Wild-Type to Modified Generalization

• The paper demonstrates that docking-based models can outperform docking-free methods by better capturing modification-specific effects in protein–ligand affinity prediction.
• Wild-Type to Modification Generalization is defined as transferring from unmodified (wild-type) training data to structurally or statistically altered variants in diverse domains.
• Improving training data diversity through augmentation, targeted masking, and domain-shift techniques boosts out-of-domain transfer performance and model robustness.

Searching arXiv for the cited papers and topic-specific benchmarks to ground the article in current literature. {"query":"id:(Gokhale et al., 2022) OR id:(Wu et al., 30 Nov 2025) OR id:(Chen et al., 5 Jun 2026) OR id:(Kmicikiewicz et al., 28 May 2025)", "max_results": 10} Wild-Type to Modification Generalization denotes a family of evaluation settings in which a model is trained on an unmodified reference regime and then tested on altered variants that were not observed during training. In the protein–ligand affinity benchmark introduced with DAVIS-complete, the reference regime is the set of wild-type kinase–ligand pairs and the target regime is the set of modified kinase variants; related work studies analogous source-to-shift transfer under deliberate data modification, proxy-to-wild transfer in deepfake speech, and active exploration beyond wild-type sequence neighborhoods in protein design (Wu et al., 30 Nov 2025, Gokhale et al., 2022, Chen et al., 5 Jun 2026, Kmicikiewicz et al., 28 May 2025). Across these settings, the central question is whether models learn modification-specific effects or merely interpolate within the support of the reference distribution.

1. Conceptual scope

The most specific use of the term arises in modification-aware drug discovery benchmarks. There, the problem is stated directly: if a model is trained only on wild-type kinase–ligand pairs, can it predict binding affinity for modified kinases it has never seen before? The motivation is biological realism, because kinases are frequently altered by substitutions, insertions, deletions, phosphorylation, and combinations of these events, and such changes can alter binding affinity, drug resistance, and selectivity (Wu et al., 30 Nov 2025).

A broader empirical formulation appears in work on data modification under distribution shift. That line of work calls the original data the source distribution and the shifted evaluation sets the target / out-of-domain (OOD) distributions, and studies whether modifying the training distribution improves transfer to altered or unseen inputs while also affecting adversarial robustness (Gokhale et al., 2022). An analogous transfer problem appears in deepfake speech, where codec resynthesized speech functions as a proxy training distribution and must generalize to more varied in-the-wild codec-based speech generation systems (Chen et al., 5 Jun 2026). In protein design, the same logic appears as the problem of moving beyond a wild-type neighborhood without losing biological plausibility or fitness (Kmicikiewicz et al., 28 May 2025).

Setting	Reference regime	Modified or shifted regime
DAVIS-complete affinity benchmark	wild-type kinase–ligand pairs	modified kinase variants with substitutions, insertions, deletions, phosphorylation, or combinations
Data-modification under distribution shift	source distribution $\mathcal{D}$	unseen target distributions $\mathcal{D}_e$
CodecFake detection	CoRS proxy data	in-the-wild CoSG systems, including unseen generative models and long-form audio
Protein design	wild-type / starting sequence neighborhood	more heavily modified, novel sequences

This suggests a useful unifying interpretation: wild-type-to-modification generalization is not limited to one modality, but denotes transfer from a reference distribution to structurally or statistically altered variants whose differences are consequential for prediction.

2. Formal problem statements

In the DAVIS-complete formulation, wild-type proteins are denoted by

$P^w = \{ p^{w_i}\},$

modified proteins by

$P^m = \{ p^{m_i} \},$

with multiple modified variants

$p^{m_i} = \{ p^{m_i}_j \},$

and the full protein set by

$P^* = P^w \cup P^m.$

Ligands are

$L = \{ l_k \},$

and affinity is written as $A(p,l)$. Modification-induced affinity shift is defined as

$\Delta pK_d = A(p^{m_i}_j, l_k) - A(p^{w_i}, l_k).$

The benchmark’s strict transfer setting is

$P^wL \rightarrow P^mL,$

with training only on wild-type protein–ligand pairs and testing only on modified protein–ligand pairs (Wu et al., 30 Nov 2025).

In the distribution-shift formulation, OOD generalization “expects a model which is trained on distribution $\mathcal{D}_e$0 to perform reliably on unseen distributions $\mathcal{D}_e$1, that differ from $\mathcal{D}_e$2.” Adversarial robustness is defined through a worst-case perturbation

$\mathcal{D}_e$3

so the same model is evaluated on in-domain accuracy, OOD accuracy, and robustness to small perturbations (Gokhale et al., 2022).

In deepfake speech, DSFA makes the source-to-shift problem explicit in latent feature space. For feature map

$\mathcal{D}_e$4

instance-level channel statistics are

$\mathcal{D}_e$5

and batch-level variability is interpreted as domain uncertainty. Perturbed statistics are then sampled and applied through an AdaIN-style transformation,

$\mathcal{D}_e$6

to simulate plausible wild-domain shifts during fine-tuning (Chen et al., 5 Jun 2026).

In ProSpero, the transition beyond a wild-type neighborhood is formalized through a target distribution

$\mathcal{D}_e$7

which combines a surrogate score with a frozen pretrained sequence prior. This encodes the idea that exploration should favor both predicted fitness and biological plausibility (Kmicikiewicz et al., 28 May 2025).

3. Modification-aware affinity prediction in DAVIS-complete

DAVIS-complete introduces Wild-Type to Modification Generalization as one of three benchmark settings, alongside Augmented Dataset Prediction and Few-Shot Modification Generalization. The dataset extends the original DAVIS benchmark of 442 kinases × 72 ligands = 31,824 affinity measurements by manually curating 56 modified amino-acid sequences for 11 kinase proteins, generating 4,032 new modified protein–ligand pairs. The curated variants include substitutions such as T315I, L858R, and V600E; insertions such as FLT3 ITD; deletions such as EGFR E746-A750del; phosphorylation; and combined modifications such as mutation plus phosphorylation (Wu et al., 30 Nov 2025).

The benchmark contains three evaluation scenarios. The global modification setting tests all modified pairs through $\mathcal{D}_e$8. It is further stratified into four cases according to whether WT and modified affinities are capped or uncapped: WT-uncapped / modification-uncapped, WT-capped / modification-uncapped, WT-uncapped / modification-capped, and WT-capped / modification-capped. A second scenario fixes a ligand and varies the modifications of the same kinase, testing whether a model can distinguish how different modifications of the same protein change binding to the same drug. A third scenario fixes a modified kinase and varies ligands, testing whether the model can rank ligands correctly for a mutant target (Wu et al., 30 Nov 2025).

The benchmark evaluates docking-free models—DeepDTA, AttentionDTA, GraphDTA, DGraphDTA, and MGraphDTA—and docking-based models—FDA and Boltz-2—using MSE, Pearson correlation coefficient $\mathcal{D}_e$9, and C-index. Two WT reference baselines are critical: reusing the measured WT affinity for the modified pair, and reusing the model’s WT prediction for the modified pair. These baselines reveal whether a model captures modification-specific effects or merely propagates WT signal (Wu et al., 30 Nov 2025).

The main empirical result is that docking-based models generalize better in zero-shot settings, especially on structurally novel cases, whereas docking-free models often overfit to wild-type proteins and struggle with unseen modifications. On the complete test set and especially the modification subset, however, Boltz-2 is not always the top performer; in some settings DeepDTA can outperform it. In the most diagnostic same-ligand/different-modifications scenario, docking-free models often have low $P^w = \{ p^{w_i}\},$0, sometimes below 0.3, C-index only slightly above random, around 0.5, and performance often not meaningfully better than WT baselines. The paper states that many docking-free models are essentially predicting values very close to the WT prediction rather than capturing modification-specific changes. In the same-modification/different-ligands setting, performance improves for all models, but in 44 out of 55 cases wild-type and modified affinity profiles remain highly consistent, with $P^w = \{ p^{w_i}\},$1, indicating that ligand effects can dominate modification effects (Wu et al., 30 Nov 2025).

A concrete quantitative subset is the global WT-uncapped / modification-uncapped case. There, DeepDTA reports MSE 0.63, $P^w = \{ p^{w_i}\},$2, C-index 0.79; MGraphDTA reports MSE 0.61, $P^w = \{ p^{w_i}\},$3, C-index 0.80; FDA reports MSE 1.47, $P^w = \{ p^{w_i}\},$4, C-index 0.72; and Boltz-2 reports MSE 0.83, $P^w = \{ p^{w_i}\},$5, C-index 0.77. The paper also reports that the mean $P^w = \{ p^{w_i}\},$6 across modified pairs is about $P^w = \{ p^{w_i}\},$7, while noting that DAVIS is heavily censored at $P^w = \{ p^{w_i}\},$8, so many changes are not exactly observable (Wu et al., 30 Nov 2025).

4. Data modification, OOD transfer, and adversarial robustness

A related empirical literature asks what happens when the training distribution itself is deliberately modified. The comparison in "Generalized but not Robust? Comparing the Effects of Data Modification Methods on Out-of-Domain Generalization and Adversarial Robustness" studies four classes of intervention: additional training datasets / multi-source training (MS), data augmentation (DA), model debiasing (DB), and dataset filtering (DF). The central claim is that methods intended to improve OOD generalization often also change adversarial robustness, so both should be evaluated rather than only one (Gokhale et al., 2022).

MS broadens empirical support by adding more real examples from other datasets for the same task. In NLP, the paper uses SNLI+MNLI for NLI and Natural Questions plus other MRQA datasets for QA; in vision, it uses MNIST+USPS. Across tasks, MS improves average OOD accuracy on all three tasks and also improves robustness on NLI and QA. DA synthetically creates label-preserving variants; the paper uses EDA for text and M-ADA for vision. DA generally helps, but the effect is task-dependent: in NLI, EDA gives small gains in average OOD accuracy and improves robustness; in QA, EDA slightly improves in-domain accuracy but does not consistently improve OOD accuracy; in vision, DA improves OOD accuracy on MNIST-M, SVHN, and SYNTH and improves adversarial robustness, with M-ADA doing especially well. DB modifies the modeling pipeline so that the model learns to ignore spurious cues; the paper uses LMH for NLI, Mb-CR for QA, and RandConv for vision. DB improves OOD accuracy on three tasks and robustness on NLI and QA, and RandConv is particularly strong on digits. DF, implemented with AFLite, removes “easy” or biased examples. It is the most inconsistent strategy: it improves OOD generalization on SNLI and is the best method for average NLI OOD accuracy, but for QA it dramatically reduces both in-domain and OOD exact match, and for digits it catastrophically lowers OOD accuracy on all target datasets and also lowers in-domain accuracy. It is also the only method that consistently degrades adversarial robustness across all three benchmarks (Gokhale et al., 2022).

The paper’s empirical synthesis is explicit: MS increases OOD accuracy on all three tasks and robustness on two tasks; DA increases OOD on two tasks and robustness on all three tasks; DB increases OOD on three tasks and robustness on two tasks; DF decreases OOD on two tasks and robustness on all three tasks. On digits, training-size sweeps from 10% to 100% show that OOD accuracy rises with dataset size and that OOD accuracy and adversarial robustness are positively correlated across these sweeps (Gokhale et al., 2022).

The toy 2D concentric-circles example makes the mechanism visible. The source data contain biased clusters, with class 0 overrepresented at $P^w = \{ p^{w_i}\},$9 and class 1 at $P^m = \{ p^{m_i} \},$0, and 20% of the 10,000 training samples biased. MS and Gaussian DA with $P^m = \{ p^{m_i} \},$1 make the training distribution more diverse, while AFLite-style filtering down to 10% of the data removes whole sectors of the distribution, including the biased clusters. A linear SGD classifier evaluated on OOD data and PGD attacks exhibits the same pattern as the larger experiments: adding diversity helps both OOD generalization and robustness, while filtering can sharply damage both (Gokhale et al., 2022).

5. Proxy-to-wild transfer in deepfake speech

An analogous modification-aware transfer problem appears in CodecFake detection, where training commonly uses codec resynthesized speech as a proxy and testing targets in-the-wild codec-based speech generation. The paper "Mitigating Proxy-to-Wild Domain Gap in Deepfake Speech" identifies three axes of proxy-to-wild domain gap: artifact mismatch, silence mismatch, and content and speaker mismatch. The argument is that CoRS helps because it exposes models to codec reconstruction artifacts, but also creates a new overfitting risk because it is generated from a limited set of codecs and a fixed corpus, allowing a detector to learn codec-specific or dataset-specific quirks that do not transfer to unseen codecs, new generative systems, new decoding and tokenization schemes, long-form audio, or more diverse acoustic and linguistic conditions (Chen et al., 5 Jun 2026).

The proposed remedy combines a post-trained SSL backbone with Domain-Shift Feature Augmentation. The backbone is post-trained Wav2Vec2-Large-AntiDeepfake. DSFA transforms deterministic latent statistics into stochastic domain-shift distributions by estimating mini-batch variability in feature means and standard deviations, sampling perturbed statistics with either Uniform or Gaussian noise, and applying them through an AdaIN-style reconstruction. DSFA is applied stochastically with probability $P^m = \{ p^{m_i} \},$2, so the model alternates between original and shifted features during fine-tuning. The training objective is

$P^m = \{ p^{m_i} \},$3

although the experiments report that supervised contrastive loss can help on CoSG Eval but may hurt transfer to CoSG ExtEval, while DSFA-only fine-tuning is the more robust configuration (Chen et al., 5 Jun 2026).

Evaluation uses CodecFake+ for training, with CoRS as training spoof data and CoSG as evaluation spoof data. CoSG Eval contains 17 codec-based generation models. CoSG ExtEval extends this with 40 unseen generative models and long-form audio up to about 149 s. The experimental setup uses raw waveforms at 16 kHz, 4 seconds segment length during training, RawBoost, Adam, learning rate 1e-6, weight decay 1e-4, batch size 14, CE weights (0.1, 0.9), and EER (%) as the metric (Chen et al., 5 Jun 2026).

The post-trained SSL model alone yields CoSG Eval: EER around 3.95% and CoSG ExtEval: EER around 22.19%. With DSFA, the best reported configuration improves to about 2.78% on CoSG Eval and about 21.80% on CoSG ExtEval. Ablation shows that $P^m = \{ p^{m_i} \},$4 is the worst ExtEval configuration, the best ExtEval result occurs around $P^m = \{ p^{m_i} \},$5, and too much augmentation as $P^m = \{ p^{m_i} \},$6 hurts performance. The best DSFA insertion layer depends on perturbation distribution: Uniform is best around Layer 24, whereas Gaussian is best around Layer 1. Feature-statistics visualization reports improved overlap between source and target distributions: mean overlap increases from 42.91% to 43.03%, and standard-deviation overlap from 65.01% to 67.09% (Chen et al., 5 Jun 2026).

This suggests that wild-type-to-modification generalization has a close analogue in proxy-to-wild generalization: the relevant failure mode is overfitting to a narrow reference distribution, and the relevant remedy is to train against a family of plausible shifts rather than a single static source.

6. Exploration beyond wild-type neighborhoods in protein design

In protein engineering, wild-type-to-modification generalization appears as a design rather than prediction problem. ProSpero addresses the task of finding sequences with both high fitness and novelty under a limited query budget, while starting from a wild-type / starting sequence and exploring beyond its neighborhood without leaving the biological manifold. The method uses a frozen pretrained generative model, EvoDiff-OADM (38M parameters), as a sequence prior, and an iteratively retrained surrogate to guide acquisition from oracle feedback (Kmicikiewicz et al., 28 May 2025).

The active-learning loop alternates between fitting a surrogate, choosing the current best sequence, identifying editable residues by targeted masking, generating candidates through biologically constrained Sequential Monte Carlo, querying the oracle on the top candidates, and updating the dataset. The surrogate is not used to generate sequences directly; instead, it guides inference-time sampling from the frozen generator. This division is central to the method’s claim that it can explore beyond wild-type neighborhoods while preserving biological plausibility (Kmicikiewicz et al., 28 May 2025).

Targeted masking acts as a sequence-space analogue of alanine scanning. The method creates partially alanine-mutated sequences, scores them with predictive mean and uncertainty, and uses UCB with $P^m = \{ p^{m_i} \},$7 for masking and $P^m = \{ p^{m_i} \},$8 for the SMC phase to choose mutations that appear informative and tolerable. Biologically constrained SMC then fills masked positions under a charge-class restriction. Masked sites are partitioned into positive, negative, and neutral classes according to the wild-type residue, and sampling is restricted to amino acids in the same charge class. The constrained prior is denoted $P^m = \{ p^{m_i} \},$9. This is the explicit mechanism that keeps exploration biologically plausible even when the surrogate is misspecified (Kmicikiewicz et al., 28 May 2025).

The empirical results indicate that ProSpero is best on 5/8 tasks, second-best on the other 3/8, and improves fitness over wild-type on every task. It ranks first in novelty on 6/8 tasks and second on 1/8, often achieving roughly 2–9× greater novelty than PEX while matching or exceeding its fitness. The benchmark tasks are AMIE, TEM, E4B, Pab1, AAV, GFP, UBE2I, and LGK. Biological plausibility is assessed through validity defined from physicochemical properties, and structural quality is supported by pTM and pLDDT consistently above 70 together with scPerplexity consistent with plausible foldability. On LGK, candidates can differ from wild type by over 70 amino acids while maintaining structural confidence comparable to wild-type (Kmicikiewicz et al., 28 May 2025).

Ablations show that the full method outperforms variants without SMC, without targeted masking, without both, and without charge-class restriction. The paper’s interpretation is that the charge-based biological prior is a major contributor to robustness under surrogate misspecification, especially at low SNR, while targeted masking and SMC become more effective as surrogate signal improves. In this setting, wild-type-to-modification generalization is therefore not merely larger mutational distance; it is structured exploration that preserves biochemical compatibility while moving into novel regions of sequence space (Kmicikiewicz et al., 28 May 2025).

7. Recurring empirical patterns and unresolved issues

Across these literatures, wild-type-to-modification generalization repeatedly exposes a distinction between models that learn modification-sensitive structure and models that merely reuse reference-distribution signal. In DAVIS-complete, this appears as the competitiveness of WT baselines and the observation that some docking-free models produce modified-pair predictions nearly identical to WT predictions. In the data-modification study, it appears as the contrast between methods that broaden support and methods that prune it. In deepfake speech, it appears as overfitting to proxy codec artifacts. In protein design, it appears as the tension between local plausibility and broader exploration (Wu et al., 30 Nov 2025, Gokhale et al., 2022, Chen et al., 5 Jun 2026, Kmicikiewicz et al., 28 May 2025).

A consistent positive pattern is that broader or better-structured coverage of relevant variation tends to help. MS, DA, and DB improve OOD generalization more reliably than DF in the distribution-shift study. DSFA narrows the proxy-to-wild gap by sampling latent shifts rather than trusting a fixed proxy distribution. ProSpero combines a frozen biological prior with constrained exploration to move beyond wild-type neighborhoods without degenerating into implausible sequences. In DAVIS-complete, docking-based models generalize better in zero-shot settings, and docking-free models improve notably when fine-tuned on a small set of modified examples in the related few-shot benchmark (Gokhale et al., 2022, Chen et al., 5 Jun 2026, Kmicikiewicz et al., 28 May 2025, Wu et al., 30 Nov 2025).

The principal caveats are equally consistent. Filtering can look strong on a benchmark with a particular shortcut, such as SNLI/HANS, yet fail badly on QA and digits. Docking-based gains in modification-aware affinity prediction may be inflated somewhat by data leakage from structure-prediction training sources. CoSG ExtEval remains difficult even with DSFA, and the best settings are sensitive to augmentation probability, noise distribution, and insertion layer. In DAVIS-complete, the heavy censoring at $p^{m_i} = \{ p^{m_i}_j \},$0 hides many true affinity shifts. These limitations show that the core problem is not solved by any single architecture or benchmark split (Gokhale et al., 2022, Wu et al., 30 Nov 2025, Chen et al., 5 Jun 2026).

The resulting picture is specific rather than generic. Wild-type-to-modification generalization is strongest when the training procedure adds coverage of biologically or statistically relevant variation, or imposes invariances that align with the actual shift, while preserving task signal. It is weakest when the model overfits to the wild-type or proxy regime, or when training-data curation removes modes that later become necessary for transfer.