A General Harness for Protein Foundation Model Fitness Prediction

Yang Tan; Qijia Tian; Gangyu Sun; Bozitao Zhong; Mingchen Li; Yuanxi Yu; Nanqing Dong; Liang Hong

cs.AISep 28, 2026

A General Harness for Protein Foundation Model Fitness Prediction

Authors: Yang Tan, Qijia Tian, Gangyu Sun, Bozitao Zhong, Mingchen Li, Yuanxi Yu, Nanqing Dong, Liang Hong

Organizations: Shanghai Jiao Tong University · Shanghai Innovation Institute · Shanghai Artificial Intelligence Laboratory

Abstract

Accurate fitness prediction is central to protein engineering and understanding sequence-function relationships. With advances in deep learning, protein foundation models (PFMs) have become widely used for this task. Recent analyses, however, show that these models share preferences reflecting their training corpora, while unreliable inputs can further distort fitness predictions. Family-specific evolutionary evidence and structural context can help address these limitations by providing complementary constraints on model scores, motivating VenusREM-Harness (VRH), a general, model-agnostic, training-free Retrieval-Enhanced Mutation harness. It fuses frozen model scores with multiple sequence alignment (MSA) evidence according to model uncertainty, then applies gated background correction and score shrinkage based on structural confidence and solvent exposure. Across 1,211 assays and 3.1 million measured variants from ProteinGym, VenusMutHub, and the newly curated viral benchmark VenusViroHub, all 71 configurations improve Spearman correlation on all 3 benchmarks by 0.073 on average, with broad gains across 5 metrics. Extended analyses relate retrieval gains to model-MSA preference differences, assess domain-level gains and immune-escape cases, and quantify computational speedups. Built with VRH, VenusREM2 is the first to rank highest in all function, taxon, MSA-depth, and mutation-depth categories, with a ProteinGym Average Spearman of 0.556, 0.038 above the prior best.

Figures & tables

Figure 1: Shared protein-wide amino-acid background preferences across representative models. The 49-model consensus closely follows UniProtKB/Swiss-Prot amino-acid abundance (Pearson $r=0.99$ ).

Figure 2: Framework of VenusREM-Harness . (a) Frozen PFM scores are refined using family evidence and structural context. (b) Adaptive MSA fusion adjusts the family contribution to model uncertainty and model–MSA agreement. (c) Coherence-gated calibration corrects model background bias while preserving family evidence. (d) Solvent exposure and structural confidence guide attenuation at each site before substitution contributions are aggregated into a variant score.

Figure 3: ProteinGym Spearman overall and across 17 strata (Appendix Tables 14 – 18 ). Gray: strongest comparator (SOTA Model); blue hatching and labels: VenusREM2 gains. Vertical labels identify comparators. Avg.: overall; Act./Bind./Expr./Org./Stab.: activity/binding/expression/organismal fitness/stability; Hum./Euk./Prok./Vir.: human/other eukaryotes/prokaryotes/viruses; L/M/H: low/medium/high MSA depth. Mutation depth counts substitutions per variant (

5+

: at least five). AIDO-RAG and PoET denote AIDO Protein-RAG (16B) and PoET (200M); AIDO-RAG and VenusREM tie overall at reported precision.

Figure 4: Raw– VRH results across five metrics on ProteinGym, VenusMutHub, and VenusViroHub (top to bottom). Points represent 71 configurations; those above the diagonal indicate gains.

Component	Setting	PG Validation (21)		PG Full (217)		VMH Full (905)		VVH (89)
Component	Setting	Score	$\Delta_{\mathrm{Raw}}$	Score	$\Delta_{\mathrm{Raw}}$	Score	$\Delta_{\mathrm{Raw}}$	Score	$\Delta_{\mathrm{Raw}}$
Raw	Native field	0.5573	—	0.5069	—	0.2126	—	0.2709	—
MSA fusion	w/o Ent. Weight.	0.5742	+0.017	0.5205	+0.014	0.2341	+0.022	0.2897	+0.019
MSA fusion	w/ Ent. Weight.	0.5774	+0.020	0.5246	+0.018	0.2336	+0.021	0.2901	+0.019
Model calibration	w/o Coh. Gating	0.5641	+0.007	0.5219	+0.015	0.2326	+0.020	0.2841	+0.013
Model calibration	w/ Coh. Gating	0.5788	+0.022	0.5284	+0.022	0.2369	+0.024	0.2942	+0.023

Table 1: ProSST -2048 ablation. PG (ProteinGym) validation/VMH (VenusMutHub): assay-macro; PG full: official Average Spearman; VVH (VenusViroHub): mutant-only phenotype–protein hierarchical Spearman.

\Delta_{\mathrm{Raw}}

reports absolute score changes (three decimals). w/o Ent. Weight. fixes

\alpha=0.8

; w/o Coh. Gating fixes

\kappa=1

, while w/ Coh. Gating uses

\kappa(R)

Figure 5: Separation and retrieval gain (

148

ProteinGym proteins). Red: quartile means and 95% bootstrap intervals; faint: proteins.

Figure 6: (a) Raw/ VRH distributions by phenotype across $71$ configurations (medians, IQRs, $1.5\times$ IQR whiskers, and outliers). Numbers indicate mean gains; red marks unresolved cases. (b) Measured escape and predicted constraint for SARS-CoV-2 XBB.1.5 RBD and ZIKV MR766 E; colors show within-panel percentiles and gray marks missing sites. (c) Parkin N232 and R396 in structural context. (d) Domain-wise Spearman. (e) N232A and R396H approach measured ranks.

Figure 7: Post-inference runtime on ProteinGym (

2.47

million variants): loop-based versus amortized lookup and MSA construction.

Appendix figures & tables68 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 8: Input distributions for ProteinGym (217 assays), VenusMutHub (905), and VenusViroHub (89). (a) Sequence length. (b) MSA depth. (c) MSA occupancy. (d) Residue pLDDT. (e) Residue RSA. Structure panels pool unique WT sequences; all selected structures fully cover the query.

Dataset	$N$ assays	Sequence length	MSA depth	MSA occupancy	Mean pLDDT	Mean RSA
ProteinGym	217	397 $\pm$ 502	8,650 $\pm$ 6,713	0.67 $\pm$ 0.23	86.6 $\pm$ 10.8	0.37 $\pm$ 0.09
VenusMutHub	905	288 $\pm$ 239	9,791 $\pm$ 6,452	0.69 $\pm$ 0.18	92.1 $\pm$ 7.5	0.32 $\pm$ 0.08
VenusViroHub	89	559 $\pm$ 383	1,217 $\pm$ 2,610	0.62 $\pm$ 0.14	83.9 $\pm$ 9.8	0.33 $\pm$ 0.07

Appendix

Table 2: Assay-level input statistics for ProteinGym, VenusMutHub, and VenusViroHub, reported as mean

\pm

standard deviation.

N

is the number of assays; MSA occupancy is the fraction of non-gap match-state entries; pLDDT and RSA are assay-level means.

Dataset	Unique WT	Structures	pLDDT $<50$	50–70	70–90	$\geq 90$
ProteinGym	187	187	18.9%	9.0%	22.6%	49.5%
VenusMutHub	527	527	6.7%	3.3%	10.6%	79.5%
VenusViroHub	52	52	9.7%	13.2%	36.0%	41.1%

Appendix

Table 3: AlphaFold2 confidence strata over unique wild-type sequences. All selected structures cover the full query; residue percentages are pooled within each benchmark.

Dataset	RSA $<0.2$	RSA $0.2$ – $0.5$	RSA $\geq 0.5$
ProteinGym	36.7%	29.4%	33.9%
VenusMutHub	45.8%	28.9%	25.3%
VenusViroHub	40.6%	30.2%	29.2%

Appendix

Table 4: RSA strata over unique wild-type sequences. Residue percentages are pooled within each benchmark.

Harmonised phenotype	Assays	Evaluated substitutions
Immune escape	21	148,468
Cell entry	18	189,048
Binding	17	90,245
Expression	15	58,120
Fitness	15	114,287
Stability	2	11,187

Appendix

Table 5: Phenotype composition of VenusViroHub after quality control. Evaluated substitutions exclude wild-type control rows.

Virus and construct	Source
HIV-1 BF520 Env	( Radford & Bloom, 2025 )
HIV-1 TRO11 Env	( Radford & Bloom, 2025 )
Influenza A/H3N2 HK19 HA	( Welsh et al., 2024 )
Influenza A/H3N2 MC22 HA	( Yu et al., 2025 )
Lassa virus Josiah GP	( Carr et al., 2024 )
Nipah virus F	( Larsen et al., 2026 )

Appendix

Table 6: VenusViroHub assays with experimental readout Monoclonal-antibody escape (15 assays). Distinct rows from the same study correspond to different viral constructs or measured phenotypes. † Public dataset repository without an associated archival article.

Virus and construct	Source
HIV-1 BF520 Env	( Radford et al., 2023 )
Human coronavirus 229E spike	( Harari et al., 2026 )
Influenza A/H5N1 HA	( Dadonaite et al., 2024a )
Lassa virus Josiah GP	( Carr & Bloom, 2026 ) †
Respiratory syncytial virus A F	( Simonich et al., 2026 )
SARS-CoV-2 XBB.1.5 RBD	( Dadonaite et al., 2024b )

Appendix

Table 7: VenusViroHub assays with experimental readout Serum-antibody escape (6 assays).

Virus and construct	Source
Chikungunya virus E3/E2/E1	( Ju et al., 2025 )
Chikungunya virus E3/E2/E1	( Ju et al., 2025 )
HIV-1 BF520 Env	( Radford & Bloom, 2025 )
HIV-1 TRO11 Env	( Radford & Bloom, 2025 )
Human coronavirus 229E spike	( Harari et al., 2026 )
Influenza A/H3N2 MC22 HA	( Yu et al., 2025 )

Appendix

Table 8: VenusViroHub assays with experimental readout Cell entry (18 assays).

Virus and construct	Source
Chikungunya virus E3/E2/E1	( Ju et al., 2025 )
Human coronavirus 229E spike	( Harari et al., 2026 )
Nipah virus Malaysia RBP	( Larsen et al., 2025 )
SARS-CoV-2 Alpha RBD	( Starr et al., 2022a )
SARS-CoV-2 Beta RBD	( Starr et al., 2022a )
SARS-CoV-2 Delta RBD	( Starr et al., 2022a )

Appendix

Table 9: VenusViroHub assays with experimental readout Receptor binding (17 assays).

Virus and construct	Source
Bat coronavirus PRD0038 RBD	( Starr et al., 2022c )
Bat coronavirus RmYN02 RBD	( Starr et al., 2022c )
Bat coronavirus RsYN04 RBD	( Starr et al., 2022c )
SARS-CoV-2 PLpro	( Wu et al., 2024 )
SARS-CoV-2 Alpha RBD	( Starr et al., 2022a )
SARS-CoV-2 Beta RBD	( Starr et al., 2022a )

Appendix

Table 10: VenusViroHub assays with experimental readout Protein expression and abundance (15 assays).

Virus and construct	Source
Coxsackievirus B3 2A	( Álvarez-Rodríguez et al., 2024 )
Coxsackievirus B3 2B	( Álvarez-Rodríguez et al., 2024 )
Coxsackievirus B3 2C	( Álvarez-Rodríguez et al., 2024 )
Coxsackievirus B3 3A	( Álvarez-Rodríguez et al., 2024 )
Coxsackievirus B3 3B	( Álvarez-Rodríguez et al., 2024 )
Coxsackievirus B3 3C	( Álvarez-Rodríguez et al., 2024 )

Appendix

Table 11: VenusViroHub assays with experimental readout Viral fitness and growth (15 assays).

Virus and construct	Source
Influenza A/H3N2 MC22 HA	( Yu et al., 2025 )
Influenza A/H5N1 HA	( Dadonaite et al., 2024a )
SARS-CoV-2 PLpro	( Wu et al., 2024 )

Appendix

Table 12: VenusViroHub assays with experimental readout Protein stability and protease activity (3 assays).

Model	Family	Evaluated variants	Readout	Source
ESM-1b	Masked sequence LM	650M	mask, wt	Official repository
ESM-1v	Masked sequence LM	650M	mask, wt	Official repository
ESM-2	Masked sequence LM	8M/35M/150M/650M/3B	mask, wt	Official repository
ESMC	Masked sequence LM	300M, 600M	mask	Official repository
CARP	Masked sequence LM	600K/38M/76M/640M	wt	Official repository
S3F	Structure-aware	S3F	wt, mask	Official repository

Appendix

Table 13: Evaluated baselines, native readouts, and implementations. Readouts follow Equation 1 ; mask

=

masked marginal, wt

=

wild-type field, and tf

=

teacher forcing.

Rank	Model	Seq	Str	Evo	Average	Activity	Binding	Expression	Organismal	Stability
—	VenusREM2 (ours)	✓	✓	✓	0.556	0.541	0.495	0.557	0.494	0.691
1	AIDO Protein-RAG (16B) ( Sun et al., 2024 )	—	✓	✓	0.518	0.517	0.426	0.522	0.491	0.635
2	VenusREM ( Tan et al., 2025b )	✓	✓	✓	0.518	0.495	0.454	0.533	0.459	0.650
3	ProSST ( $K{=}2048$ ) ( Li et al., 2024 )	✓	✓	—	0.507	0.476	0.445	0.530	0.431	0.653
5	S3F-MSA ( Zhang et al., 2024 )	—	✓	✓	0.496	0.502	0.440	0.479	0.477	0.581
8	Protriever ( Weitzman et al., 2025 )	—	—	✓	0.479	0.487	0.396	0.496	0.479	0.537

Appendix

Table 14: ProteinGym zero-shot substitution leaderboard: official 217-assay Average Spearman. Bold red, purple, and black mark the top 3 distinct scores per column at reported precision.

Rank	Model	Seq	Str	Evo	Average	Activity	Binding	Expression	Organismal	Stability
—	VenusREM2 (ours)	✓	✓	✓	0.556	0.541	0.495	0.557	0.494	0.691
1	AIDO Protein-RAG (16B) ( Sun et al., 2024 )	—	✓	✓	0.518	0.517	0.426	0.522	0.491	0.635
2	VenusREM ( Tan et al., 2025b )	✓	✓	✓	0.518	0.495	0.454	0.533	0.459	0.650
3	ProSST ( $K{=}2048$ ) ( Li et al., 2024 )	✓	✓	—	0.507	0.476	0.445	0.530	0.431	0.653
5	S3F-MSA ( Zhang et al., 2024 )	—	✓	✓	0.496	0.502	0.440	0.479	0.477	0.581
8	Protriever ( Weitzman et al., 2025 )	—	—	✓	0.479	0.487	0.396	0.496	0.479	0.537

Appendix

Table 15: Complete zero-shot ProteinGym substitutions comparison under the official 217-assay Average Spearman protocol, split by biological category. The same models appear by taxon and MSA depth in Table 16 . Bold red, purple, and black mark the top 3 distinct scores per column at reported precision.

Rank	Model	Seq	Str	Evo	Human	Other euk.	Prok.	Virus	Low	Medium	High
—	VenusREM2 (ours)	✓	✓	✓	0.562	0.615	0.601	0.540	0.553	0.557	0.623
1	AIDO Protein-RAG (16B) ( Sun et al., 2024 )	—	✓	✓	0.531	0.587	0.558	0.522	0.498	0.534	0.585
2	VenusREM ( Tan et al., 2025b )	✓	✓	✓	0.529	0.582	0.549	0.492	0.495	0.524	0.577
3	ProSST ( $K{=}2048$ ) ( Li et al., 2024 )	✓	✓	—	0.516	0.573	0.549	0.454	0.465	0.507	0.580
5	S3F-MSA ( Zhang et al., 2024 )	—	✓	✓	0.502	0.558	0.521	0.502	0.469	0.509	0.547
8	Protriever ( Weitzman et al., 2025 )	—	—	✓	0.480	0.542	0.492	0.516	0.464	0.498	0.512

Appendix

Table 16: Complete zero-shot ProteinGym substitutions comparison for the same models as Table 15 , split by official taxon and MSA-depth bins. MSA depth uses

N_{\mathrm{eff}}/L

: Low

<1

, Medium

[1,100)

, High

\geq 100

. Bold red, purple, and black mark the top 3 distinct scores per column at reported precision.

Stratification	Group	Raw	VRH	$\Delta$
Taxon	Human	0.524	0.562	+0.038
	Other eukaryotes	0.596	0.615	+0.019
	Prokaryote	0.579	0.601	+0.022
	Virus	0.476	0.540	+0.064
MSA depth	Low	0.498	0.553	+0.055
	Medium	0.517	0.557	+0.040

Appendix

Table 17: VenusREM2 ProteinGym Average Spearman stratified by taxon and MSA depth, recomputed with the official ProteinGym evaluator. MSA depth uses

N_{\mathrm{eff}}/L

: Low

<1

, Medium

[1,100)

, High

\geq 100

\Delta

is VRH minus Raw.

Rank	Model	Seq	Str	Evo	1	2	3	4	$5+$
—	VenusREM2 (ours)	✓	✓	✓	0.576	0.473	0.468	0.424	0.473
1	AIDO Protein-RAG (16B) ( Sun et al., 2024 )	—	✓	✓	0.527	0.414	0.419	0.394	0.414
2	VenusREM ( Tan et al., 2025b )	✓	✓	✓	0.534	0.397	0.355	0.322	0.368
3	ProSST ( $K{=}2048$ ) ( Li et al., 2024 )	✓	✓	—	0.521	0.394	0.317	0.277	0.332
5	S3F-MSA ( Zhang et al., 2024 )	—	✓	✓	0.499	0.333	0.378	0.346	0.383
8	Protriever ( Weitzman et al., 2025 )	—	—	✓	0.473	0.320	0.427	0.387	0.427

Appendix

Table 18: Complete zero-shot ProteinGym substitutions comparison for the same models as Table 15 , split by official mutation-depth bins. Depth is the number of amino-acid substitutions per variant;

5+

denotes at least five substitutions. Bold red, purple, and black mark the top 3 distinct scores per column at reported precision.

Figure 9: Native-background diagnostics on the original 37-model cohort. (a) Protein-level amino-acid profiles versus UniProtKB/Swiss-Prot composition. (b) Pairwise Pearson correlations among those profiles. (c) Readout-scale comparison across scoring interfaces. (d) Mean directional alignment $c=L^{-1}\sum_{i}\operatorname{corr}_{a}(R_{i},b)$ over protein–configuration pairs.

Dataset	Pairwise	Corpus	Native–MSA		Complete sites	Aligned sites
	$r$	$r$	Median $r$	Pooled $r$	$N$	(%)
ProteinGym	0.995	0.991	0.965	0.758	27,763	19.7
VenusMutHub	0.996	0.989	0.964	0.863	148,983	24.8
VenusViroHub	0.990	0.885	0.976	0.621	29,416	38.1

Appendix

Table 19: Cross-benchmark summary of native-background agreement and local MSA preference separation. Complete sites are unique WT positions with all required profiles.

Figure 10: ProteinGym replication. (a) Profiles from 49 models and UniProt abundance (gray bars). (b) The original 37-model native-background consensus versus MSA abundance (central 99%; correlations use all

217\times 20

pairs). (c) Raw– VRH Spearman by family across 71 configurations.

Figure 11: VenusMutHub replication. (a) Profiles from 49 models and UniProt abundance (gray bars). (b) The original 37-model native-background consensus versus MSA abundance (central 99%; correlations use all

905\times 20

pairs). (c) Raw– VRH Spearman by family across 71 configurations.

Figure 12: VenusViroHub replication. (a) Profiles from 37 models and UniProt abundance. (b) Their native-background consensus versus MSA abundance. (c) MSA site preferences versus corpus and native backgrounds. (d) Raw– VRH Spearman by family across 71 configurations.

Family	$N$	Raw	VRH	Mean change	Median change	Wins / $N$
Structure-aware	26	0.450	0.485	+0.035	+0.029	26 / 26
Masked sequence LM	20	0.359	0.440	+0.081	+0.062	20 / 20
Inverse folding	10	0.382	0.447	+0.065	+0.070	10 / 10
Autoregressive LM	15	0.302	0.410	+0.108	+0.098	15 / 15

Appendix

Table 20: ProteinGym Average Spearman by backbone family across 71 paired configurations and 217 assays.

N

is the number of configurations. Mean and median change are paired VRH –Raw differences; Wins /

N

counts configurations with positive change.

Metric	Raw	VRH	$\Delta$
Spearman	0.541 $\pm$ 0.180	0.575 $\pm$ 0.159	+0.034
NDCG	0.794 $\pm$ 0.138	0.813 $\pm$ 0.130	+0.019
AUC	0.796 $\pm$ 0.100	0.815 $\pm$ 0.091	+0.019
MCC	0.422 $\pm$ 0.164	0.450 $\pm$ 0.151	+0.028
Top-recall	0.261 $\pm$ 0.131	0.270 $\pm$ 0.131	+0.008

Appendix

Table 21: VenusREM2 ProteinGym metrics, reported as assay-macro mean

\pm

standard deviation over 217 assays. These means are not official Average Spearman.

\Delta

is VRH –Raw; the VRH column is the six-

K

ProSST ensemble; top-recall denotes top-variant recall.

Category	Structure (26)		Masked LM (20)		Inverse folding (10)		AR LM (15)		All (71)
	Raw	VRH	Raw	VRH	Raw	VRH	Raw	VRH	Raw	VRH
Activity	0.442	0.486	0.371	0.453	0.318	0.414	0.321	0.434	0.379	0.456
Binding	0.381	0.407	0.288	0.353	0.327	0.384	0.251	0.313	0.320	0.369
Expression	0.459	0.487	0.369	0.433	0.382	0.429	0.334	0.400	0.397	0.445
Organismal Fitness	0.378	0.423	0.295	0.394	0.294	0.373	0.334	0.397	0.334	0.402
Stability	0.589	0.623	0.474	0.567	0.588	0.635	0.272	0.507	0.489	0.584

Appendix

Table 22: Mean ProteinGym Average Spearman by biological category and backbone family across the 71 configurations. Counts in the header are the number of configurations in that family. Structure

=

structure-aware; Masked LM

=

masked sequence LM; AR LM

=

autoregressive LM.

Family	$N$	Raw	VRH	Mean change	Median change	Wins / $N$
Structure-aware	26	0.172	0.214	+0.042	+0.032	26 / 26
Masked sequence LM	20	0.134	0.196	+0.062	+0.050	20 / 20
Inverse folding	10	0.210	0.256	+0.046	+0.049	10 / 10
Autoregressive LM	15	0.102	0.182	+0.080	+0.070	15 / 15

Appendix

Table 23: VenusMutHub macro-averaged assay-level Spearman by backbone family across 71 paired configurations and 905 assays. Column definitions match Table 20 .

Family	$N$	Raw	VRH	Mean change	Median change	Wins / $N$
Structure-aware	26	0.215	0.280	+0.065	+0.063	26 / 26
Masked sequence LM	20	0.072	0.222	+0.151	+0.147	20 / 20
Inverse folding	10	0.271	0.307	+0.036	+0.036	10 / 10
Autoregressive LM	15	0.099	0.206	+0.107	+0.110	15 / 15

Appendix

Table 24: VenusViroHub hierarchically aggregated Spearman by backbone family across 71 paired configurations and 89 assays, excluding wild-type controls. Column definitions match Table 20 .

Benchmark	Resampling unit	Units	Mean $\Delta$	Wins / $N$
ProteinGym	UniProt ID	186	+0.068	71 / 71
VenusMutHub	wild-type sequence	527	+0.056	71 / 71
VenusViroHub	phenotype–backbone cell	52	+0.094	71 / 71

Appendix

Table 25: Mean paired VRH –Raw changes across 71 configurations.

\Delta

follows each benchmark’s primary aggregation: Average Spearman for ProteinGym, assay-macro Spearman for VenusMutHub, and hierarchical Spearman for VenusViroHub. Units follow the paired bootstrap in Appendix B.1 ; Wins /

N

counts positive changes.

Figure 13: Mean assay Spearman on the 21-assay validation split ( ProSST -2048) under a fixed mixing coefficient

\alpha

, for retrieval only and the full pipeline; bands are 95% assay-bootstrap intervals. Dashed lines: the entropy-adaptive

\alpha

of Equation 8 . Both endpoints degrade, and the adaptive coefficient matches or exceeds the best fixed coefficient.

Comparison	Mean $\Delta$	Median $\Delta$	Wins / $N$
Gated $-$ no correction	+0.012	+0.010	71 / 71
Ungated $\kappa=1$ $-$ no correction	+0.003	+0.003	43 / 71
Gated $-$ ungated $\kappa=1$	+0.010	+0.009	71 / 71

Appendix

Table 26: Coherence-gate ablation on 71 ProteinGym configurations (official Average Spearman).

Figure 14: Leave-one-site-out coherence across 217 ProSST-2048 assays. (a)

\kappa_{\mathrm{LOSO}}

preserves the full-protein gate. (b) Calibration gain over MSA retrieval versus coherence; red points show quartile means and 95% bootstrap intervals.

Interface	Context at site $i$	WT residue visible	Field construction	Multi-mutant scoring
Masked marginal	$x_{\setminus i}$	No	$L$ masked inputs	Additive native field
Wild-type field	Full native $x$	Yes	1 forward pass	Additive native field
Teacher forcing	Prefix $x_{<i}$	No (prefix only)	1 forward pass	Additive unless rescored

Appendix

Table 27: Conditioning and cost of three scoring interfaces. Masked inputs can be batched; teacher forcing hides the target residue at its own conditional.

Dataset	Masked			Wild-type
	Raw	VRH	$\Delta$	Raw	VRH	$\Delta$
ProteinGym	0.389	0.468	+0.079	0.395	0.458	+0.062
VenusMutHub	0.157	0.218	+0.061	0.142	0.187	+0.045
VenusViroHub	0.082	0.239	+0.156	0.076	0.263	+0.188

Appendix

Table 28: Mean assay Spearman across eight paired checkpoints: ESM-1b , ESM-1v , ESM-2 8M/35M/150M/650M/3B, and SaProt 650M-AF2. VRH uses entropy weighting. ProteinGym uses assay means (not official Average Spearman); both other benchmarks use macro-averages. Wild-type control rows are excluded from VenusViroHub evaluation.

Family	Model	Full sequence	Native field	$\Delta$	Wins
Inverse folding	ESM-IF1	0.422	0.421	$-0.001$	95/217
Autoregressive	ProGen2 -S	0.336	0.288	$-0.048$	47/217
Autoregressive	ProGen2 -M	0.380	0.337	$-0.043$	54/217
Autoregressive	ProGen2 -B	0.378	0.330	$-0.048$	58/217
Autoregressive	ProGen2 -L	0.380	0.327	$-0.053$	53/217
Autoregressive	ProGen2 -XL	0.391	0.351	$-0.040$	67/217

Appendix

Table 29: ProteinGym interface comparison on 217 paired assays. Full-sequence values follow the official aggregate;

\Delta

is native field minus full-sequence log-likelihood, and wins are paired at assay level.

Route	Context	Evaluations	Order	Multi-mutant score
Official random order	Mutant + structure	Per mutant	Yes	May be non-additive
Native TF	Native + structure	Per protein	No	Additive
Conditional TF	Mutant background + structure	Per background	No	Non-additive

Appendix

Table 30: ProteinMPNN official random-order versus native teacher-forcing (TF) scoring.

Figure 15: ProteinMPNN teacher forcing versus official random-order scoring on AlphaFold2 structures. (a) Mean assay Spearman across eight checkpoints. (b) Per-assay Spearman at v_48_020.

Subset	Quartets	Single TF	Single official	Double TF	Double official	$\rho_{\epsilon}$ official	$\rho_{\epsilon}$ conditional TF
All	535,917	0.306	0.105	0.284	0.165	0.005	0.012
Non-floor	400,118	0.312	0.091	0.292	0.152	0.012	0.022

Appendix

Table 31: ProteinMPNN v_48_020 on Olson GB1;

\rho_{\epsilon}

is Spearman correlation with measured pairwise epistasis.

Rank	Model	Readout	VRH	Overall	Activity	Binding	Expression	Organismal	Stability
1	VenusREM2	wt	✓	0.556	0.541	0.495	0.557	0.494	0.691
2	ProSST (K=4096)	wt	✓	0.541	0.516	0.494	0.543	0.475	0.674
3	ProSST (K=2048)	wt	✓	0.534	0.528	0.460	0.541	0.469	0.673
4	ProSST (K=1024)	wt	✓	0.532	0.513	0.467	0.533	0.476	0.670
5	ProSST (K=128)	wt	✓	0.525	0.508	0.472	0.517	0.463	0.663
6	VenusREM2	wt	$\times$	0.524	0.480	0.479	0.538	0.444	0.681

Appendix

Table 32: Unified Raw and VRH leaderboard on ProteinGym (official Average Spearman, UniProt-unit aggregation), ranked by the Overall score over the 71 configurations. Each configuration contributes a Raw (

\times

) and a VRH (

\checkmark

) entry. Readout is the native scoring interface (mask

=

masked marginal, wt

=

wild-type field, tf

=

teacher forcing). Categories are the five official selection types; the best value in each column is in red.

Rank	Model	Readout	VRH	Overall	stability	activity	PPI bind.	selectivity	DTI bind.
1	ProteinMPNN (v_48_020)	tf	✓	0.271	0.399	0.108	0.024	0.012	0.198
2	ESM-IF1	tf	✓	0.267	0.356	0.152	0.108	0.018	0.269
3	ProteinMPNN-Soluble (v_48_020)	tf	✓	0.265	0.396	0.106	0.036	0.000	0.093
4	ProteinMPNN-Soluble (v_48_030)	tf	✓	0.265	0.383	0.120	0.039	0.022	0.168
5	ProteinMPNN (v_48_030)	tf	✓	0.259	0.388	0.106	0.019	-0.020	0.137
6	VenusREM2	wt	✓	0.258	0.351	0.122	0.113	-0.017	0.274

Appendix

Table 33: Unified Raw and VRH leaderboard on VenusMutHub (assay-macro Spearman), ranked by the Overall score over the 71 configurations. Each configuration contributes a Raw (

\times

) and a VRH (

\checkmark

) entry. Readout is the native scoring interface (mask / wt / tf). Categories are the five assay tasks; the best value in each column is in red.

Rank	Model	Readout	VRH	Overall	fitness	activity	expression	cell entry	binding	stability	immune esc.
1	SaProt (650M_PDB, mask)	mask	✓	0.320	0.394	0.577	0.569	0.484	0.145	0.045	0.026
2	S3F (wt)	wt	✓	0.318	0.404	0.582	0.540	0.493	0.111	0.052	0.046
3	ESM-IF1	tf	✓	0.318	0.359	0.603	0.566	0.439	0.173	0.060	0.026
4	MIF-ST	tf	✓	0.317	0.391	0.542	0.553	0.476	0.144	0.068	0.042
5	ProSST (K=2048)	wt	✓	0.310	0.366	0.570	0.561	0.469	0.126	0.048	0.032
6	SaProt (650M_PDB, wt)	wt	✓	0.307	0.343	0.608	0.564	0.432	0.131	0.052	0.022

Appendix

Table 34: Unified Raw and VRH leaderboard on VenusViroHub (mutant-only hierarchical Spearman: means within phenotype–backbone cells, then across the seven phenotypes), ranked by the Overall score over the 71 configurations. Each configuration contributes a Raw (

\times

) and a VRH (

\checkmark

) entry. Readout is the native scoring interface (mask / wt / tf); the best value in each column is in red.

Model	Raw	Adaptive mix	+ Ungated CCD	+ Gated CCD	+ RSA	+ pLDDT (Full)
ESM-IF1	0.421	0.466	0.457	0.467	0.473	0.474
MIF-ST	0.430	0.440	0.441	0.444	0.456	0.458
ProteinMPNN (v_48_002)	0.378	0.425	0.417	0.431	0.441	0.443
ProteinMPNN (v_48_010)	0.385	0.431	0.424	0.439	0.449	0.451
ProteinMPNN (v_48_020)	0.389	0.434	0.430	0.444	0.454	0.456
ProteinMPNN (v_48_030)	0.381	0.428	0.431	0.440	0.451	0.453

Appendix

Table 35: Per-configuration staged component scores for 10 inverse-folding configurations on ProteinGym (official 217-assay Average Spearman).

Model	Raw	Adaptive mix	+ Ungated CCD	+ Gated CCD	+ RSA	+ pLDDT (Full)
ESM3	0.442	0.461	0.460	0.469	0.481	0.483
ProSST (K=1024)	0.485	0.517	0.513	0.522	0.530	0.532
ProSST (K=128)	0.469	0.506	0.503	0.512	0.522	0.525
ProSST (K=20)	0.438	0.486	0.481	0.495	0.507	0.510
ProSST (K=2048)	0.507	0.525	0.522	0.528	0.533	0.534
ProSST (K=4096)	0.498	0.525	0.522	0.531	0.538	0.541

Appendix

Table 36: Per-configuration staged component scores for 26 structure-aware configurations on ProteinGym (official 217-assay Average Spearman).

Model	Raw	Adaptive mix	+ Ungated CCD	+ Gated CCD	+ RSA	+ pLDDT (Full)
ProGen2-B	0.330	0.377	0.387	0.396	0.419	0.422
ProGen2-L	0.327	0.371	0.382	0.390	0.413	0.416
ProGen2-M	0.337	0.380	0.391	0.398	0.420	0.423
ProGen2-S	0.288	0.362	0.378	0.386	0.404	0.406
ProGen2-XL	0.351	0.373	0.375	0.385	0.415	0.419
ProGen3-112M	0.232	0.341	0.367	0.381	0.398	0.401

Appendix

Table 37: Per-configuration staged component scores for 15 autoregressive language model configurations on ProteinGym (official 217-assay Average Spearman).

Model	Raw	Adaptive mix	+ Ungated CCD	+ Gated CCD	+ RSA	+ pLDDT (Full)
CARP-38M	0.314	0.342	0.356	0.367	0.409	0.415
CARP-600K	0.159	0.221	0.249	0.271	0.345	0.354
CARP-640M	0.390	0.427	0.433	0.437	0.450	0.451
CARP-76M	0.360	0.380	0.391	0.399	0.432	0.436
ESM-1b (mask)	0.389	0.430	0.436	0.440	0.453	0.455
ESM-1b (wt)	0.394	0.407	0.414	0.418	0.445	0.447

Appendix

Table 38: Per-configuration staged component scores for 20 masked sequence language model configurations on ProteinGym (official 217-assay Average Spearman).

Model	Raw	Adaptive mix	+ Ungated CCD	+ Gated CCD	+ RSA	+ pLDDT (Full)
ESM-IF1	0.216	0.260	0.243	0.260	0.268	0.267
MIF-ST	0.196	0.209	0.203	0.210	0.228	0.230
ProteinMPNN (v_48_002)	0.191	0.233	0.232	0.241	0.245	0.244
ProteinMPNN (v_48_010)	0.218	0.240	0.246	0.247	0.254	0.256
ProteinMPNN (v_48_020)	0.221	0.257	0.254	0.262	0.269	0.271
ProteinMPNN (v_48_030)	0.210	0.244	0.251	0.251	0.260	0.259

Appendix

Table 39: Per-configuration staged component scores for 10 inverse-folding configurations on VenusMutHub (assay-macro Spearman over 905 assays).

Model	Raw	Adaptive mix	+ Ungated CCD	+ Gated CCD	+ RSA	+ pLDDT (Full)
ESM3	0.185	0.210	0.214	0.213	0.226	0.226
ProSST (K=1024)	0.162	0.223	0.236	0.231	0.241	0.242
ProSST (K=128)	0.139	0.213	0.225	0.221	0.231	0.230
ProSST (K=20)	0.143	0.213	0.216	0.224	0.237	0.238
ProSST (K=2048)	0.213	0.234	0.233	0.237	0.250	0.249
ProSST (K=4096)	0.161	0.225	0.228	0.230	0.238	0.237

Appendix

Table 40: Per-configuration staged component scores for 26 structure-aware configurations on VenusMutHub (assay-macro Spearman over 905 assays).

Model	Raw	Adaptive mix	+ Ungated CCD	+ Gated CCD	+ RSA	+ pLDDT (Full)
ProGen2-B	0.124	0.168	0.183	0.180	0.193	0.194
ProGen2-L	0.120	0.159	0.175	0.175	0.189	0.188
ProGen2-M	0.115	0.154	0.173	0.167	0.182	0.182
ProGen2-S	0.091	0.149	0.179	0.179	0.191	0.192
ProGen2-XL	0.123	0.148	0.149	0.153	0.178	0.176
ProGen3-112M	0.072	0.141	0.160	0.162	0.176	0.177

Appendix

Table 41: Per-configuration staged component scores for 15 autoregressive language model configurations on VenusMutHub (assay-macro Spearman over 905 assays).

Model	Raw	Adaptive mix	+ Ungated CCD	+ Gated CCD	+ RSA	+ pLDDT (Full)
CARP-38M	0.073	0.102	0.136	0.132	0.159	0.158
CARP-600K	0.021	0.063	0.120	0.115	0.150	0.151
CARP-640M	0.157	0.187	0.191	0.194	0.206	0.206
CARP-76M	0.131	0.149	0.172	0.174	0.192	0.194
ESM-1b (mask)	0.166	0.198	0.200	0.200	0.214	0.211
ESM-1b (wt)	0.153	0.166	0.174	0.171	0.194	0.193

Appendix

Table 42: Per-configuration staged component scores for 20 masked sequence language model configurations on VenusMutHub (assay-macro Spearman over 905 assays).

Model	Raw	Adaptive mix	+ Ungated CCD	+ Gated CCD	+ RSA	+ pLDDT (Full)
ESM-IF1	0.285	0.305	0.300	0.307	0.316	0.318
MIF-ST	0.288	0.299	0.304	0.303	0.316	0.317
ProteinMPNN (v_48_002)	0.269	0.285	0.283	0.291	0.304	0.305
ProteinMPNN (v_48_010)	0.271	0.285	0.285	0.292	0.305	0.306
ProteinMPNN (v_48_020)	0.270	0.284	0.285	0.291	0.304	0.306
ProteinMPNN (v_48_030)	0.266	0.280	0.284	0.288	0.302	0.304

Appendix

Table 43: Per-configuration staged component scores for 10 inverse-folding configurations on VenusViroHub: mutant-only hierarchical Spearman over 89 assays, 52 cells, and 7 phenotypes; calibration and shrinkage use

\beta=1-\alpha

Model	Raw	Adaptive mix	+ Ungated CCD	+ Gated CCD	+ RSA	+ pLDDT (Full)
ESM3	0.137	0.183	0.196	0.211	0.244	0.249
ProSST (K=1024)	0.204	0.243	0.247	0.250	0.272	0.275
ProSST (K=128)	0.179	0.223	0.232	0.234	0.261	0.264
ProSST (K=20)	0.161	0.211	0.213	0.224	0.254	0.259
ProSST (K=2048)	0.271	0.290	0.284	0.294	0.309	0.310
ProSST (K=4096)	0.200	0.237	0.242	0.244	0.268	0.271

Appendix

Table 44: Per-configuration staged component scores for 26 structure-aware configurations on VenusViroHub: mutant-only hierarchical Spearman over 89 assays, 52 cells, and 7 phenotypes; calibration and shrinkage use

\beta=1-\alpha

Model	Raw	Adaptive mix	+ Ungated CCD	+ Gated CCD	+ RSA	+ pLDDT (Full)
ProGen2-B	0.076	0.123	0.132	0.136	0.172	0.175
ProGen2-L	0.064	0.121	0.136	0.137	0.169	0.173
ProGen2-M	0.117	0.152	0.159	0.165	0.208	0.212
ProGen2-S	0.045	0.114	0.128	0.133	0.167	0.171
ProGen2-XL	0.162	0.179	0.179	0.186	0.232	0.235
ProGen3-112M	0.034	0.117	0.152	0.159	0.178	0.182

Appendix

Table 45: Per-configuration staged component scores for 15 autoregressive language model configurations on VenusViroHub: mutant-only hierarchical Spearman over 89 assays, 52 cells, and 7 phenotypes; calibration and shrinkage use

\beta=1-\alpha

Model	Raw	Adaptive mix	+ Ungated CCD	+ Gated CCD	+ RSA	+ pLDDT (Full)
CARP-38M	0.027	0.067	0.096	0.112	0.213	0.222
CARP-600K	0.006	0.047	0.087	0.098	0.201	0.210
CARP-640M	0.092	0.162	0.185	0.190	0.212	0.215
CARP-76M	0.050	0.090	0.113	0.130	0.221	0.230
ESM-1b (mask)	0.093	0.169	0.192	0.197	0.220	0.223
ESM-1b (wt)	0.094	0.127	0.150	0.159	0.240	0.248

Appendix

Table 46: Per-configuration staged component scores for 20 masked sequence language model configurations on VenusViroHub: mutant-only hierarchical Spearman over 89 assays, 52 cells, and 7 phenotypes; calibration and shrinkage use

\beta=1-\alpha

Model	Spearman		NDCG		AUC		MCC		Top-recall
	Raw	VRH	Raw	VRH	Raw	VRH	Raw	VRH	Raw	VRH
ESM3	0.442	0.483	0.767	0.780	0.744	0.765	0.346	0.377	0.228	0.236
ProSST (K=1024)	0.485	0.532	0.761	0.790	0.764	0.788	0.372	0.410	0.230	0.251
ProSST (K=128)	0.469	0.525	0.754	0.789	0.757	0.786	0.363	0.407	0.227	0.253
ProSST (K=20)	0.438	0.510	0.745	0.786	0.739	0.777	0.336	0.393	0.210	0.242
ProSST (K=2048)	0.507	0.534	0.757	0.778	0.777	0.791	0.398	0.416	0.236	0.249

Appendix

Table 47: ProteinGym structure-aware models with complete five-metric Raw– VRH exports.

Model	Spearman		NDCG		AUC		MCC		Top-recall
	Raw	VRH	Raw	VRH	Raw	VRH	Raw	VRH	Raw	VRH
ESM-IF1	0.421	0.474	0.747	0.777	0.729	0.759	0.328	0.372	0.223	0.239
MIF-ST	0.430	0.458	0.772	0.778	0.736	0.752	0.339	0.362	0.233	0.235
ProteinMPNN (v_48_002)	0.378	0.443	0.741	0.770	0.706	0.741	0.302	0.345	0.218	0.239
ProteinMPNN (v_48_010)	0.385	0.451	0.747	0.774	0.709	0.745	0.303	0.350	0.225	0.244
ProteinMPNN (v_48_020)	0.389	0.456	0.746	0.774	0.710	0.747	0.304	0.353	0.226	0.245

Appendix

Table 48: ProteinGym inverse-folding models with complete five-metric Raw– VRH exports.

Model	Spearman		NDCG		AUC		MCC		Top-recall
	Raw	VRH	Raw	VRH	Raw	VRH	Raw	VRH	Raw	VRH
CARP-38M	0.314	0.415	0.709	0.755	0.675	0.729	0.248	0.322	0.186	0.212
CARP-600K	0.159	0.354	0.649	0.740	0.590	0.694	0.127	0.271	0.144	0.194
CARP-640M	0.390	0.451	0.747	0.767	0.716	0.748	0.307	0.355	0.208	0.223
CARP-76M	0.360	0.436	0.728	0.759	0.699	0.740	0.282	0.339	0.196	0.214
ESM-1b (mask)	0.389	0.455	0.742	0.765	0.715	0.750	0.306	0.360	0.202	0.218

Appendix

Table 49: ProteinGym masked protein LMs with complete five-metric Raw– VRH results.

Model	Spearman		NDCG		AUC		MCC		Top-recall
	Raw	VRH	Raw	VRH	Raw	VRH	Raw	VRH	Raw	VRH
ProGen2-B	0.330	0.422	0.736	0.765	0.682	0.731	0.259	0.332	0.193	0.212
ProGen2-L	0.327	0.416	0.737	0.764	0.681	0.728	0.258	0.328	0.192	0.214
ProGen2-M	0.337	0.423	0.736	0.764	0.687	0.732	0.266	0.333	0.191	0.213
ProGen2-S	0.288	0.406	0.708	0.747	0.658	0.721	0.223	0.314	0.175	0.207
ProGen2-XL	0.351	0.419	0.755	0.772	0.694	0.730	0.275	0.329	0.198	0.207

Appendix

Table 50: ProteinGym autoregressive LM models with complete five-metric Raw– VRH exports.

Model	Spearman		NDCG		AUC		MCC		Top-recall
	Raw	VRH	Raw	VRH	Raw	VRH	Raw	VRH	Raw	VRH
ESM3	0.185	0.226	0.837	0.842	0.598	0.619	0.147	0.187	0.219	0.219
ProSST (K=1024)	0.162	0.242	0.822	0.845	0.585	0.627	0.131	0.187	0.201	0.254
ProSST (K=128)	0.139	0.230	0.817	0.840	0.575	0.620	0.103	0.176	0.181	0.234
ProSST (K=20)	0.143	0.238	0.820	0.841	0.576	0.627	0.114	0.189	0.191	0.225
ProSST (K=2048)	0.213	0.249	0.840	0.847	0.614	0.630	0.170	0.199	0.228	0.246

Appendix

Table 51: VenusMutHub structure-aware models with complete five-metric Raw– VRH exports.

Model	Spearman		NDCG		AUC		MCC		Top-recall
	Raw	VRH	Raw	VRH	Raw	VRH	Raw	VRH	Raw	VRH
ESM-IF1	0.216	0.267	0.833	0.846	0.613	0.640	0.164	0.206	0.203	0.235
MIF-ST	0.196	0.230	0.837	0.842	0.599	0.618	0.141	0.172	0.219	0.229
ProteinMPNN (v_48_002)	0.191	0.244	0.830	0.844	0.598	0.626	0.149	0.183	0.199	0.230
ProteinMPNN (v_48_010)	0.218	0.256	0.836	0.848	0.611	0.631	0.163	0.199	0.229	0.253
ProteinMPNN (v_48_020)	0.221	0.271	0.835	0.851	0.611	0.638	0.164	0.194	0.225	0.266

Appendix

Table 52: VenusMutHub inverse-folding models with complete five-metric Raw– VRH exports.

Model	Spearman		NDCG		AUC		MCC		Top-recall
	Raw	VRH	Raw	VRH	Raw	VRH	Raw	VRH	Raw	VRH
CARP-38M	0.073	0.158	0.816	0.831	0.543	0.587	0.057	0.120	0.195	0.206
CARP-600K	0.021	0.151	0.804	0.830	0.513	0.581	0.013	0.109	0.160	0.198
CARP-640M	0.157	0.206	0.830	0.838	0.584	0.609	0.128	0.152	0.192	0.212
CARP-76M	0.131	0.194	0.826	0.837	0.567	0.600	0.099	0.152	0.200	0.224
ESM-1b (mask)	0.166	0.211	0.831	0.838	0.589	0.612	0.137	0.162	0.208	0.215

Appendix

Table 53: VenusMutHub masked protein LMs with complete five-metric Raw– VRH results.

Model	Spearman		NDCG		AUC		MCC		Top-recall
	Raw	VRH	Raw	VRH	Raw	VRH	Raw	VRH	Raw	VRH
ProGen2-B	0.124	0.194	0.823	0.837	0.564	0.598	0.101	0.159	0.184	0.210
ProGen2-L	0.120	0.188	0.823	0.835	0.560	0.598	0.091	0.147	0.187	0.210
ProGen2-M	0.115	0.182	0.825	0.835	0.558	0.595	0.078	0.133	0.189	0.206
ProGen2-S	0.091	0.192	0.815	0.838	0.546	0.600	0.069	0.157	0.174	0.223
ProGen2-XL	0.123	0.176	0.827	0.836	0.561	0.591	0.083	0.134	0.195	0.202

Appendix

Table 54: VenusMutHub autoregressive LM models with complete five-metric Raw– VRH exports.

Model	Spearman		NDCG		AUC		MCC		Top-recall
	Raw	VRH	Raw	VRH	Raw	VRH	Raw	VRH	Raw	VRH
ESM3	0.137	0.249	0.690	0.733	0.573	0.632	0.103	0.200	0.124	0.152
ProSST (K=1024)	0.204	0.275	0.673	0.727	0.608	0.645	0.157	0.219	0.125	0.153
ProSST (K=128)	0.179	0.264	0.671	0.725	0.596	0.640	0.139	0.212	0.127	0.152
ProSST (K=20)	0.161	0.259	0.669	0.721	0.586	0.637	0.122	0.212	0.126	0.155
ProSST (K=2048)	0.271	0.310	0.691	0.730	0.643	0.664	0.223	0.252	0.135	0.157

Appendix

Table 55: VenusViroHub structure-aware models with mutant-only hierarchical phenotype–backbone means; AUC/MCC use 82 assays (45 cells), and the other metrics use 89 assays (52 cells).

Model	Spearman		NDCG		AUC		MCC		Top-recall
	Raw	VRH	Raw	VRH	Raw	VRH	Raw	VRH	Raw	VRH
ESM-IF1	0.285	0.318	0.720	0.755	0.648	0.666	0.226	0.258	0.160	0.171
MIF-ST	0.288	0.317	0.736	0.752	0.653	0.669	0.234	0.259	0.166	0.173
ProteinMPNN (v_48_002)	0.269	0.305	0.729	0.754	0.647	0.663	0.229	0.252	0.170	0.180
ProteinMPNN (v_48_010)	0.271	0.306	0.728	0.753	0.647	0.664	0.225	0.253	0.169	0.178
ProteinMPNN (v_48_020)	0.270	0.306	0.722	0.749	0.647	0.663	0.225	0.255	0.164	0.175

Appendix

Table 56: VenusViroHub inverse-folding models with mutant-only hierarchical phenotype–backbone means; AUC/MCC use 82 assays (45 cells), and the other metrics use 89 assays (52 cells).

Model	Spearman		NDCG		AUC		MCC		Top-recall
	Raw	VRH	Raw	VRH	Raw	VRH	Raw	VRH	Raw	VRH
CARP-38M	0.027	0.222	0.630	0.727	0.515	0.621	0.028	0.178	0.103	0.145
CARP-600K	0.006	0.210	0.614	0.722	0.504	0.615	0.013	0.170	0.099	0.143
CARP-640M	0.092	0.215	0.658	0.709	0.547	0.614	0.071	0.170	0.115	0.135
CARP-76M	0.050	0.230	0.637	0.728	0.527	0.625	0.040	0.181	0.109	0.151
ESM-1b (mask)	0.093	0.223	0.649	0.710	0.545	0.616	0.064	0.180	0.111	0.138

Appendix

Table 57: VenusViroHub masked LMs with mutant-only hierarchical phenotype–backbone means; AUC/MCC use 82 assays (45 cells), and the other metrics use 89 assays (52 cells).

Model	Spearman		NDCG		AUC		MCC		Top-recall
	Raw	VRH	Raw	VRH	Raw	VRH	Raw	VRH	Raw	VRH
ProGen2-B	0.076	0.175	0.674	0.716	0.543	0.596	0.064	0.141	0.123	0.141
ProGen2-L	0.064	0.173	0.675	0.724	0.537	0.594	0.049	0.139	0.124	0.144
ProGen2-M	0.117	0.212	0.688	0.730	0.562	0.612	0.090	0.164	0.128	0.149
ProGen2-S	0.045	0.171	0.648	0.711	0.528	0.593	0.039	0.140	0.113	0.137
ProGen2-XL	0.162	0.235	0.715	0.744	0.588	0.627	0.128	0.190	0.146	0.153

Appendix

Table 58: VenusViroHub autoregressive LMs with mutant-only hierarchical phenotype–backbone means; AUC/MCC use 82 assays (45 cells), and the other metrics use 89 assays (52 cells).

Figure 16: New prediction. The README dashboard preview places sequence, structure, and MSA inputs on the left and model selection on the right. The readiness checklist and Start scoring button guide submission of a prediction.

Figure 17: Review and export. The ranked candidate table is linked to the structure viewer and residue-level evidence. Selected rows can be exported using Add to CSV . Screenshot reproduced from the README preview.

Figure 18: Benchmark browsing. Dataset, metric, property, and input-type filters support inspection of Raw and harness-adjusted scores. The README screenshot shows the 71-configuration ProteinGym comparison.

Explore similar work

Better Protein Function Prediction by Modeling Survivorship Bias
May 7, 2026Zhongmou Chao, Poompol Buathong, Ekaterina Selivanovitch +2Survival Predictions
Smith School of Chemical and Biomolecular Engineering, Cornell University, USA · Center for Applied Mathematics, Cornell University, USA · School of Operations Research and Information Engineering, Cornell University, USA
ProtSent: Protein Sentence Transformers
May 7, 2026Dan Ofer, Oriel Perets, Michal Linial +1Language Modeling Transformer Architectures
Department of Computer and Information Science Ben-Gurion University of the Negev
PFArena: Benchmarking Language Models for Protein Modification
Sep 24, 2026Yawen Ouyang, Xinbo Zhang, Ziyuan Ma +15Protein Quality-Driven Selective Mutation
Shanghai Artificial Intelligence Laboratory · Generative Symbolic Intelligence Lab (GenSI), Tsinghua University · Institute for AI Industry Research (AIR), Tsinghua University +1

May 7, 2026, cs.LGJ/K move · Enter open · S save

Better Protein Function Prediction by Modeling Survivorship Bias

Zhongmou Chao, Poompol Buathong, Ekaterina Selivanovitch, Susan Daniel+1

Smith School of Chemical and Biomolecular Engineering, Cornell University, USA · Center for Applied Mathematics, Cornell University, USA · School of Operations Research and Information Engineering, Cornell University, USA

Survival Predictions

Open paper AskPDF arXiv

Protein sequence data from nature exhibits survivorship bias: we only observe data from those organisms that survive and reproduce, while non-functional protein mutations are eliminated by natural selection. Thus, predicting whether a protein sequence is functional often requires learning from positive examples alone. While positive-unlabeled (PU) learning frameworks offer a generic solution to this problem, existing PU methods ignore the evolutionary processes that shape sequence observability and cause survivorship bias. Consider a sequence that is one mutation away from a commonly-observed protein variant in a well-surveilled organism. If the sequence were functional, it would likely be observed. If it is not observed, this suggests non-functionality. In contrast, sequences that are unlikely to arise through mutation may be missing simply because they never arose. Thus, these two kinds of missing sequences should be treated differently when training models. In this work, we propose Evo-PU, a PU learning framework that uses a scientific understanding of nucleotide mutation to model survivorship bias for well-surveilled single-organism sequence data. On three prediction tasks using single-organism uniform-coverage surveillance data -- predicting results from held-out influenza and respiratory syncytial virus (RSV) mutagenesis studies, and predicting future SARS-CoV-2 variants -- Evo-PU outperforms standard PU learning, one-class classification (OCC), and protein language models (PLMs). On prediction tasks from multi-organism ProteinGym datasets with more heterogeneous surveillance coverage, we identify opportunities to generalize our approach.

Abstract

Explore similar work

Better Protein Function Prediction by Modeling Survivorship Bias

ProtSent: Protein Sentence Transformers

PFArena: Benchmarking Language Models for Protein Modification