Anna-Maria Liebhoff, Thiagarajan Venkataraman, William R Morgenlander, Miso Na, Tomasz Kula, Kathleen Waugh, Charles Morrison, Marian Rewers, Randy Longman, June Round, Stephen Elledge, Ingo Ruczinski, Ben Langmead, H Benjamin Larman
9 min
Phage Immunoprecipitation Sequencing (PhIP-Seq) is a powerful tool for profiling antibody responses across massive antigenic spaces, such as the human gut virome. However, as proteomes grow larger, traditional tiling methods—which synthesize overlapping peptides across the entire length of proteins—become prohibitively expensive and inefficient. Researchers in this study developed Dolphyn, a computational strategy designed to compress peptide libraries while maintaining high sensitivity for antibody detection.
Dolphyn works in two primary stages. First, it employs a Random Forest (RF) classifier to identify and prioritize 15-amino-acid sequences (15-mers) that are highly likely to function as antibody epitopes based on their amino acid composition. The model was trained on public epitope data, revealing that amino acid side-chain properties are crucial for distinguishing reactive regions. Second, Dolphyn utilizes an epitope-stitching technique, which combines up to three of these high-probability 15-mer epitopes into a single 56-amino-acid peptide construct. This effectively compacts the library by focusing resources on the most likely immunogenic regions.
The researchers validated Dolphyn by applying it to the Gut Phage Database, finding that it could represent the phage proteome with 78% fewer peptides than standard tiling methods. Despite this significant reduction in library size, Dolphyn-designed libraries yielded three times more reactive peptides in a cohort of 51 healthy individuals. Notably, the study confirmed a prevalent immune response in healthy populations to E.coli-infecting Myoviridae phages, successfully recapitulating findings from larger, traditional Pepsyn-designed libraries with far greater efficiency.
This method addresses a major barrier in high-throughput immunology. By making it computationally and financially tractable to screen massive, poorly characterized proteomes, Dolphyn facilitates the exploration of the gut-immune axis and other complex host-microbiome interactions. Its modular design allows researchers to substitute the prediction model, making it a flexible tool for various future applications in antibody profiling.
We investigate a relatively underexplored component of the gut-immune axis by profiling the antibody response to gut phages using Phage Immunoprecipitation Sequencing (PhIP-Seq). To cover large antigenic spaces, we develop Dolphyn, a method that uses machine learning to select peptides from protein sets and compresses the proteome through epitope-stitching. Dolphyn compresses the size of a peptide library by 78% compared to traditional tiling, increasing the antibody-reactive peptides from 10% to 31%. We find that the immune system develops antibodies to human gut bacteria-infecting viruses, particularly E.coli-infecting Myoviridae. Cost-effective PhIP-Seq libraries designed with Dolphyn enable the assessment of a wider range of proteins in a single experiment, thus facilitating the study of the gut-immune axis.
Sam: They started with viral peptides already known to attract antibodies frequently. In one experiment, they substituted small stretches of amino acids to see where binding broke down. In another, they tested shorter fragments of the original peptide.
Alex: Those experiments address different things: where recognition depends on the sequence, and how much sequence you need to retain. What did they suggest?
Sam: Most public linear epitopes could be captured in about fifteen amino acids. Linear means recognition of a stretch of sequence, rather than distant regions brought together by folding. The authors used fragments of that length to train their predictor.
Alex: What does the predictor learn? Is it recognizing particular viral sequences, or something more transferable?
Sam: It's a random forest, a classifier combining many decision trees. Its inputs describe amino acid frequencies, pairs, and chemical groupings. Charged side chains and other composition features help separate reactive fragments from nonreactive ones.
Alex: Related fragments could make that look stronger than it really is. How did they check performance on sequences outside the training examples?
Sam: They kept fragments from the same original peptide out of both sides of the training-test split. They also tested previously screened Epstein-Barr virus peptides absent from training, and separately screened selected enterovirus fragments. Those tests showed moderate discrimination, not near-perfect prediction.
Alex: So this is a ranking aid, not a reliable declaration that every selected fragment is an antibody target. How does a modest predictor produce such a large efficiency gain?
Sam: Selection is only one component. Dolphyn also stitches three promising regions from the same protein into one displayed peptide, separated by short linkers. In the pilot design, the highest-scoring region comes first, followed by the next-ranked regions.
Alex: That buys more candidate targets per synthesized fragment. But joining regions changes their surroundings. Did they test whether stitching interferes with antibody detection?
Sam: The pilot included individual predicted regions and their stitched counterparts. Usually, only one of the individual regions was reactive. Binding signals for the stitched version were similar to those for the reactive individual region, supporting effective representation in the examples tested.
Alex: And the readout becomes protein-level recognition. A stitched fragment can tell you that antibodies recognize something from that protein, without necessarily identifying which joined region they bind.
Sam: That's the intended use. The authors then compared the designs using the same collection of prevalent gut phages from North American individuals. The results report screening plasma from fifty-one healthy individuals, rather than testing only computational predictions.
Alex: Give me the central experimental result. How much more useful was each fragment in the compressed library?
Sam: In the uniform library, about ten percent of peptides were bound by antibodies in at least one screened sample. In the Dolphyn library, that rose to thirty-one percent. That's library-wide reactivity across the cohort, not the fraction bound by every individual.
Alex: More hits per fragment is useful, but it could still erase the biological pattern. What happened when they compared antibody responses to the gut phages themselves?
Sam: Both designs identified stronger responses to phages associated with E. coli, especially those annotated as Myoviridae, a phage group. Most individuals showed reactivity to that group. The compressed library also preserved the broad clustering of phage reactivity seen with the uniform design.
Alex: These viruses infect bacteria, not the people providing plasma. Does antibody binding tell us that the immune response changes the gut ecosystem?
Sam: No. This study finds antibody recognition, not its consequences. The authors explicitly leave open whether these antibodies neutralize phages, enhance anything, or associate with health-related traits. Some host and phage classifications also depend on predictions and sequence alignments, rather than direct infection experiments.
Alex: Let's return to what compression loses. If I want a complete inventory of antibody targets, how close does Dolphyn get to the uniform screen?
Sam: It recovers only a little over half of the proteins reactive in Pepsyn. Some proteins are absent because Dolphyn requires enough predicted epitopes to assemble a stitched peptide. Smaller proteins can therefore disappear from the library entirely.
Alex: That makes the trade-off concrete. The broad pattern survives, but an individual protein I care about could be missing before the experiment starts.
Sam: Dolphyn also detects some protein targets that Pepsyn misses, so the uniform library isn't a perfect reference. Still, neither result supports comprehensive equivalence. My reading is that this is convincing evidence for efficient pattern discovery, not for replacing exhaustive coverage whenever target completeness matters.
Alex: How far should researchers carry that conclusion beyond this particular gut phage collection?
Sam: The training focused on public viral epitopes, and the pilot focused on locally prevalent gut phages. Performance elsewhere remains something to test. Dolphyn is modular, so researchers can replace the predictor, training data, or other design settings, but flexibility isn't evidence of generalization.
Alex: Who should read the full document, and where should they start to decide whether the compromise fits their work?
Sam: Researchers designing large antibody libraries should start with the Dolphyn library-design section, then the stitching sensitivity comparison. Read the protein-target recovery comparison before treating the compression as a substitute for uniform coverage. Gut microbiome researchers should also examine the healthy-individual phage responses and the discussion of unknown antibody function.
Alex: And what should everyone else remember when a smaller screen claims to preserve discovery power?
Sam: Ask what kind of discovery it preserves. Here, selecting and combining likely targets retains key immune patterns with far fewer peptides. But a smaller screen can preserve the pattern without preserving every target.
Alex: Thanks for listening.