Principal Genomic Parts for Durable Engraftment

When doctors transplant a microbial community into a patient — by fecal microbiota transplantation (FMT) or a probiotic cocktail — the introduced strains usually take hold briefly and then vanish, and the patient's native gut ecosystem snaps back like a stretched rubber band. In a five-year follow-up of 227 IBS patients, the clinical response rate after FMT peaked at 74.62% at three months and drifted down to 62.71% by sixty months (Long-Term Follow-Up Results of Fecal Microbiota Transplantation for Irritable Bowel Syndrome: A Single-Center, Retrospective Study), and in a blinded hypertension trial of 126 patients the blood-pressure benefit was gone by Day 30 (Fecal microbiota transplantation for hypertension: an exploratory, multicenter, randomized, blinded, placebo-controlled trial). Everyone agrees that some strains persist and most do not, but we cannot look at a strain's genome today and say which fate it will meet. Meanwhile, an unrelated literature has solved a structurally identical problem: computational morphology works out, for languages where one root balloons into hundreds of inflected forms, which small set of forms ("principal parts") lets you deduce all the others (Principal Parts Detection for Computational Morphology: Task, Models and Benchmark).

This proposal imports that machinery into microbiology. A bacterial pangenome has the same shape as a morphologically rich language: a conserved core genome plays the role of the lexical root, and variable accessory cassettes — operons, mobile elements, glycoside-hydrolase repertoires — play the role of affixes. The claim we will test is that engraftment durability is governed by a small, structured subset of accessory cassettes, that these "principal genomic parts" can be discovered by the same characterize–cluster–select algorithms used on inflectional paradigms, and that a representation built from them predicts strain survival half-life in the gut better than any bag-of-genes or sequence-similarity model.

Background and Why the Intersection Is Empty

Two ML systems already learn from raw microbial genomes, and both stop just short of this question. The set-aggregated genome embedding model (SAGE) represents a community as a hierarchy of sets — genes pooled into taxa, taxa pooled into communities — using pre-trained DNA language models, and predicts static relative abundance; it explicitly does not model longitudinal persistence or strain-level dynamics (Exploring Set-Aggregated Genome Embeddings for Microbiome Abundance Prediction). The whole genome transformer encodes each gene with the protein language model ESM-2 and classifies a genome's habitat as Host, Soil, or Aquatic, reaching 71% accuracy and 0.82 F1 on host-associated microbes (Whole Genome Transformer for Gene Interaction Effects in Microbiome Habitat Specificity) — but it randomly shuffles contig order during training, deliberately destroying any positional grammar, and it never evaluates persistence inside a host. On the linguistics side, the principal-parts benchmark reports 55.05% average F1 for its best model against a 21.20% random baseline across ten languages (Principal Parts Detection for Computational Morphology: Task, Models and Benchmark), and the SIGMORPHON 2021 shared task showed that grammar-based structure, not raw string similarity, drives paradigm discovery (Findings of the SIGMORPHON 2021 Shared Task on Unsupervised Morphological Paradigm Clustering). Nobody has connected the two: treating a pangenome as an inflectional system whose "paradigm cells" are accessory-gene configurations, and asking which cells determine survival in a host.

Two more results shape the design. ChronoStrain tracks strain abundances through time as a Bayesian state-space model with a Wiener-process (Gaussian random walk) prior on trajectories, so it can attach uncertainty to a strain's persistence estimate — but it takes strain identities as given reference genomes and does not predict why a strain survives (Strain tracking with uncertainty quantification). And Dehouck and Denis found that raw morphological productivity (how many forms a language has) does not predict whether morphology helps parsing; what matters is an information-theoretic quantity, Head POS Entropy, with a threshold at 0.65 below which explicit morphology yields gains up to 15.20 UAS points in Finnish (A Framework for Understanding the Role of Morphology in Universal Dependency Parsing). This is a warning: pangenome size or accessory-genome count is probably as uninformative about engraftment as form-count is about parsing, and we will need the analogue of their informativeness metric, not their productivity metric.

Core Claim

Engraftment durability is a function of a sparse, structured set of accessory cassettes. Concretely, if we represent each strain as (core genome → lemma, accessory cassettes → inflected slots), then (a) a minimal "principal genomic parts" subset exists from which the rest of the strain's colonization-relevant profile is predictable, (b) it is discoverable by adapted principal-parts algorithms, and (c) predictions from it beat bag-of-genes and k-mer baselines on longitudinal persistence data. The quantity we call engraftment entropy — the reduction in uncertainty about persistence gained by conditioning on accessory content:

$$E = H(P \mid \text{core}) - H(P \mid \text{core}, \text{accessory}),$$

where $P$ is persistence status at a fixed horizon and $H$ is entropy — plays the role Head POS Entropy plays for parsing: it tells us, per species and per host context, whether accessory structure matters at all before we invest in modeling it.

Research Program

The work has three components. First, we formalize the genomic morphology: from pangenome gene presence/absence matrices, define the core backbone as the "lemma" and accessory cassettes as slot-valued "affixes," then run the three-stage principal-parts pipeline — characterization (analogue of edit scripts: cassette-addition and cassette-loss operations between strains), clustering (hierarchical k-means on the characterization, as in the linguistic benchmark), and selection (minimal subset of cassettes whose configuration renders the rest predictable). Second, we build the durability label layer: run ChronoStrain-style Bayesian tracking on longitudinal metagenomic cohorts to extract, per donor or administered strain, a posterior survival half-life with credible intervals rather than a binary engrafted/not-engrafted label. Third, we join them: a model that conditions a survival-time prediction on the selected principal parts (and on nothing else in the accessory genome), trained jointly rather than as annotate-then-predict — the design lesson coming from LemmaTag, where a shared encoder resolving tagging and lemmatization together beat sequential pipelines on Czech (98.37% vs 97.86% lemmatization accuracy) and Arabic (96.08% vs 92.60%), precisely because errors in the first stage stopped propagating into the second (LemmaTag: Jointly Tagging and Lemmatizing for Morphologically Rich Languages with BRNNs).

Two design choices are motivated by constraining results in the file. The persistent "dysbiotic scar" left by meropenem, still unresolved 30 days after treatment (Absence of gut microbiota restoration following meropenem treatment in a systemic infection model), warns that durability depends on the resident community state too, so host context enters as a conditioning variable, not an afterthought. And the single-strain evidence that survival requires specific equipment — the B. adolescentis strain AF91-08b2A, whose stress-resistance genes (uspA, dnaK, groS) and glycoside-hydrolase repertoire preserved colon length at 7.15 cm versus 5.79 cm in untreated colitic mice (Genome analysis of Bifidobacterium adolescentis and investigation of its effects on inflammation and intestinal barrier function) — is our existence proof that a handful of cassettes, not the whole genome, carries the signal.

Proposed Experiments

The evaluation runs in three stages of increasing difficulty. Stage one is a representation sanity check: on the ProGenomes v3 habitat dataset (~30,000 genomes), we verify that slot-structured cassette representations do not lose — and ideally improve on — the 71% accuracy and 0.82 host F1 reported for the contig-shuffled transformer (Whole Genome Transformer for Gene Interaction Effects in Microbiome Habitat Specificity), under the same split protocol and a matched compute and tuning budget. Stage two is the core test: using the UTI Microbiome Project (longitudinal E. coli) and the Early Life Microbiota Colonisation study (longitudinal Enterococcus faecalis) cohorts that ChronoStrain was validated on, we extract survival half-lives for administered or colonizing strains and evaluate concordance-index and time-dependent AUROC for half-life prediction from principal parts. ChronoStrain outperformed StrainGST, StrainEst, and mGEMS on these cohorts, but no reported scalar RMSE or AUROC value was found on Lacuna, so our numbers are generated by re-running it ourselves as the label oracle. Stage three is out-of-sample clinical transfer: the 748-sample metagenomic series from the hypertension FMT trial, where donor-strain persistence at Day 7 versus Day 30 is measured against the published clinical anchors (−4.34 mmHg systolic difference at Day 7, $p=0.024$; gone by Day 30, $p=0.62$) (Fecal microbiota transplantation for hypertension). We test whether predicted-durable strains are the ones whose persistence correlates with the sustained-benefit subgroup (the >48-year-old group retained a 7.65 mmHg difference at Day 30). For the principal-parts discovery step itself, the calibration target is the structure of the linguistic benchmark: our best configuration should beat a random cassette-selection baseline by a margin comparable to the 55.05% versus 21.20% F1 gap reported there, evaluated by held-out reinflection — can the selected cassettes predict the rest of the colonization-relevant accessory profile?

Baselines

The naive reference point is "predict every administered strain washes out by Day 30," whose prevalence is high given the clinical record of transient benefit; any model must beat its AUROC on the persistence labels before being taken seriously, and this baseline costs nothing to run. The strongest existing-method competitor is the whole-genome transformer recipe: ESM-2 gene embeddings, ~4,000 gene context, transformer encoder (Whole Genome Transformer for Gene Interaction Effects in Microbiome Habitat Specificity) — not for its original habitat task (where it reports 71% accuracy, 0.82 F1 on host microbes) but retrained, with our survival-half-life labels replacing the habitat labels. It is the strongest competitor because it is the only prior architecture that consumes whole genomes and produces host-relevant predictions; we run it on identical strain sets, identical label horizons, and a matched tuning budget (same sweep over learning rate, depth, and context length), and beating it means a higher concordance index on half-life with non-overlapping bootstrap confidence intervals, plus better Day-30 persistence AUROC in the hypertension cohort. A third comparator is a two-step classical pipeline — Prodigal-style gene calls plus k-mer MinHash distances into a gradient-boosted classifier — which represents everything the field would do without representation learning. The critical ablation isolates our one ingredient: the same joint encoder and the same survival head, but with the structured cassette-slot representation replaced by permutation-invariant pooling of gene embeddings, SAGE-style (Exploring Set-Aggregated Genome Embeddings for Microbiome Abundance Prediction). If slot structure and principal-parts selection carry the signal, this ablation should fall back toward the pooled baseline; if it does not, the morphological framing is decoration and we will say so.

Why not just…?

Why not just use k-mer or substring similarity between strains? This is the cheapest serious objection, and the linguistic evidence is directly against it at the structural level: in the SIGMORPHON 2021 shared task, grammar-based paradigm models beat string-similarity baselines by at least 23.06% F1 across test languages (Findings of the SIGMORPHON 2021 Shared Task on Unsupervised Morphological Paradigm Clustering), and in the principal-parts benchmark, raw edit distance underperformed transformational edit scripts because surface similarity misses functional structure. We still run the k-mer baseline honestly rather than argue from analogy, because genomes could differ from words; if MinHash-plus-boosted-trees matches principal-parts performance on the UTI and infant cohorts, the structural claim is dead and the simpler method wins.

Why not just scale a pretrained genome or protein language model with standard pooling? The whole genome transformer already tried the shuffled-order version of this and plateaued at 71% on a three-way habitat task it was designed for, struggling to separate even soil from aquatic (Whole Genome Transformer for Gene Interaction Effects in Microbiome Habitat Specificity), and SAGE's own authors found unstructured aggregation misses higher-order relationships unless the hierarchy is built in (Exploring Set-Aggregated Genome Embeddings for Microbiome Abundance Prediction). The morphology literature diagnoses why: frequency-based subword segmentation slices across structural boundaries and hides shared roots, the failure Park et al. documented for BPE in morphologically rich languages. Still, this is the real threat — ESM-2 embeddings may implicitly encode cassette structure. Our pooling ablation is designed to settle exactly this, on equal data and budget, rather than by argument.

Why not just fix durability with dosing protocols and skip genomics? Clinically, repeated dosing genuinely works sometimes — the capsule subgroup in the IBS cohort held an 88.46% response at 36 months (Long-Term Follow-Up Results of FMT for Irritable Bowel Syndrome) — so this objection has real force, and we concede that dosing and strain selection are complementary, not rival, levers. But dosing is not sufficient on its own: identical capsule schedules in the hypertension trial (20 capsules at Days 1, 7, 14) produced Day-7 benefit that vanished by Day 30 for most patients, with persistence of specific taxa such as Parabacteroides merdae tracking sustained benefit, and 60 days of continuous probiotic administration in mice was fully reversed on cessation because host-genetic selection dominated the introduced strain. The within-trial contrast is our experiment: under a fixed dosing protocol, strain-level genomic features must explain who persists and who does not, or the dosing-only story wins.

Decisive Checks

Three outcomes would decide the proposal. First, the pooling ablation: if permutation-invariant pooling matches slot-structured principal parts on concordance index across both longitudinal cohorts, the morphological framing adds nothing and should be abandoned. Second, the informativeness test: if engraftment entropy $E$ is near zero for most species — accessory content tells us nothing about persistence beyond the core — then durability is a host-state phenomenon, not a strain-genome phenomenon, and the program should pivot to community-state predictors. Third, transfer: if principal parts selected on mouse and UTI cohorts fail to rank strains in the independent human FMT cohort, the discovered cassettes are cohort-specific artifacts rather than general colonization machinery. Any one of these failures is informative; all three passing would give the field its first principled answer to "which strains will stay."