When a virus jumps from its natural reservoir into a new species, one decisive question is whether its proteins can switch off the new host's front-line immune alarms — the signaling cascades that detect viral DNA or RNA and call for reinforcements. We already have a screened library measuring how strongly roughly 800 viral proteins suppress human immune signaling (High-Content Screening and Computational Prediction Reveal Viral Genes That Suppress the Innate Immune Response), and we know that at least some viral suppressors work straight out of the box against immune proteins from humans, mice, and chickens (Unique and complementary suppression of cGAS-STING and RNA sensing- triggered innate immune responses by SARS-CoV-2 proteins). What we cannot yet do is predict, from sequence alone, how strongly a viral protein will suppress a given host's immune pathway. The suppression measurements themselves span orders of magnitude, and every existing viral-prediction benchmark sidesteps this by scoring models only on rank order, which is blind to everything this proposal cares about.
The core observation comes from an unexpected corner. The When and How of Target Variable Transformations showed that different regression algorithms respond oppositely to rescaling the prediction target: linear models improve by 11–63% median Relative Squared Error under quantile normalization, kernel regressors improve 48% under a uniform mapping, while gradient-boosted tree ensembles are indifferent or even hurt. Separately, Can Tabular Foundation Models Predict Algorithm Runtime Distributions? showed a tabular transformer that collapses on heavy-tailed targets unless they are log-scaled, and that caps out beyond 2,048 training examples. And in the viral world, Variant effect prediction with reliability estimation across priority viruses and ProteinGym: Large-Scale Benchmarks for Protein Fitness Prediction and Design both found that no single model family wins: ensembles of sequence, alignment, and structure models do. These three facts have never been put together, because the viral benchmarks never look at absolute-scale error, and the transformation study never left the UCI repository.
Core Claim
On heavily skewed, multi-order-of-magnitude viral bioactivity targets, a heterogeneous stacked ensemble in which each base learner is trained on the target transformation that matches its inductive bias — raw values for trees, log or Yeo-Johnson for linear and kernel models, log-scaled for the tabular transformer — will beat both any single model and any uniformly-transformed ensemble on original-scale error and uncertainty calibration. Calibration, in turn, is what makes the ensemble useful in the regime that actually matters experimentally: choosing the next viral protein to assay against a new host species, and doing so far more efficiently than random screening.
Concretely, the predictor takes a viral protein sequence plus a host pathway identifier (say, human STING versus chicken STING), and outputs the expected fold-suppression of that pathway, with an honest uncertainty interval. The aggregation step is:
$$\hat{y}(x) \;=\; \sum_{m=1}^{M} w_m(x)\; c_m\, f_m^{-1}\!\big(h_m(x)\big)$$
where $h_m$ is base learner $m$, $f_m$ is the target transformation chosen for that learner (the identity for trees, $\log(1+y)$ or Yeo-Johnson for linear and kernel models), $f_m^{-1}$ maps its prediction back to the original suppression scale, $c_m$ is an empirical back-transformation bias correction estimated from cross-fitted residuals (naive inversion of a convex transform like $\exp$ systematically underestimates the mean, by Jensen's inequality), and $w_m(x)$ are weights learned by a supervised meta-model, which Assessment of Ensemble Learning to Predict Wheat Grain Yield Based on UAV-Multispectral Reflectance showed can beat simple averaging by trusting different members in different input regions.
This differs from the closest prior work in one essential way. Variant effect prediction with reliability estimation across priority viruses built a hybrid ensemble (SaProt-EVE) for viral fitness, but scored it with rank correlations on single mutations, with no host in the input and no target scale in the loss; ProteinGym does the same across 2.7 million variants; Predicting high-fitness viral protein variants with Bayesian active learning and biophysics does use calibrated continuous predictions, but only for a single Gaussian process on spike-receptor binding, never for intracellular immune suppression and never with target transformations or model diversity. Nobody has asked how absolute-scale prediction of immune suppression behaves under transformation, because nobody has built the dataset where that question is answerable.
Research Program
Three components. First, repackage the high-content screening library of Ng et al. as a continuous regression dataset rather than a hit list: the per-cell nuclear-translocation readouts for IRF3, NF-κB, and pSTAT1 in cGAS-knockout fibroblasts become graded suppression scores for 800 viral genes, which will need to be re-derived from the imaging data since the published analysis thresholds them into 79 binary hits. Second, extend the target across species: using the luciferase-reporter design of Rui et al. as the assay template, measure suppression by a stratified subset of the library (roughly 150 proteins, enriched in the 79 hits plus matched non-hits) against STING orthologs from human, mouse, and chicken, yielding a train-on-human, test-on-divergent-host split that directly probes cross-species generalization. Third, build the ensemble: embeddings from protein language models and family-specific alignments as input features (the pairing that EVEREST showed is necessary because universal sequence models underperform on viruses), with base learners spanning trees, kernel ridge regression, a linear model, and the tabular transformer run under its 2,048-example context ceiling, each given its matched transformation, and a stacking layer on top. A smaller stretch goal uses the same machinery on two-protein combinations, since Rui et al. showed ORF3a and 3CL suppress complementary pathways and Identification of the key amino acid mutations in the PB2 and PA proteins of classical swine H1N1 influenza A virus in mammalian adaptation showed non-additive 11.2-fold synergy between proteins — non-additivity that no single-site benchmark can express.
Proposed Experiments
The primary evaluation is original-scale root mean squared error and mean absolute error on held-out suppression fold-changes, plus 90% interval coverage, with Spearman reported only as a secondary diagnostic. The decisive comparisons are the cross-species splits: train on human-ortholog suppression, test on mouse and chicken. Then a simulated active-learning campaign in the style of VIRAL: an upper-confidence-bound selector picks which of the unmeasured library genes to assay next, and we measure the enrichment of true high-suppressors among the picks relative to random selection; VIRAL's enrichment factor of 5 after ten acquisition rounds is the reference point for what a well-calibrated selector achieves on a different task, and the fivefold acceleration it reports over random sampling sets the plausible-effect scale. A dedicated experiment quantifies back-transformation bias: predictions from log-trained learners, naively exponentiated versus corrected, checked for systematic underestimation of the conditional mean. Finally, an epistasis probe on the small set of available two-protein measurements tests whether the ensemble's pairwise features predict non-additive combined suppression better than an additive baseline.
Baselines
The "do nothing" reference is predicting the training-set median suppression for every protein, which establishes the floor any model must clear. The strongest simple competitor is a gradient-boosted tree ensemble trained directly on raw suppression values: Nuyts and Davis showed tree ensembles are largely indifferent to target transformation, which makes this the natural cheap alternative — it requires no transformation machinery, no back-transformation correction, and no ensemble bookkeeping. It will be tuned with the same hyperparameter search budget and the same training folds as the full ensemble. The strongest existing method for supervised protein-phenotype regression is ProteinNPT, which reports a mean Spearman correlation of 0.613 on the ProteinGym supervised benchmark (ProteinGym); no reported result found on Lacuna for ProteinNPT on any immune-suppression or cross-species target, so we will run it ourselves on our dataset under the same budget. We also include a fine-tuned universal protein language model as a single-model baseline, since EVEREST found these systematically underperform alignment-based models on viral tasks — a claim our setting will re-test. The key ablation is the full ensemble with one shared transformation applied to every member (identity, log, and Yeo-Johnson variants), which isolates the per-learner transformation matching as the ingredient under test; a second ablation replaces the learned stacker with simple averaging, following the finding in A Unified Comparison of User Modeling Techniques for Predicting Data Interaction and Detecting Exploration Bias that averaging behaves as a smoother and can already be strong. Beating the tree baseline means lower original-scale error and better interval coverage on the cross-species splits; beating the shared-transform ablation means the gain comes specifically from matched transformations, not from ensemble size.
Why not just…?
Why not just use gradient-boosted trees on the raw targets? This is the most serious objection, and Nuyts and Davis give it real support: trees tolerated raw skewed targets on every dataset they tested. If trees match the full ensemble on original-scale error in our experiments, the proposal's contribution honestly narrows to the uncertainty and active-learning regime, where trees are weakest — they do not produce the calibrated predictive variance that an upper-confidence-bound selector needs, and VIRAL demonstrates that acquisition quality collapses without it. The tree baseline is therefore run in every experiment, not argued away; the cross-species splits and coverage metrics are designed to be exactly where a variance-calibrated, transformation-matched ensemble should earn its complexity.
Why not just use one large protein language model? The corpus answers this directly rather than by argument: EVEREST found universal sequence models underperform alignment-based models on viral proteins because viral diversity is filtered out of their training corpora, and ProteinGym found its best zero-shot model was itself a hybrid ensemble, TranceptEVE. Our ensemble deliberately includes both model types as members. That said, a fine-tuned single model might still win on a small 800-protein dataset where extra members overfit, which is why the single-model baseline is included with an equal tuning budget.
Why not just apply log(1+y) to everything? Uniform log scaling is the field's default instinct, and Ibrahimli and Adriaensen confirm it is essential for the tabular transformer member. But Nuyts and Davis found quantile-normal and Yeo-Johnson consistently beat the natural log for linear models, and support-vector regressors improved 48% only under a uniform mapping, not a log; meanwhile trees can be actively degraded by transformation. Forcing one transform on all members forfeits these gains and imposes back-transformation bias even on members that never needed it. The uniform-log ablation settles this empirically: if it matches per-learner matching, the simpler recipe wins and we will say so.
Why not just rank proteins and ignore the continuous scale? Ranking suffices if the only goal is a hit list, and it is exactly what let ProteinGym and EVEREST ignore transformations entirely, since Spearman is invariant to any monotone rescaling. But ProteinGym itself notes high rank correlation does not guarantee success at prioritizing the top variants, and an acquisition selector like VIRAL's needs calibrated values and variances, not orderings, to decide what to assay next against a new host species. Rank-only evaluation would make the active-learning experiment unrunnable, which is the point.
Decisive Checks
Three results would decide the proposal. First, on the cross-species splits, the transformation-matched ensemble must beat the raw-target tree ensemble and the shared-transform ablation on original-scale error at matched compute and tuning budget; if the tree ensemble wins on error, the fallback claim must survive on 90% interval coverage. Second, the simulated acquisition campaign must enrich true cross-species suppressors over random selection by a margin attributable to calibration — measured by comparing the matched ensemble's selector against a selector driven by the tree ensemble with bootstrap variance. Third, the back-transformation correction must measurably remove the Jensen bias in exponentiated predictions; if the residual bias persists, that is itself a publishable negative constraint on log-space training for bioactivity targets. A failure on the first two checks would reduce this work to a benchmark paper, but one that still fills a real gap: no dataset currently exists where continuous, cross-species immune-suppression prediction can be measured at all.
References
- High-Content Screening and Computational Prediction Reveal Viral Genes That Suppress the Innate Immune Response
- Unique and complementary suppression of cGAS-STING and RNA sensing- triggered innate immune responses by SARS-CoV-2 proteins
- The When and How of Target Variable Transformations
- Can Tabular Foundation Models Predict Algorithm Runtime Distributions?
- Variant effect prediction with reliability estimation across priority viruses
- ProteinGym: Large-Scale Benchmarks for Protein Fitness Prediction and Design
- Assessment of Ensemble Learning to Predict Wheat Grain Yield Based on UAV-Multispectral Reflectance
- Predicting high-fitness viral protein variants with Bayesian active learning and biophysics
- Identification of the key amino acid mutations in the PB2 and PA proteins of classical swine H1N1 influenza A virus in mammalian adaptation
- A Unified Comparison of User Modeling Techniques for Predicting Data Interaction and Detecting Exploration Bias