Variant effect prediction

May 10, 2026

Model type less important than data quantity when predicting catalytic properties of protein sequences

variant-effect-prediction

Apr 21, 2026

No one-size-fits-all approach to scoring antibody structures

Apr 21, 2026

Accounting for substitution likelihood improves variant effect prediction relative to log-likelihood

variant-effect-prediction

Apr 21, 2026

High-confidence predictions from protein language models co-cluster together in embedding space and correlate with performance on variant effect prediction tasks

Apr 21, 2026

Missense mutation effects are predictive of effects of indels

variant-effect-prediction

Apr 21, 2026

Most ML quality metrics cannot effectively predict enzyme activity after controlling for similarity to native

Apr 21, 2026

Number of sequences with 90% homology is a greater predictor of viral variant effect prediction than alignment depth

variant-effect-prediction

Apr 21, 2026

PLMs downweigh probability of sequences with multiple mutations

variant-effect-prediction

Apr 21, 2026

Precision decreases and recall increases as variant effect prediction focuses on top-performing point mutations

variant-effect-prediction

Apr 21, 2026

Short insertions are common in evolution

Apr 21, 2026

Single deletion mutations are better tolerated than larger (2-3 residue) deletions

Apr 21, 2026

Structural tokens improve zero-shot variant effect prediction in ESM3, but only when structures are not computationally derived

Apr 21, 2026

Taking the minimum logit can outperform averaging logits when ensembling models for pathogenicity prediction

variant-effect-prediction

Apr 21, 2026

Variant effect prediction quality correlates with DMS prediction quality

variant-effect-prediction

Apr 21, 2026

Coupling strength of two substitutions is correlated with backbone distance

Apr 21, 2026

Fitness prediction using inverse folding can be improved with auxiliary B-factor and SASA losses

Apr 21, 2026

Including nonfunctional sequences during training improves prediction of negatives but not positives

variant-effect-prediction

Apr 21, 2026

ML-based protein design methods outperform physics-based methods at identifying deleterious sequences but not top sequences

variant-effect-prediction

Apr 21, 2026

No one-size-fits-all best approach to zero-shot or few-shot protein fitness prediction

variant-effect-prediction

Apr 21, 2026

Spearman correlations of protein property prediction methods do not correlate perfectly with absolute error

variant-effect-prediction

Apr 21, 2026

Weighting training sequences by fitness improves identification of other fit sequences

variant-effect-prediction

Quartz 4

Explorer

Variant effect prediction

Model type less important than data quantity when predicting catalytic properties of protein sequences

No one-size-fits-all approach to scoring antibody structures

Accounting for substitution likelihood improves variant effect prediction relative to log-likelihood

High-confidence predictions from protein language models co-cluster together in embedding space and correlate with performance on variant effect prediction tasks

Missense mutation effects are predictive of effects of indels

Most ML quality metrics cannot effectively predict enzyme activity after controlling for similarity to native

Number of sequences with 90% homology is a greater predictor of viral variant effect prediction than alignment depth

PLMs downweigh probability of sequences with multiple mutations

Precision decreases and recall increases as variant effect prediction focuses on top-performing point mutations

Short insertions are common in evolution

Single deletion mutations are better tolerated than larger (2-3 residue) deletions

Structural tokens improve zero-shot variant effect prediction in ESM3, but only when structures are not computationally derived

Taking the minimum logit can outperform averaging logits when ensembling models for pathogenicity prediction

Variant effect prediction quality correlates with DMS prediction quality

Coupling strength of two substitutions is correlated with backbone distance

Fitness prediction using inverse folding can be improved with auxiliary B-factor and SASA losses

Including nonfunctional sequences during training improves prediction of negatives but not positives

ML-based protein design methods outperform physics-based methods at identifying deleterious sequences but not top sequences

No one-size-fits-all best approach to zero-shot or few-shot protein fitness prediction

Spearman correlations of protein property prediction methods do not correlate perfectly with absolute error

Weighting training sequences by fitness improves identification of other fit sequences