28 of 30 contrast × method slots have a scored model (held out: donors); best held-out AUC 1.000 (tumor vs immune cells · new screen 1 all k-mer); the combined hash + new screen 1 model beats the hash model in 4 of 5 contrasts

What Yost et al. 2019 found, what P2G found, and which genes overlap

The study (Yost et al. 2019, Nat Med): scRNA-seq and scTCR-seq of BCC/SCC biopsies before and after anti-PD-1. After treatment, exhausted CD8 T cells expanded. Their TCRs showed these were mostly new clonotypes not seen in the tumour before treatment (clonal replacement), not pre-existing tumour-infiltrating T cells being reinvigorated. A co-expression analysis gave a core exhaustion signature: HAVCR2, TIGIT, the tissue-resident markers ITGAE and CXCR6, and ENTPD1/CD39. On the diffusion map, one axis tracked exhaustion genes (PDCD1, HAVCR2) and another tracked activation genes (IFNG, TNF). The paper reports no classifier AUC.

What P2G found (best held-out-donor AUC per contrast): CD8 exhausted vs activated 0.978 · before treatment only 0.965 · after treatment only 0.955 · after vs before treatment 0.935 · tumour vs immune 1.000. SPLASH-only models are much weaker (0.46–0.71) and put 91–97% of their weight on HLA genes. The pooled exhausted vs activated test set has only 47 activated cells, so its AUC is fragile. The before-only and after-only splits are more balanced.

Bottom line: From sequence alone, P2G recovers the paper's exhaustion vs activation split, with TIGIT and CXCL13 on one side and IFNG and early-activation genes on the other. It misses PDCD1, HAVCR2 and ENTPD1. The after vs before treatment model is mostly HLA, so it is likely learning donors. The paper's main result, clonal replacement, is TCR-based and outside what these contrasts measure.

28 models over 5 contrasts, scored on held-out donors

Dataset: GEO GSE123813 (Yost et al. 2019) · basal and squamous cell carcinoma biopsies before and after anti-PD-1 · 48 SRA runs · 78,707 cells in the metadata · 14 donors. Counted over: every contrast × method with an output under the S3 prefixes sent for this page. Each contrast keeps only the cells of its classes. P2G keeps only donors with cells of both classes, then holds out whole donors.

Each row is one model. hash is the P2G stage 3 model; new screen 1 ranks targets by presence ratio between the classes; combined fits one lasso on the union of both candidate sets; combined dict averages an all-k-mer model and a SPLASH model. Binary contrasts report held-out AUC; regressions report Spearman ρ (rank agreement of predicted and measured values) and R² (share of the variance explained compared with predicting the mean).

Yost et al. 2019 reports no classifier result, so no published value sits beside P2G's

The classes are the authors' cluster labels (exhausted vs activated CD8, tumour vs immune) and the treatment time point

Counted over: the cells of each contrast, joined to the exact train and test cell IDs in each run's run_split.rds. Labels come from the authors' cell metadata (cluster, cluster_tp, treatment). Donor, tumour type and sort gate are confounded in this dataset.

Stage 1, all anchors: 15 to 3,643 of about 143,000 anchors pass |r| > 0.2

Counted over: every anchor in each hash run's stage1_pearson_all.tsv.gz, before any selection. Passes per contrast: CD8 exhausted vs activated 1,323 · CD8 exhausted vs activated, before treatment 348 · CD8 exhausted vs activated, after treatment 1,794 · after vs before treatment 74 · tumor vs immune cells 3,643. r is computed on read count × f, so its sign follows f; the page shows |r|. An anchor's gene is the gene of its top target in stage2_cv_pass.tsv.gz.

New screen 1 keeps targets present in enough cells, ranks them by presence ratio, and picks the cut K by training cross-validation

Counted over: every anchor-target pair in each contrast's new screen 1 run, read from its run.log. Exclusive targets are present in one class only; graded targets are present in both and ranked.

Selected sequences: 16,059 of 25,125 selected rows (summed over the 28 models) match the genome exactly

Counted over: the sequences each model selected (non-zero weight): 25,125 across the 28 models, out of 38,467 rows in their tables. 1,785 of them have no target recorded.

Selected genes vs expression: how much of each model's weight sits on highly expressed genes

Counted over: the genes the selected sequences are drawn on, per model, weighted by summed |weight|. Expression: GSE123813_bcc_scRNA_counts.txt.gz and GSE123813_scc_scRNA_counts.txt.gz (GEO GSE123813), normalised to counts per 10,000 per cell.

Literature: what PubMed says about each heavy gene with T-cell exhaustion and PD-1 blockade

Counted over: the 15 heaviest named genes of each model. PubMed title/abstract searches: gene + T-cell exhaustion, gene + PD-1 blockade, and gene + both (a paper that names the gene and both topics).

79 alternative-splicing, 3,242 mismatch (1 or more) and 1,009 BLAST-only features (18 of them imperfect) among the selected sequences of the 28 models

Counted over: the selected sequences (non-zero weight) of every model, after the 1% rule: a sequence present in fewer than 1% of training cells is left out (407 selected sequences, almost all from hash models; new screen 1 already requires at least 5%). (a) crosses a STAR junction that is not in the reference transcript; compatible transcripts have that exact intron, incompatible ones span it but splice differently or keep it. (a) also counts a soft-clipped sequence that is full length only on alternative isoforms. (b) edit distance of 1 or more from the aligner that placed it, on a full-length placement (split into 1 and 2 or more); soft-clipped and unplaced sequences have no comparable edit distance and are left out of (b). (a) or (b) is counted once per sequence, with the overlap shown, over the selected sequences and over all stage 3 rows. This follows section 3h of the GSE123813 instructions (Julia, 2026-10-08). (c) placed by neither aligner, but hit by local BLAST (GENCODE, genome, IPD-IMGT/HLA) or NCBI core_nt; perfect = 100% coverage and 100% identity. A sequence can be in more than one group. The same features for every study are collected on the running feature ledger.

Each selected sequence drawn on its gene, marked by how it matches the genome and how other isoforms read it

Drawn: the selected sequences with a genome position. Listed in the table under the track: every row on that gene, selected or not.

Choose a model, then a gene. Genes are listed by the total weight of their selected sequences. Click a mark or a table row for its card.

4,225 selected sequences (summed over the 28 models) did not align confidently and were BLASTed locally

Counted over: selected sequences of the chosen model (section 4's model menu). The pool is every sequence without a confident full-length alignment with at most 1 mismatch; sequences with no target recorded are counted but cannot be searched.

Every sequence with 2+ mismatches or soft-clipping, BLASTed against mRNAs, the whole genome and HLA alleles

Counted over: every row with 2 or more mismatches (full length) or any soft-clipping, selected or not. Coverage = aligned bases ÷ sequence length; imperfect values are amber.

5c · Selected sequences never BLASTed, and the ones at a known SNP or with a base change

Counted over: the selected sequences of the model chosen in section 4.

Selected sequences at a known dbSNP site, a ClinVar or COSMIC record, or with a base change called by P2G

Examples for every prediction: the heaviest variant-bearing sequences of each contrast and model

Counted over: the variant table of each of the 28 models (heaviest 60 selected sequences each). For each contrast and model, the table shows the 3 heaviest distinct gene + variant pairs. HLA rows are marked: HLA alleles differ between donors, so they may track donors rather than cell state. Example: CD8 exhausted vs activated, after treatment. Known variants in HLA and Fc-gamma genes, for every model, are in section 6c.

The chosen model, in full

Counted over: the selected sequences of the model chosen in section 4 (heaviest 60).

How much weight sits on exhaustion markers, Fc gamma receptors and HLA genes?

Counted over: the selected sequences of every model, by summed |weight|. Groups: canonical exhaustion markers, the Fc gamma receptors with FcRγ, and HLA class I and II genes. HLA genes differ between donors by allele, so weight on them can track donors rather than cell state. The lookup counts every row, selected or not. Section 6c lists every HLA and Fc-gamma sequence per model, with known variants; it also counts HLA sequences found only by BLAST, so its HLA shares can differ from the ones here.

Immune genes that vary between people: how much of each model’s score is in HLA and Fc-gamma genes?

Why this matters: a model can learn which person a cell came from instead of which condition it is in, and HLA is the most variable region of the human genome. This page tests every model on patients it did not train on, which limits how much a model can gain from recognising one training patient.

25 of 28 models are flagged: HLA and Fc-gamma genes make up 5% or more of their score.

1 model selects no sequence in these genes: tumor vs immune cells · hash SPLASH.

The selected sequences in these genes, model by model

Click a model to open its sequences. 27 models have at least one.

Checks run on this build (they inform, they never block)

Counted over: every model table read for this page.

Annotation coverage per model (rows with a value)

Every file this page was built from, with its S3 path

Read: the files below, the GEO expression matrix and guide barcodes, GENCODE v46, GRCh38.p13, IPD-IMGT/HLA 3.65.0, NCBI gene_info and MSigDB hallmark sets.

Every rule this page followed

Source

Built to match the GSE123813 sequence-review page (artifact 6ChpFqMvKXfBMEbrJT1LsV) and the page instructions in rosabiocom/rosa-skills (population-comparison-site, *_PAGE_INSTRUCTIONS.md). The original build scripts are not in the skill, so this page was rebuilt: Julia's page code (renderer, gene_track_annot.js, styles) is reused, and every data file is regenerated by the shared build in website-artifact/sites/study-build (build_rows.py, run_blast.sh, blast_classify.py, expr_stats.py, lit_fetch.py, build_page.py, assemble.py).

Models

One page per study, every contrast on it. Each contrast has up to six models, chosen in every dropdown: hash all k-mer and hash SPLASH (P2G stage 3, stage3_annotated_target.tsv.gz of each dictionary), new screen 1 all k-mer and new screen 1 SPLASH (presence-ratio final-selected tables), combined (Combined_selected_seq: hash + new screen 1 candidates in one lasso) and combined dict (all-k-mer + SPLASH dictionaries, equal-weight). A slot without an output is listed in section 1 with the reason.

Scores

Binary contrasts: held-out AUC (hash: roc_auc of test_evaluation.tsv at feature_level = anchor_target; the anchor-level AUC of summary_row.tsv is shown for reference). Regressions: Spearman ρ between predicted and measured values on held-out cells, and R², the share of the variance explained compared with predicting the mean.

Weight and direction

hash: lasso_coef × f; the other three: the model coefficient. Positive pushes toward the first class of the contrast, negative toward the second. For a regression, positive raises the predicted value and negative lowers it.

Rows

One aligner per sequence (minimap2 when it placed the sequence, otherwise bowtie2); CIGAR, edit distance and cs all from that aligner. A dotted or empty cell is "none". Gene = GENCODE v46 gene overlapping the drawn position, protein-coding before lncRNA before pseudogene, named before ENSG-only. Rows with no target sequence are "no target recorded" and never placed or searched. SPLASH features of the combined-dictionary model take their sequence from the SPLASH new screen 1 table when it has one.

Isoforms

GENCODE v46 comprehensive. The reference row is MANE Select, else Ensembl canonical. An isoform row is kept only when it reads a selected sequence differently (stopping short is not a different reading); identical readings are collapsed; at most 8 rows; every transcript is accounted for in the caption.

Junctions

STAR star_junction matched exactly to the introns of every GENCODE v46 transcript: reference, annotated alternative, or not annotated (new pairing, one new site, no annotated site).

BLAST

Local first: GENCODE v46 transcripts (blastn-short, dust off, E ≤ 1, ≥ 90% identity, up to 500 targets), GRCh38.p13 top level with alternate haplotypes and patches (word size 11, E ≤ 10, ≥ 95% identity) and IPD-IMGT/HLA 3.65.0. Pool = no confident full-length alignment with ≤ 1 mismatch. Rescue bar: ≥ 30 bases at ≥ 97% identity; transcripts first, then the genome (E ≤ 1e-4). A tie across genes is ambiguous; a genome locus is a position, never a gene. Every sequence no local search explains, and every 5b "no closer match" or "clipped bases unexplained" sequence, was then searched remotely in NCBI core_nt (megablast, word size 16, E ≤ 1e-3, up to 25 targets); the card shows the best hit, its organism and the tie count. Other pool sequences were not sent remotely; a sequence still unsearched would be marked "would need a remote search".

Expression

GEO GSE123813_bcc_scRNA_counts.txt.gz and GSE123813_scc_scRNA_counts.txt.gz (GEO GSE123813), CP10K per cell. Each model's two groups are its own train + test cells from run_split.rds (a regression is split at the median of the response). Matrix columns are matched to P2G cells by barcode. Ranks are among genes expressed in the two groups together; a GENCODE name missing from the matrix is matched through NCBI synonyms.

Literature

PubMed title/abstract counts for gene + T-cell exhaustion, gene + PD-1 blockade, and both, for the 15 heaviest named genes of each model (symbol, full name, NCBI synonyms of 4+ characters with a digit or hyphen). Up to three papers per gene with the abstract sentence that names the gene, verbatim. Each linked paper has a one-line conclusion written from its title and abstract only, with a relevance tag (central, finding, gene list only, unrelated match, no abstract).

Gene census

Groups: canonical exhaustion markers, the Fc gamma receptors with FcRγ, and HLA class I and II genes. HLA genes differ between donors by allele, so weight on them can track donors rather than cell state. The lookup counts every row, selected or not.

Literature: splicing, mismatches and BLAST per gene

Requested 2026-10-08. For each of the 15 heaviest genes in the literature table (3f), two columns summarise that gene's selected sequences (non-zero weight) in the chosen model: alternative splicing (STAR junctions classed against GENCODE v46, plus sequences exonic only in non-reference isoforms and the mRNA-search junction verdict) and mismatches · BLAST (edit-distance strata 0 / 1 / 2+, soft-clipped and unplaced counts, the local BLAST verdicts, and NCBI core_nt hits). A sequence belongs to the gene when it lies in the gene's track, or its stage-3 gene_name, junction annotation or BLAST transcript hit names the gene.

Feature report (section 3g) and the running ledger

Requested 2026-10-08. Every rebuild runs feature_report.py on the selected sequences (non-zero weight) of every model and keeps the ones in three groups: (a) alternative splicing, a STAR junction not in the reference transcript, with the GENCODE v46 transcripts compatible with it (that exact intron) and incompatible (span it, splice differently or keep it); (b) 1 or more mismatches from the aligner that placed the sequence; (c) no aligner placement but a BLAST hit (local, then NCBI core_nt), perfect or imperfect. The 1% rule is applied: a sequence present in fewer than 1% of training cells is left out (hash rows: Train_prev weighted by class size; new screen 1 already requires 5%; combined rows take the share of the same sequence in the hash or new screen 1 table). build_ledger.py then collates every study into the running feature ledger with study, phenotype, label basis, model, gene, direction, target rank, coefficient, weight and share of cells.

Not done

Cell-type marker status and LitVar2 variant papers were not run for this build. Never downloaded raw reads.

Published study

Yost et al. 2019 (PMID 31359002) reports no classifier AUC for these contrasts.