental and final chains; chain numbering such as H35 in Both, CDR grafting, and Designer, residue position in Sapiens and complete-report modes
V germline, J germline, and Germline %	BioPhi germline assignments and germline content
Non-human peptides	OASis 9-mers that do not meet the selected prevalence threshold
Humanized sequence and Length	Final amino-acid sequence and residue count

OASis values are repertoire-based humanness evidence. They are not clinical immunogenicity predictions or universal pass/fail thresholds.

A heavy-chain-only job is treated as an unpaired VH unless VHH / nanobody input is enabled. BioPhi is not VHH-specific, so declared VHH results include a warning to review framework mutations and validate single-domain solubility and function independently.

Reading OASis results

OASis cuts each variable domain into every overlapping 9-mer and counts a peptide as human when it occurs in at least the selected percentage of human subjects in the Observed Antibody Space. OASis identity (%) is the share of 9-mers that pass, so it only has meaning at the threshold that produced it. Identities from Loose, Relaxed, Medium, Strict, or a custom prevalence cannot be compared with each other. Keep the threshold fixed across a project, and record it with each result.

OASis percentile (%) places the identity within a reference set of therapeutic antibodies. In the paper, that set is 544 therapeutic antibodies from IMGT/mAb-DB, and the median percentile was 5 to 7% for murine, 37 to 40% for humanized, and 80 to 81% for human antibodies. Percentiles are therefore more stable across thresholds than identities, but they rank a sequence against past therapeutics. A high percentile is not a pass mark.

The overall value hides where the problem sits. Use the residue table or the OASis workbook to find the Non-human peptides, then check which mutations fall in those windows. Because the windows overlap, one substitution changes up to nine peptides. When comparing parental and final chains, read Mutation details alongside the identity change: a few substitutions in a Vernier position, the VH/VL interface, or a buried core can cost more than many conservative surface changes. Germline assignments and Germline % describe similarity to human segments, not folding, affinity, or immunogenicity.

Preparing inputs

Submit complete variable domains rather than whole heavy and light chains. Antibody numbering decides which residues are scored, so truncated domains and non-antibody sequences can fail recognition. Record the Numbering scheme and CDR definition with every design, because a residue can move between framework and CDR depending on the definition, which changes whether Sapiens may edit it and how a numbered Designer mutation is read.

In parser-backed modes, a FASTA with both chains needs matching base names such as candidate_01_VH and candidate_01_VL; otherwise BioPhi reports an unpaired chain. A heavy-chain-only nanobody should be submitted with VHH / nanobody input enabled. Coding DNA and PDB files work only in the parser-backed modes.

An unpaired-chain warning does not mean the score is missing. It means BioPhi processed a valid VH or VL without a recognized partner. Keep the submitted FASTA, alignments.txt when present, and the normalized input with the result files, so each mutation can be traced to a numbered position later.

Validation and follow-up

In the paper, Sapiens was compared with expert humanization on a benchmark of 177 antibodies and gave comparable results. OASis medium identity separated human from non-human therapeutic sequences with a ROC AUC of 0.966, and correlated with reported anti-drug antibody incidence across 217 therapeutics (R = -0.53, R² = 0.28). Both results are retrospective and population-level; they do not show that a given identity predicts one patient's immune response.

Predict the paired Fv of the parental and humanized sequences with ABodyBuilder3 and inspect the CDRs, Vernier positions, and VH/VL interface. TAP2 adds developability context for the modeled Fv. Binding, kinetics, and function still need measurement for the parental and designed variants, ideally across a small panel of conservative and deeper Sapiens variants and reviewed grafts. For germline and rearrangement annotation without design, use IgBLAST.

Sources

• Prihoda D, Maamary J, Waight A, Juan V, Fayadat-Dilman L, Svozil D, Bitton DA. "BioPhi: A platform for antibody design, humanization, and humanness evaluation based on natural antibody repertoires and deep learning." mAbs 14, 2020203 (2022). doi:10.1080/19420862.2021.2020203