Ranking neoantigens#

What you get. One ordered table from a candidate list, scored by the fitted EPIC aggregate: nine terms in four blocks — Expression, Presentation, Immunogenicity, Complementarity.

mhcmatch genes pairs.tsv --out annotated.tsv                     # if your table has no gene column
mhcmatch rank pairs annotated.tsv --tumor SKCM --out ranked.tsv  # the score
mhcmatch rank --coefficients                                     # every term and its coefficient

What this page is. What each term is, how it is computed, what it was fitted on, and what it does not do. One page per block is linked at the bottom and from the sidebar. Every number here is recorded in the benchmark repository under bench/results/ and cited from there rather than restated from memory.

Two things to know before reading a score. The blocks enter in pipeline order, so a later block’s coefficient is what that term is worth after the earlier ones — never in competition with them. And a candidate missing a term is scored on the terms it has, flagged, never dropped.

The model #

mhcmatch rank scores each candidate with the fitted aggregate vendored at data/aggregate_mhc1.json, which declares itself EPIC: nine terms in four hierarchical blocks, one unpenalised intercept per screen. The artifact carries its own version — a model version, an integer, distinct from the package version. Read the feature list from mhcmatch.rank.AGGREGATE_FEATURES and the grouping from mhcmatch.rank.AGGREGATE_BLOCKS rather than typing either out.

The blocks are entered in pipeline order, each on top of the last, so a recognition coefficient is what that term is worth after presentation and expression rather than in competition with them.

This page is about the terms — what each one is and what it is computed from, for the nine-term mhc1.human.neoantigen fit. The other seven artifacts are the same machinery on a different term set: six for both class-II neoantigen cells, and for the pathogen cells two at class I and three at class II. For the eight fitted cells themselves — their coefficients with bootstrap intervals, what each was fitted on, how well each scores and what each cannot be asked — see The shipped models.

block	term	what it is
`presentation`	`binder`	`-log10` of the calibrated combined %rank — the Fisher statistic over the presentation rank and the Potts affinity rank, read as a percentage. Allele-relative: where this peptide sits in its own allele’s distribution. The presentation rank alone is the separate key `pres`, which is still computed and is not fitted.
	`log10a`	The density axis on its log-odds scale, `log10(a)` for `a = [P]/Kd`. This is exactly the logit of `mhcmatch.rank.occupancy()`, since `occ/(1-occ) == a` identically. Absolute, where a %rank is allele-relative, and defined without a wild type. Occupancy itself is still computed and emitted; a probability entered linearly in a log-odds model is the mis-specification, not the axis.
`expression`	`expr_lvl`	`log2(1 + TPM/c)` for this candidate’s source-gene abundance — the cohort’s own measurement where it has one, else the tumour type’s reference value, else the gene’s matched-normal or cross-tissue level. `c` is the 25th percentile of the tumour type’s own non-zero gene medians, 0.1400 to 0.2400 TPM over the fit’s seven screens. See `mhcmatch.rank.expr_level()`.
`expression`	`expr_norm`	The same gene’s median in the tumour’s matched normal tissue, on the same floor, falling back to that gene’s pan-tissue median and never to missing. Free rather than subtracted: a ratio would need equal and opposite coefficients, and both are positive. See `mhcmatch.rank.expr_norm_level()`.
`physchem`	`C_phys_buried`	`mhcmatch.complement.burial()`, the Rose burial propensity averaged over the TCR face. An imported scale, so zero fitted residue parameters. See The recognition axis, reduced.
	`C_phys_charge`	the same, on Atchley AF5 (electrostatic charge). v3 paired burial with Kidera KF4 hydropathy instead, and that pair was collinear — r = −0.837 per peptide, one chemistry axis in two columns — so burial was not identified. AF5 is orthogonal to burial at r = +0.008, which is what resolves it. The recognition axis, reduced owns the selection.
`corpus`	`C_corpus_thymus`	`mhcmatch.mimicry.corpus_R()` on the thymic channel — the exact Łuksza density over the TCR face, label-free. Reads as danger, coefficient positive. See Corpus complementarity: what the repertoire was shaped by.
	`C_corpus_self`	the same against the host proteome, coefficient negative. Not a tolerance measurement on its own — alone it is p = 0.69 and its marginal AUROC is below chance. It is the corpus block’s background term, the reference level the other two are read against; remove it and both of them fall to non-significant. Corpus complementarity: what the repertoire was shaped by has the subset ladder that shows this.
	`C_corpus_viral`	the same against a foreign presented ligandome — a thymocyte never sees this during selection, so a hit is about peripheral priming.

What changed in v3, and why. C_corpus used to be the thymic channel alone, computed by a radius-2 trie walk that captured a median 0.4999 of the sum its own definition calls for, from a cache that had no entry at all for three of the nine fitting screens. It is now the exact sum, evaluated as a k-mer table contraction, for every row; self and viral came back into the model because the contraction removed the ~7.5 GB index that had priced them out, not because anything about them changed. C_corpus_missing left with the cache — there is no gap left to flag. And C_phys was a sum over a face of width L − 5 on a strictly positive scale, so it correlated +0.954 with peptide length; averaging it fixed that and made the two scales comparable for the first time.

An earlier arrangement carried the 30-column mhcmatch.complement.score() as one term, the Łuksza viral_R as R and the three TCR-face mimicry densities as T. luksza.viral_r and complement.score still ship and are still computable; they are no longer terms of the shipped model. Every alternative is re-measured in bench/results/epic_recognition_terms.md.

Standardisation (mu, sigma) travels inside the artifact, so a caller reproduces the score exactly. A feature you cannot supply contributes its training mean — which is what “no information” should do — so a candidate with no expression value is scored on the terms it has rather than dropped.

What the coefficients are #

Standardised, so a coefficient is the log-odds shift per standard deviation of its own column and the sizes are directly comparable. One unpenalised intercept per screen, ridge \(\tau\) = 0.25; z, p and sign stability come from a cluster bootstrap over (patient, screen) — 400 resamples — because rows from one patient share tumour, HLA and run.

The values are not typed into these docs. Six pages used to carry their own copy of this table and all six went stale together the first time the model was refitted, because nothing read them. The shipped models prints all eight cells in full — generated from the artifacts on every build, which is the only form of that table that has ever stayed correct. The vendored artifact is the record, and the command line prints it too:

mhcmatch rank --coefficients    # every term, its block, its coefficient, z, p, sign stability
mhcmatch rank --holdout         # per-screen held-out AUROC, both grouped CVs, the fitted corpus

import json, importlib.resources as R
d = json.loads(R.files("mhcmatch.data").joinpath("aggregate_mhc1.json").read_text())
d["version"], d["features"], d["coef"], d["fit"]["rows"], d["fit"]["screens"]

The manuscript’s tables/epic_terms.tex is generated from that same artifact and is the citable form. What these docs carry instead is what each term is, which does not move when it is refitted.

term	block	what it is
`expr_lvl`	expression	this candidate’s own abundance as `log2(1 + TPM/c)` on the tumour type’s floor
`expr_norm`	expression	the same gene in the tumour’s matched normal tissue, on the same floor
`binder`	presentation	the calibrated Fisher combination of presentation `%rank` with the Potts affinity `%rank` — a within-allele competition statement
`log10a`	presentation	groove occupancy at `PEPTIDE_NM` on its log-odds scale, an absolute surface-density statement
`C_corpus_thymus`	corpus	danger — density against the thymic immunopeptidome
`C_corpus_self`	corpus	the block’s background — see Corpus complementarity: what the repertoire was shaped by
`C_corpus_viral`	corpus	peripheral priming — density against the foreign ligandome
`C_phys_buried`	physchem	Rose burial over the TCR face, per residue
`C_phys_charge`	physchem	Atchley AF5 charge over the TCR face — see The recognition axis, reduced

binder and log10a are two necessary conditions, not one quantity measured twice. A %rank asks whether a peptide out-competes the self peptidome its allele normally loads; occupancy asks how many copies reach the surface at a stated free-peptide concentration. Winning a groove does not imply reaching the copy number a T cell needs, and reaching it does not imply winning the groove — which is why they sit at Spearman \(\rho\) = +0.7431 rather than 1.

Two terms are doing something other than what their name suggests, and both are documented rather than tidied away:

C_phys_charge’s own p is not the statistic to read on it. It earns its place by what it does to its partner: burial beside Kidera KF4 was not identified (one axis at r = −0.837), and beside charge (r = +0.008) burial’s standard error halves and it resolves on a smaller coefficient. See The recognition axis, reduced.
C_corpus_self’s large negative coefficient is a background subtraction, not tolerance.

Two more things no coefficient table says, so read them here.

A coefficient is conditional on its block being entered. The blocks go in pipeline order, so a recognition coefficient is what the term is worth after presentation and expression. Adding the whole corpus block gives the largest held-out gain of any recognition block.

No term was dropped for being small. The rule is replace-and-recalibrate: a term with a mechanism stays and gets a better basis, it does not get deleted for a p-value. C_corpus_viral stays because dropping it costs the other two channels their significance.

Terms that are computed and never scored. pres, occupancy, dai, agretopicity, d_occupancy, wt_absent and any Kidera scale reachable through complement.burial(..., scale="KIDERA:KF4") are all emitted for comparison and none is a fitted term. That separation is asserted in the test suite, not merely intended: nothing in that list may appear in rank.AGGREGATE_FEATURES.

From a score to a probability #

score is a log-odds ranking, and rank is that ranking as a dense 1-based integer. Neither is a probability, because the fit deliberately has no shared intercept: every screen got its own, unpenalised, precisely so prevalence and candidate generation stayed out of the slopes.

The p_response column supplies the missing constant, by the only rule that needs no new data. Given a pool prevalence \(\pi\) you declare, it picks the single offset b with

\[\frac{1}{n}\sum_i \sigma(s_i + b) = \pi\]

and reports \(\sigma(s_i + b)\). See mhcmatch.rank.probability().

$ mhcmatch rank fasta candidates.fasta --alleles donor.txt --prevalence 0.06

It is a prior shift, not a recalibration. b is additive and \(\sigma\) is monotone, so it preserves the ordering exactly — halving \(\pi\) roughly halves every probability and moves no rank. What it buys is portability: a raw-score cut-off means nothing across cohorts whose base rates differ by four orders of magnitude, and “P ≥ 0.2 at an assumed 6 % pool prevalence” is a statement another cohort can be held to.

The default is TESLA’s 37 immunogenic of 615 tested candidates — the community benchmark whose whole design is “a pipeline nominated these; which respond”. It is a prior, not a measurement of your cohort, and it is the single number the emitted probability is most sensitive to. Anchors:

pool	responds	what it is
Neopep	0.0060 %	19 of 318,197 — an exhaustive scan, nothing filtered
TESLA (the default)	6.0 %	37 of 615 — nominated candidates, tested
ITSNdb	59.7 %	89 of 149 — a curated, positive-enriched set
assayed vaccine units	~19 %	41 of 216 units, 13 patients (Sahin et al., Nature 2026;651:1088–1096) — already through a cassette selection

Why occupancy and not agretopicity #

The obvious term for “is the mutant better presented than its wild type” is the differential agretopicity index, log10(Kd_WT / Kd_MT). It does not work, and the reason is worth stating because the quantity is widely used.

Fitted on the cleaned corpus it carried a negative coefficient with an interval crossing zero, and the marginal was negative too: within-screen median AUROC 0.4986 against the binder %rank’s 0.6383, below 0.5 in nearly every screen. Seven parameterisations were tried — the raw ratio, the Łuksza pseudocount at five saturation scales, a logistic squashing, gating on anchor substitutions, gating on genuine binders, the interaction with binder, and the concentration-free share A/(1+A). None resolves.

One appears to. Tightening the pseudocount drives the coefficient to z +3.32 — but at that setting the term is 0.9955 correlated with -log10(Kd_MT). It improves by deleting the agretopicity from itself, leaving mutant affinity under another name, and the wild-type residual sits at 0.4526 at every setting. See bench/results/neoag_dai_terms.md.

Occupancy and agretopicity derives what the equilibrium supplies instead. Three things we had been adding by hand fall out of it rather than being imposed: the binder gate, because a mutant that does not bind occupies nothing whatever its wild type does; the pseudocount, because the free-MHC 1 in the denominator is Łuksza’s ε — which is why that ε carries units of inverse concentration; and a bounded scale whose steepness is fixed at 1 rather than fitted.

agretopicity is still emitted as a column. It is not a term of the model.

What the fit rests on #

The corpus is the union of every human and mouse neoantigen deposit, harmonised and cleaned, with each exclusion measured rather than assumed:

host keyed on the MHC genus, never a host_species column;
pathogen epitopes dropped — a peptide recorded only under Influenza A, M. tuberculosis or SARS-CoV-2 is not a neoantigen whatever its immunogenicity;
peptides identical to a host-proteome window dropped — no somatic alteration, no neoantigen;
label conflicts resolved positive-wins and counted;
CEDAR and Gfeller retained but held out of the fit, their contamination characterised.

Counts, the full filter cascade and the per-cohort breakdown are in bench/results/neoag_corpus_grand.md and neoag_corpus_noise.md.

What to feed it #

Two entry points with different contracts. rank table re-ranks rows another tool scored; rank fasta predicts de novo from variant windows.

`rank table` — the four things it actually reads#

A CSV in the pipeline .scored.csv shape, but only these columns are consulted:

column	needed?	what it is
`epitope`	yes	the peptide. Rows with a missing or non-alphabetic value are skipped silently
`best_allele`	yes to re-score	the restricting allotype, in the panel’s own form. A paired class-II molecule is one composite key, not two columns — `HLA-DPA10103-DPB10401`, `HLA-DQA10501-DQB10301` — while DR is the beta chain alone, `DRB1_1501`, because DRA is effectively monomorphic. Without it, presentation and `binder` are `NaN` and only the sequence-derived terms survive
`tpm`	no	expression. Absent, the fallback below runs and `expr_imputed` is set
`gene_name`	no	the parent gene’s HGNC symbol, used to look expression up when `tpm` is absent and `--tissue` is given, and by `expr_norm` always. `gene` — what `mhcmatch genes` writes and what `rank pairs` reads — is accepted under that name too, so an annotated table needs no rename; see The parent gene, when the deposit does not name one

ref_seq / seq are read when present, and an incoming score is preserved in components['score_builtin'] so the two rankings can be compared. Everything else in the 57-column schema is ignored — passing it is harmless, omitting it costs nothing.

`rank fasta` — expression, in three ways#

The window FASTA header is the pipeline’s Somatic: record, and of its thirteen fields exactly two reach the ranker: tpm and gene_name. When tpm is in the header, it is used and expr_imputed is 0. When it is not, mhcmatch.rank._expression_for() falls back, in order:

--tumor SKCM — TCGA, keyed on the peptide, so it is specific to the epitope;
--tissue pancreas — GTEx, keyed on the gene;
neither — expression is NaN, expr_imputed is 1, and the row still ranks. A missing covariate never drops a candidate: a non-finite term standardises to z = 0, the fitting corpus’s own mean, which is what “no information” means on that scale. The gap is recorded twice and in two senses — expr_imputed says the abundance was not measured, and the imputed column names every term that fell back. The retired expr_missing fitted indicator is a different thing — its source was very nearly a screen label, so the per-screen intercept already carried it (bench/results/epic_expr_arms.md).

--tumor is worth passing even when the abundance column is present, because it sets c. A tumour’s floor is roughly half its matched normal’s, so the pooled fallback is not a neutral choice. Where the origin arrives as free text, mhcmatch.expression.resolve_context() maps it — "liver", "LIHC" and "hepatocellular" all resolve — and raises on a string it does not recognise, rather than returning a plausible number computed from the wrong distribution.

Note

TPM or FPKM: the unit cancels, but only while the floor comes from the same column.

log2(1 + x/c) is unit-free whenever c is a quantile of the same measurement as x: multiply the column by any constant and c moves with it. The shipped ``c`` is a quantile of the reference, in TPM, so that condition holds for a submitted TPM column and fails for one in FPKM or in counts — silently, since nothing in the number says which it was.

The repair is mhcmatch.expression.batch_scale(), a median of ratios against the reference for the tumour type, and it is gated on covering at least half that context’s expressed genes. Handed a whole transcriptome it recovers a known factor exactly across \(10^{-3}\) to \(10^{6}\); handed a candidate list it refuses, and should. A mutation reaches such a list only where the gene was seen in RNA, so the ratio measures that conditioning rather than the library: on screens whose columns are all deposited as TPM, the ungated estimator returns 1.78, 2.18 and 3.15, never below 1. Counting more candidates cannot fix it, which is why the guard is coverage and not count.

The arithmetic below is why the two metrics differ by one constant in the first place.

The two differ by a single library-wide constant, identical for every transcript:

\[\mathrm{TPM}_i = \mathrm{FPKM}_i \cdot \frac{10^6}{\sum_j \mathrm{FPKM}_j}\]

Length enters when going from counts to either metric and cancels in the ratio between them. Checked on a 20,000-gene simulation: the per-gene TPM/FPKM ratio is constant to \(2.2\times10^{-15}\) and uncorrelated with transcript length.

Two consequences. Re-ranking one sample is safe either way — a constant factor shifts every candidate by the same amount and cannot reorder them within a patient. Converting exactly needs the whole table, not the FASTA: renormalise the sample’s FPKM column to sum to \(10^6\). If you hold only per-candidate values the constant is not recoverable, and it matters for cross-sample comparison and for any absolute reading of the score, since expr_lvl and expr_norm are both fitted with positive weight — run mhcmatch rank --coefficients for the sizes the installed artifact uses, which move at every refit.

The parent gene, when the deposit does not name one#

Both fitted expression terms are keyed on an HGNC gene symbol: expr_lvl needs it whenever tpm is absent, and expr_norm needs it always. Most published neoantigen deposits ship the peptide and not the gene — over the benchmark corpus, 356,387 of 695,811 rows (51.2%) and 5,205 of 5,833 positives (89.2%) carried no symbol. Those rows do not drop out; they all collapse onto one mean-imputed constant, and a constant cannot order anything. On the VACCIMEL screen expr_norm had standard deviation exactly 0.0000 and AUROC exactly 0.5000 while carrying +0.4950 log-odds per standard deviation in the then-shipped EPIC v10 artifact — a fitted parameter paid for on no information. Repaired, it fits +0.2222 in v12. (Which term is largest moves with every refit; mhcmatch rank --coefficients prints the current set.)

The symbol is recoverable from the peptide, because a neoantigen is a near-copy of a self peptide:

mhcmatch genes candidates.tsv --species human --max-subs 2 --out annotated.tsv
mhcmatch rank pairs annotated.tsv --tumor SKCM --out ranked.tsv

genes searches the reference proteome for the nearest self peptide (mhcmatch.proteome.Proteome.assign_genes()), names it by its UniProt GN= field and writes a gene column back beside every column the table already had. Both rank pairs and rank table read that column, so the two commands compose with no join and no rename.

Over the same corpus that leaves coverage at 692,349 of 695,811 rows (99.5%), up from 339,424 (48.8%), and 4,511 of the 5,833 positives gain a symbol they were not deposited with — which is what takes expr_norm’s standard deviation on VACCIMEL from 0.0000 to 2.520, i.e. from no ordering to an ordering. bench/results/gene_resolution.md.

Three properties of the annotation, each of which is a way the axis would otherwise lose information:

The radius is 2, because a neoantigen can carry more than one mutation. The first shell does nearly all of the work — 349,921 of 695,811 corpus rows find a parent one substitution out — and the second is not rounding: 3,004 further rows need two, and on VACCIMEL it is the difference between 87 of 93 rows resolved and 90 of 93. Those are precisely the rows a radius-1 search would have left on the imputed constant.
Only the nearest shell votes. A radius-2 shell is roughly 85 times the size of the radius-1 shell inside it, so pooling the two lets a distant coincidence outvote a genuine single-substitution parent.
A tie becomes several rows, and an unresolved peptide keeps its row with an empty cell. Which of several equally-near parents a peptide should be scored under is a question the expression reference answers and the search cannot, so every tied gene is emitted and the caller takes the best score per peptide — group_by(peptide).agg(max(score)). Expect them, and expect them shortest-first: over the corpus’s 345,478 (host, peptide) pairs, 22,172 of 70,485 8-mers (31.5%) name more than one nearest gene against 7,448 of 97,995 11-mers (7.6%), because the shorter the peptide the more of the proteome sits one substitution away. Nothing in the ranker assumes one row per (peptide, allele).

Reading the output #

mhcmatch rank fasta candidates.fasta --alleles donor.txt --tumor SKCM --out ranked.tsv

score is the aggregate; higher is better. rank is that score as a dense 1-based integer and p_response is it on a probability axis at --prevalence (above). Every one of the model’s nine features is a column, because a row should report what produced it: binder and log10a, expr_lvl and expr_norm (with expression, expr_pct and expr_imputed beside them), the two chemistry scales C_phys_buried and C_phys_charge, and the three corpus channels C_corpus_thymus / _self / _viral. pres and occupancy are emitted too and are not fitted. agretopicity, physchem, variant_type and n_alleles_presenting / alleles_presenting are reported beside them and are not in the model.

variant_type is carried for the cassette layer rather than for the score: a frameshift or fusion product is foreign over a stretch rather than at one position, so it fails differently from a missense and earns a quota of its own in mhcmatch.portfolio.compose() (Cassette composition). It is the product class — missense, frameshift, inframe_deletion, fusion, isoform, cnv — and not the header’s type field, which is provenance (Somatic) and says nothing about what the variant makes; see mhcmatch.predict.variant_product(). --extended appends the remaining mimicry channels and --annotate what each candidate resembles; both add columns only and never change the ordering.

The binding core #

--core appends core, core_offset and core_source to rank, predict and neoag; the cassette map (cassette build --map) carries core unconditionally, beside the core_start / core_end it already had. It follows NetMHCpan’s definition — “the minimal 9 amino acid binding core directly in contact with the MHC (i.e. excluding potential insertions)” — with core_offset its Of, 0-based.

The core is residues, never a padded frame. The parenthesis in that definition is the operative part: where an alignment to a 9-mer motif needs a gap, the inserted position is not part of the core. So it is nine residues whenever the peptide can fill nine — every class-II core, and a class-I 9-, 10- or 11-mer — and the peptide’s own residues when it cannot, so a class-I 8-mer’s core is the 8-mer. A gap character would not be neutral in an amino-acid column in any case: B is Asx in IUPAC, and a reader would take it for a real ambiguity code.

Class I holds both anchors and lets the middle give way. The footprint is mhcmatch.diffusion.MHC1_CORE resolved by mhcmatch.store.mhc1_positions() — the same mapping the scorer uses, so the reported core is the residues the model actually read. A 9-mer is its own core. A 10- or 11-mer drops one or two central residues, which is NetMHCpan’s Gp/Gl deletion. Below nine the +5 and -4 positions collide and the losing slot is dropped — not a residue — so every residue still appears exactly once and an 8-mer’s core is the 8-mer. core_offset is 0: the footprint is anchored at both ends, so there is no N-terminal protrusion to report.

Class II is the register-anchored 9-mer, peptide[Of:Of+9], matching NetMHCIIpan’s Core and Of. Which register produced it is a column and not a footnote, because the two available registers disagree often on real ligands: core_source reads model when it came from the per-allele mhcmatch.diffusion.AnchorModel.best_register() (predict and rank fasta, where that register was already computed to score with), heuristic when it came from the allele-agnostic one-pass scan (neoag and rank table, which have no allele), and footprint for class I, where there is no register to choose.

Reported, never scored — the aggregate reads the peptide, not the core, and --core cannot move a ranking.

Warning

A model emits the features it used, and refuses to run without them. EPIC’s corpus term reads three reference deposits as three 64 KB k-mer tables, so an aggregate score builds no trie at all and --no-self is allowed with --score aggregate. --extended and --annotate do build the reference index, because they report the self channels — paid once for the whole candidate list, cached on disk and stageable prebuilt, so once per machine rather than once per run.

The imputed column names any feature that had to take its training mean for that row — a candidate with no IC50 has no occupancy, a frameshift has no wild type. Those are candidates with incomplete data, not a different model, so they are scored and the substitution is declared.

Occupancy and agretopicity #

Both come from the same competitive-binding equilibrium and differ in what the peptide is competing against — the allele’s own self-ligandome for one, its own wild type for the other:

\[\theta_{MT} = \frac{[P]/K_{MT}}{1 + [P]/K_{MT} + \sum_i [P_i]/K_i} \qquad \phi = \frac{[P]/K_{MT}}{[P]/K_{MT} + [P]/K_{WT}} = \frac{K_{WT}}{K_{WT}+K_{MT}}\]

Three properties of \(\theta\) come out of the physics rather than being imposed on it. A mutant that does not bind occupies nothing whatever its wild type does, so a binder gate is automatic. The free-MHC 1 in the denominator is exactly the pseudocount Łuksza applies to both dissociation constants — which is why that \(\varepsilon\) carries units of inverse concentration. And \(\theta\) is bounded in \([0,1]\), so the four-decade tail of the raw ratio cannot set a slope.

\(\theta\) is additive to the binder %rank, not redundant with it: a %rank says where a peptide sits in its allele’s own distribution, occupancy says how much groove it actually holds, and an allele with a permissive groove has a large self load its candidates must out-compete. Fitted together on the retired ten-screen grand-corpus fit, binder held z +6.5 while occupancy carried z +3.6 to +3.8, stable across \([P]\) from 1 to 1,000 nM.

\(\phi\) does not resolve — z −0.48, and 0.4979 on its own. Neither do the raw ratio, the pseudocount amplitude, a logistic squashing of it, or gating on anchor substitutions and genuine binders; the benchmark records all seven parameterisations. It is emitted as a column and is not a term of the fitted model.

Limits #

Held-out performance is well below in-fit performance. Leave-one-twin-group-out on the gfeller group gave 0.5781 under the retired ten-screen fit, because Gfeller and Gfeller-GBM share 96.5 % of their peptides — which is why Gfeller_GBM left the corpus for v11. Quote the twin-group column: the shipped artifact’s cv_twin mean over decided screens is 0.6957.
Not every mimicry channel is established in direction. viral_tcr and thymus_tcr flip sign in 22 % and 35 % of bootstrap resamples, which is why EPIC does not carry them. Its own sign stabilities are in bench/results/epic_recognition_terms.md.
The prior is a property of your candidate pool, not of biology. The fitted prevalence is 0.18 % (597 positives of 339,599 fitted rows), which is how these screens were assembled. Supply your own.
--score gate is the two-term product-of-sigmoids, kept for when a candidate failing either axis should not be rescuable by the other.

The four blocks, one page each

Each letter of EPIC has its own page. Read this one for how the nine terms combine; read those for what a single term is and how to get it.

Ranking neoantigens#

rank table — the four things it actually reads#

rank fasta — expression, in three ways#

The parent gene, when the deposit does not name one#

`rank table` — the four things it actually reads#

`rank fasta` — expression, in three ways#