Cell-Line PICRUSt: Reconstructing Unmeasured Pathway States and Drug Responses on a Cancer Lineage Tree

When microbiologists sequence a mixed bacterial community, they usually read only a cheap, short "barcode" gene rather than full genomes. PICRUSt — a method whose name means "phylogenetic investigation of communities by reconstruction of unobserved states" — exploits the fact that functional traits evolve slowly: an unsequenced microbe is placed on an evolutionary tree of sequenced relatives, statistical models of trait evolution reconstruct what gene families its ancestors carried, and from that the microbe's likely metabolic capabilities are inferred without ever sequencing its genome. Cancer pharmacogenomics has, in mirror image, the opposite problem: it has rich molecular profiles for the cell lines it has already screened, but models trained on those lines collapse when moved to new ones, because they latch onto tissue-of-origin and culture quirks instead of mechanisms of drug response. This proposal applies PICRUSt's three-step logic — build a reference tree, reconstruct unobserved states along it, and normalize for lineage-specific "copy number" artifacts — to cancer cell lines, treating the cell line panel as an evolving clade whose unmeasured functional states and drug sensitivities can be reconstructed from where a line sits on a lineage tree.

The immediate motivation is quantified in the literature. The pharmacogenomics guidelines paper reports that even after best-practice fixes, a Random Forest drug-sensitivity model's correlation with held-out truth falls from roughly 0.41 within one screening panel to roughly 0.33 when trained on one panel and tested on another (Drug Sensitivity Prediction From Cell Line-Based Pharmacogenomics Data: Guidelines for Developing Machine Learning Models). Meanwhile, the one method that pools related lineages — CHER, which shares information across cell lines of common developmental origin — beats an Elastic Net in 60% of drug–phenotype pairs with a mean Pearson improvement of 0.24, but is restricted to a single binary split between two lineage groups (Context Sensitive Modeling of Cancer Drug Sensitivity). No one has built a full lineage tree of cell lines and run ancestral state reconstruction on it. That is the gap this proposal fills.

Core Claim

The claim is that a cancer cell line panel can be organized into a lineage tree — using developmental lineage as topology and multi-omic similarity as branch lengths — and that PICRUSt-style machinery on this tree does two things static methods cannot: it reconstructs pathway-level functional states for cell lines whose molecular profiles are missing or sparse, and it replaces binary pan-cancer-versus-lineage pooling with continuous, distance-weighted transfer of drug response patterns along branches. A third ingredient is a direct transplant of PICRUSt's 16S copy-number correction: PICRUSt divides each microbe's abundance by its predicted number of barcode-gene copies before aggregating function, and we analogously divide each cell line's expression by its local aneuploid copy number before computing pathway activity, so that pathway scores reflect functional state rather than genomic amplification. The testable prediction is that tree-structured reconstruction shrinks the cross-panel generalization gap and enables prediction for sparsely profiled lines, at a cost the flat and static baselines cannot reach.

Research Program

The first component is constructing the reference tree. PICRUSt's tree is built from the 16S barcode of sequenced organisms; our analog is a cell-line tree in which topology comes from the Cell Ontology's hierarchical cell-type organization — the same resource POPPY uses to compute distances between cell types in clinical bulk tumors (Cross-modal transfer learning for mapping bulk transcriptomes at cellular level) — constrained by developmental-origin relationships of the kind CHER identified (e.g., melanoma and glioma sharing ectodermal origin, breast and ovarian lines sharing genomic similarity), with branch lengths estimated from CCLE copy-number, expression, and methylation distances. This is not a true phylogeny: cell lines are not vertically inherited organisms. The tree is a structured similarity hierarchy, and ancestral state reconstruction on it is mathematically a Gaussian smoothing process along its branches. We state this plainly, because it sharpens what the tree must earn: it must beat unstructured similarity (nearest-neighbor imputation on expression correlation) to justify its existence.

The second component is trait reconstruction. Following PICRUSt's second stage, where statistical models of evolution estimate gene-family copy numbers on unsequenced branches, we treat each of the 1,329 GSVA pathway activity scores used by Precily as a quantitative trait (Gene expression based inference of cancer drug sensitivity), fit a Brownian-motion model of trait change along the cell-line tree, and use it to reconstruct pathway states at unmeasured tips and internal nodes. These reconstructed states serve two roles: as imputed features for lines with missing transcriptomes, and as smoothed, lineage-aware pathway inputs to drug response models.

The third component is copy-number-corrected pathway aggregation. PICRUSt's third stage normalizes for 16S copy-number variation across taxa before scaling function; the analog is to divide each gene's expression in a cell line by its aneuploid copy number in that line before computing pathway scores, removing the amplification artifact that makes heavily aneuploid lineages look functionally hyperactive. The guidelines paper shows that failing to correct basal lineage differences makes models predict lineage rather than mechanism (Drug Sensitivity Prediction From Cell Line-Based Pharmacogenomics Data: Guidelines for Developing Machine Learning Models); copy-number normalization is an evolutionary correction aimed at exactly this artifact.

The fourth component is phylogenetic partial pooling for drug response. CHER uses L0 regularization and split variables to decide, per feature, whether a predictor is shared pan-cancer or specific to one of two lineage groups; we generalize this from a binary split to a continuous kernel in which a drug–feature coefficient for one lineage borrows strength from other lineages with a weight that decays with tree distance. SiamCDR's contrastive grouping of cell lines by cancer type is the categorical limit of this idea (Enhancing drug and cell line representations via contrastive learning for improved anti-cancer drug prioritization); the tree replaces its discrete cancer-type labels with graded evolutionary distance.

Proposed Experiments

The primary benchmark is the cross-panel stress test designed by the guidelines paper (Drug Sensitivity Prediction From Cell Line-Based Pharmacogenomics Data: Guidelines for Developing Machine Learning Models): train on GDSC, test on GDSCv2, CTRPv2, and gCSI, predicting Area Above the Curve (AAC) rather than IC50, and excluding haematopoietic lines when modeling solid tumors. The reported reference points are a Random Forest within-domain correlation of roughly 0.41 falling to roughly 0.33 cross-domain from that paper, Precily's independent-test Pearson correlation of 0.84 on CTRPv2, and CellHit's Pearson correlation of 0.89 on GDSC (Gene expression based inference of cancer drug sensitivity; Learning and actioning general principles of cancer cell drug sensitivity). Success is defined as a smaller cross-domain correlation drop than every baseline under identical data splits, and the honest target is the drop relative to Precily, since Precily's 0.84 on CTRPv2 is the strongest cross-panel number in the file.

The second experiment is the missing-state test, which isolates the claim nothing else in the literature can make. We mask the transcriptomes of a growing fraction of CCLE lines (the same panel used by CHER and Precily), reconstruct their pathway states via ancestral state reconstruction, and predict their drug sensitivities from reconstructed features, comparing against flat k-nearest-neighbor imputation by expression similarity, tissue-mean imputation, and dropping the lines entirely. Varying the masked fraction from 10% to 90% traces the curve of how much profile sparsity each method tolerates — the setting where static GSVA and LLM-selected features structurally cannot help, because they require complete transcriptomes.

The third experiment is the lineage-structure test on CHER's own turf: the CCLE subsets of 70 blood, 52 breast/ovary, and 63 skin/glioma lines, comparing CHER's binary split against continuous tree-weighted pooling under identical Elastic Net base learners and identical tuning budgets, with the reference point being CHER's reported mean Pearson improvement of 0.24 over Elastic Net in 60% of drug–phenotype pairs (Context Sensitive Modeling of Cancer Drug Sensitivity).

The fourth experiment is the copy-number ablation: pathway-based response models trained with and without aneuploidy correction, evaluated on whether correction specifically reduces the tendency to separate lineages rather than mechanisms, measured as the change in prediction error when a drug's response pattern is lineage-dependent versus mechanism-dependent.

Baselines

The do-nothing reference is a model that predicts each cell line's AAC for a drug as the average of its tissue type, which captures how much of the cross-domain signal is pure lineage identity. If our method does not beat this within each tissue, it has not modeled anything beyond tissue labels. The strongest flat baseline is Precily's recipe (Gene expression based inference of cancer drug sensitivity): GSVA pathway scores plus SMILESVec drug descriptors, reported at a Pearson correlation of 0.84 on CTRPv2 independent testing, the best cross-panel figure in the file. We will re-implement it, tune it with the same hyperparameter budget and identical AAC targets and line exclusions, and treat beating its cross-domain correlation as the primary success criterion — noting that Precily's within-panel numbers (0.88 on GDSC) are not the relevant comparison, since our claim concerns generalization and missing profiles, not in-domain accuracy. The strongest lineage-aware baseline is CHER (Context Sensitive Modeling of Cancer Drug Sensitivity), run on its original lineage subsets and on the cross-panel task, since it is the only published method that transfers information between related lineages. The Random Forest under the guidelines paper's protocol (Drug Sensitivity Prediction From Cell Line-Based Pharmacogenomics Data: Guidelines for Developing Machine Learning Models) provides the calibrated stress-test reference of roughly 0.41 within-domain and 0.33 cross-domain. The critical ablation is flat k-nearest-neighbor imputation on expression correlation in the missing-state experiment: it isolates whether the tree's hierarchy and branch lengths add anything over unstructured similarity, and it is the ablation that decides whether the proposal's central ingredient is real.

Why not just…?

Why not just add tissue-of-origin covariates or regress out tissue effects? This is the first thing a reviewer will propose, and the file contains direct evidence against it: CHER's authors evaluated additive tissue correction and found it fails because lineage context enters through non-linear interactions, not additive baseline shifts — the textbook example being BRAF V600E, which confers inhibitor sensitivity in melanoma but not colon cancer, a pattern that subtracting tissue means destroys (Context Sensitive Modeling of Cancer Drug Sensitivity). Our method keeps interactions but weights their sharing by lineage distance. Still, this alternative could match us on drugs whose response is purely additive in tissue; the cross-panel experiment will show whether it does.

Why not just use adversarial domain invariance to erase cell line identity? The domain-adversarial autoencoder approach works in principle, but it has been demonstrated only on two triple-negative breast cancer lines, and it treats all lineage divergence as noise to be removed (Domain-invariant features for mechanism of action prediction in a multi-cell-line drug screen). Erasing lineage identity removes exactly the information needed to impute a missing line's profile or to explain why a background drives differential sensitivity; it also cannot produce estimates for lines with no data at all. It is a real competitor for the mechanism-of-action classification task, and we will run it on our multi-line setting, but it cannot address the missing-state problem that motivates half this proposal.

Why not just use static GSVA scores or LLM-selected genes? Precily's static pathway scores reach a Pearson correlation of 0.88 on GDSC and 0.84 on CTRPv2, and CellHit's LLM-primed features reach 0.89 on GDSC (Gene expression based inference of cancer drug sensitivity; Learning and actioning general principles of cancer cell drug sensitivity). This is the strongest threat to the proposal, and we concede that if the profile is complete and the test panel is close to training, static GSVA likely matches us. Our claim is therefore deliberately narrower: the masking experiment, where transcriptomes are absent, is designed so that static GSVA and LLM selection have no input at all, and our cross-panel test is run under the guidelines' hardest splits. If tree-based reconstruction does not beat Precily on the cross-domain drop, the static alternative wins and we will report that.

Why not just fix the response metric (AAC) and filter cell lines? The guidelines paper shows that switching from IC50 to AAC and removing haematopoietic lines from solid-tumor models stabilizes predictions (Drug Sensitivity Prediction From Cell Line-Based Pharmacogenomics Data: Guidelines for Developing Machine Learning Models), and we adopt these choices throughout — but the same paper reports that even after these fixes the cross-domain correlation still falls from roughly 0.41 to roughly 0.33. Metric hygiene is necessary and free, and it is already in every experiment we run; it is not sufficient to close the gap, which is why something structural is worth trying.

Decisive Checks

Three results would settle the proposal's fate. First, in the cross-panel stress test, a smaller correlation drop than re-tuned Precily under identical protocol — or a clear failure to close the gap, which we would report as evidence that lineage structure does not help generalization. Second, in the masking experiment, ancestral state reconstruction predicting masked pathway states and drug sensitivities better than flat nearest-neighbor imputation across masking fractions; if flat similarity matches the tree, the PICRUSt transplant adds nothing and the tree should be discarded in favor of a similarity kernel. Third, in the copy-number ablation, a measurable reduction in lineage-separating shortcuts after aneuploidy correction, echoing the guidelines' finding that uncorrected lineage baselines dominate model behavior. Each check has a pre-registered failure mode, because the proposal's claim is specific enough to be wrong: that evolutionary structure, not just similarity, carries signal about unmeasured functional states in cancer cell lines.