The trained classifier's purpose was never only to score well on held-out validation genes - it exists to be pointed at the rest of the genome. Once the Phase 2 checkpoints (described in the training methodology document) were finalized, every model variant was applied to roughly 19,500 genes that carry no IRD module label at all, producing, for each gene, a predicted module and a confidence score. This is the step where the project moves from "does the model work" to "what does the model find."
Each prediction is an ensemble average: for a given gene, the five cross-validation folds' softmax probability vectors are averaged together before taking the highest-probability module as the prediction. This is a standard way to reduce the influence of any single fold's particular training/validation split on the final call.
Rather than trusting a single model's output, genome-wide predictions were generated from multiple independently-trained variants - most importantly the PPI-only and full three-tower fusion models, the two strongest performers from the training ablation. Where two structurally different models, trained on different (or differently-weighted) input signals, independently arrive at the same predicted module for a gene, that agreement is treated as meaningfully stronger evidence than either model's confidence score alone. This concordance approach is, in effect, a free ensemble-level sanity check layered on top of each model's own probability estimate.
It's worth being explicit that confidence scores between variants are not directly comparable in absolute terms. The fusion model's predictions carry systematically higher average confidence than the PPI-only model's, but the training results show the fusion model also has a substantially larger gap between its training and validation performance - a sign of mild overconfidence rather than genuinely better calibration. The PPI-only and a third, weaker NPP-only baseline produce lower but better-calibrated confidence estimates. This is mentioned here because it directly affects how the concordance results below should be read: agreement between models matters more than either model's raw confidence number in isolation.
A module's genome-wide predictions are only as trustworthy as that module was during training validation. Each of the 17 modules was assigned a reliability tier - high, medium, or low - based directly on its validation F1 score from the training ablation (full numbers in the training methodology document). Six modules reached high reliability (validation F1 ≥ 0.55); a few modules remained weak under every tested configuration and were assigned low reliability. Every genome-wide prediction inherits its module's tier as a built-in caveat: a confident-looking prediction in a low-reliability module is, by construction, less trustworthy than the same confidence score in a high-reliability one, and is treated accordingly in everything downstream.
Across the full unlabeled gene set, just over half of all genes (51.9%) received the same predicted module from at least two independently-evaluated model variants. A meaningfully smaller subset - 6.7% of all genes - received unanimous agreement across every variant tested. Both numbers matter for different reasons: the larger figure shows that the models are converging on a shared, non-random structure across most of the genome rather than producing scattered, uncorrelated guesses; the smaller figure identifies the genes with the strongest possible statistical support available from this analysis.
Among the different ways two models can agree, agreement between the PPI-based model and the NPP-based model is treated as the single strongest form of evidence, even though it is numerically the rarest (roughly 2% of genes). NPP and PPI are about as independent as two biological signals can be - one reflects evolutionary co-conservation across roughly 1,900 species, the other reflects direct or indirect physical interaction network topology - so when both agree on a module assignment despite drawing on essentially unrelated information, that convergence is unlikely to be a shared artifact of either method.
Restricting attention specifically to the six high-reliability modules narrows the candidate pool sharply: only a small fraction of the genome's concordant predictions fall within them. That narrowing is intentional rather than a limitation - the high-reliability modules are, by definition, the ones where the model's predictions have been shown to actually track ground truth during validation, so this smaller, more selective set is treated as the project's primary genome-wide output, with the much larger pool of medium- and especially low-reliability concordant calls treated as exploratory rather than reportable.
Model agreement and high confidence are necessary but not sufficient grounds for calling a gene a genuine IRD candidate. The project's working criteria for a convincing candidate require all of the following: agreement between independent model variants, membership in a high-reliability module, genuine (not imputed) PPI network coverage for that gene, and at least one independent line of supporting evidence external to the model itself - retinal or photoreceptor tissue expression data, genomic proximity to a known IRD locus, or existing literature suggestive of a retinal or ciliary function not yet formally annotated as IRD-related. "Genuine coverage" here means only that a gene cleared STRING's minimum connectivity threshold rather than falling back to imputation; it says nothing about how rich or complete that gene's documented network neighborhood actually is, a distinction discussed further below.
This biological sanity-checking pass against the high-reliability candidate pool is an active, ongoing part of the project rather than a finished step, and its results are intentionally not part of this document - this document describes the inference methodology and the statistical evidence it produces, not specific gene-level findings, which belong to the project's scientific output rather than its engineering documentation.
As a separate, qualitative check, a dimensionality-reduction projection (UMAP) of the trained three-tower embedding space was generated, with known training genes colored by their true module. The six high-reliability modules form visually distinct, well-separated clusters in this projection - consistent with their strong quantitative validation performance - and genes flagged as novel candidates predominantly embed within the region belonging to their predicted module rather than appearing scattered across the space. This is not formal evidence on its own, but it is a useful, intuitive cross-check that the model's internal representation lines up with its stated predictions.
A few limitations are worth stating directly rather than leaving implicit. A high-confidence, multi-model-concordant prediction is a statistically motivated hypothesis, not a validated biological finding - the gap between the two is exactly what the biological sanity-checking step above exists to close. Reliability tier reflects each module's measured validation performance, not a judgment about whether genes belonging to it are biologically interesting; a low-reliability module simply means the model has not yet demonstrated it can recover that module's genes correctly, which is a statement about the model and the available training data, not about the module's biology. Finally, roughly 7% of all genes genome-wide have no usable STRING network entry and rely on an imputed PPI vector; predictions for these specific genes should be weighted more toward the NPP- and ESM2-based models, where genuine signal exists, rather than the PPI-only model's score.
A more structural asymmetry is worth flagging on its own, and is discussed at length in the training methodology document: curated interaction databases like STRING reflect how much a gene has been studied as much as its true biological connectivity. The training genes are themselves well-characterized IRD genes and are, almost by construction, better-represented in PPI space than an average, arbitrarily-chosen gene. This means the PPI-based models' strong validation performance may not transfer with equal strength to the genuinely novel, poorly-characterized genes that genome-wide inference is specifically trying to surface - precisely because those genes are the ones least likely to have an extensively documented interaction neighborhood, independent of whether they are true IRD genes. A practical consequence: a moderate-confidence PPI-only prediction for an obscure, understudied gene may be just as scientifically interesting as a high-confidence one for a well-studied gene, and should not be discounted on confidence score alone during the biological sanity-checking step - if anything, it may deserve closer attention.
The next planned validation layer is independent tissue expression data - specifically from the Human Protein Atlas rather than GTEx, given GTEx's known gap in retinal tissue coverage - applied as a post-hoc check against the high-reliability candidate pool rather than as additional model input, preserving it as an external, independent line of evidence.