From Protein Language to Medicine

Why these tools, why now?

Two persistent problems in protein engineering motivate this pair of papers.

Both papers use ESM-2 as a starting point — but in opposite ways. PepMLM uses it to generate. EVOLVEpro uses it as a coordinate system. They address different problem classes.

ESM-2 — the shared foundation

ESM-2 (Meta AI) is a protein language model trained on ~250 million sequences using masked language modelling (MLM): tokens are randomly hidden and the model is trained to predict them from context. The largest version has 15 billion parameters. No structural or functional labels are used — the training signal is the sequence itself.

Two outputs — why the distinction matters

PepMLM — peptide binders from sequence alone

Chen, Quinn, Dumas, Peng et al. · Nature Biotechnology, 2025

PepMLM fine-tunes ESM-2 (650M) to reconstruct peptide binders from target sequence context — without structural input at any stage. The core use case is disordered targets where structure-based tools cannot operate.

Training strategy

Training data: 10,000 known peptide–protein pairs (PepNN + Propedia). Each pair is concatenated as one sequence — target first, peptide at the end. The entire peptide region is masked. The model is trained to reconstruct the binder from target context alone.

Benchmarking (computational)

Head-to-head with RFdiffusion (Baker lab) — current structure-based benchmark. Scored by AlphaFold-Multimer ipTM. One peptide generated per target.

Experimental validation

1. Binding — NCAM1 (ELISA)

All 4 PepMLM peptides showed concentration-dependent binding response against recombinant NCAM1 down to ~60 nM (Fig. 2a). Best peptide statistically distinct from BSA, polyG, IgG controls at 30 nM (P = 0.0051, Fig. 2b). RFdiffusion peptides for the same target showed substantially weaker signal.

2. Targeted degradation — mutant huntingtin in patient cells

Target: mutant huntingtin (mHTT), an IDP with polyglutamine expansion >36 repeats. Cells: TruHD fibroblasts (patient-derived, Q43/Q17 CAG configuration — clinically pathological repeat length, not an overexpression model).

Caveats

EVOLVEpro — guided protein optimisation

Jiang, Yan, Di Bernardo et al. · Science, 2025

Given a functional protein and a quantitative activity assay, improve performance with a small number of experimental measurements.

The fitness–activity divergence

Interpretation: ESM-2's training signal is the distribution of sequences that natural selection maintained. Evolutionary fitness reflects reproductive success. Engineering goals — editing efficiency at a human locus, dsRNA contamination in a manufacturing process — have never been selection pressures. The embedding retains value as a geometric structure; the fitness score is not a reliable proxy for engineering goals.

Architecture and loop

Results — PsaCas12f (compact CRISPR nuclease)

PsaCas12f (~529 aa) fits in a single AAV vector alongside its guide RNA — SpCas9 (~1,368 aa) does not. Four rounds of optimisation using an indel assay at RNF2 in HEK293 cells.

In vivo: epPsaCas12f + guide RNA co-packaged in single AAV, injected into mice targeting PCSK9 in liver. Blood PCSK9 fell to ~50% of baseline by day 14 (Fig. 2F). Single-vector AAV delivery not achievable with SpCas9.

Results — T7 RNA polymerase (epT7)

T7 RNAP synthesises therapeutic mRNA in vitro. Problem: dsRNA contamination activates RIG-I/MDA5, driving reactogenicity. Six rounds optimised translation output, dsRNA contamination, and mRNA quality score simultaneously.

In vivo: Luciferase mRNA (epT7 vs. WT T7) in LNPs, injected in mice; the LNPs traffic to the liver. Hepatic bioluminescence at 24 h: ~10× higher for epT7-produced mRNA (Fig. 6H).

Results — REGN10987 antibody

Chosen because prior computational tools had failed to improve it. Five rounds produced 63% improvement in binding affinity (IC50 = 11.9 nM, Fig. 1E). Used primarily to demonstrate the active learning success rate trajectory (Fig. 1C): ~9% round 1 → ~54% round 5.

Caveats

Side by side

What does the ESM-2 embedding contribute to EVOLVEpro if fitness scores are misleading?

The embedding is not the same as the fitness score. The embedding encodes positional and contextual relationships across the full sequence — it tells the random forest which variants are geometrically nearby in protein space. This provides a smoothness assumption: similar sequences in embedding space are likely to have similar activities. Without the embedding, the random forest operates in raw sequence space with no such structure. The embedding does not indicate which direction to move; it establishes that the activity surface is probably locally continuous, making interpolation feasible with few training points.

What these papers contribute

PepMLM repurposes the MLM objective for conditional binder generation

The model is ESM-2 with a different training setup: target protein + fully masked peptide are concatenated, and the model is trained to reconstruct the binder from target context alone. This means the standard masked language modelling objective — designed for sequence completion — is used to learn a conditional relationship between a target sequence and its binders. No structural input, no task-specific architecture change. The novelty is in the training framing, not the model itself.

EVOLVEpro's main finding: ESM-2 fitness scores do not predict engineering activity

Across five proteins evolved in the wet lab, the variants ESM-2 rates as most evolutionarily plausible were weakly or negatively correlated with measured activity (Figs. 2I, 3F, 4G, 5H). The reason is straightforward: ESM-2 learned what natural selection has preserved across species. Engineering goals — editing efficiency at a specific human genomic locus, dsRNA output during in vitro transcription — have never been selection pressures. The paper is the first to demonstrate this divergence experimentally across multiple protein classes.

The uAb format applies targeted degradation to proteins with no known structure

Targeted protein degraders (PROTACs, molecular glues) require a ligand for the target. For intrinsically disordered proteins, no such ligand exists. Fusing a PepMLM-designed peptide to an E3 ubiquitin ligase creates a targeting construct that doesn't depend on a known binding pocket or 3D structure. The 5/5 degradation result in TruHD patient cells is the first demonstration of this approach for mHTT specifically, though the generalisability to other IDP targets is not yet established.

EVOLVEpro nominates mutations outside the rational design search space

E643G in T7 RNAP is located near the DNA template binding site, a region not previously targeted for T7 engineering. The model arrived at this position by round 3 and converged on it after exploring the residue multiple times in round 4. This position would not have been a high-priority candidate from structural inspection alone. The same pattern holds for the PsaCas12f triple mutant: the combination I178A/K333V/K454P produces higher activity than any of the three mutations individually, an interaction that cannot be predicted additively.

Where the evidence is incomplete

PepMLM: the computational hit rate is measured on structured targets only

The 38% hit rate (vs. 29% for RFdiffusion, Fig. 1d) is calculated on 203 test targets, all drawn from structural databases. None are intrinsically disordered. AlphaFold-Multimer — used to score predicted binding — is not reliable for disordered proteins, so there is no computational benchmark for IDP targets. The 38% figure and the IDP wet lab results are therefore measuring different things. The computational number says nothing about how the model performs on the use case that motivated the paper.

The 5/5 Huntington's result — control is sound; the caveat is scale

Five uAb constructs reduced mHTT in TruHD cells. The negative control is a polyG–uAb: a polyglycine peptide fused to the same E3 ligase domain (paper Fig. 3 legend and Methods). Because it carries the identical E3 component but no real targeting sequence, it does control for E3-ligase-driven non-specific degradation — it is exactly the non-targeting control needed to show the PepMLM peptide is load-bearing. The real limits are scale (five constructs, one IDP target) and that degradation is doxycycline-induced in a patient fibroblast line rather than a neuronal model. Replication on a second, structurally unrelated IDP would strengthen the generality claim more than any additional control.

HMPV phosphoprotein clearance is not the same as antiviral activity

Fig. 4d shows reduced phosphoprotein immunofluorescence signal in HMPV-infected cells treated with four uAb constructs. This confirms the uAb can degrade the phosphoprotein in the context of active infection, but it does not show whether viral replication was reduced. Measuring antiviral activity requires a direct virology readout: plaque reduction, viral titre (TCID50), or quantitative PCR for viral genomic RNA.

EVOLVEpro's DMS benchmark uses a favourable protein set

The 12 DMS datasets used for in silico benchmarking are all from published, well-characterised proteins. Proteins that have published DMS data tend to be tractable — they express well, give clean activity readouts, and have been studied enough to generate comprehensive mutational data. The benchmark may not reflect how EVOLVEpro performs on proteins that are harder to characterise, have nosier activity landscapes, or belong to families not represented in the DMS panel.

The 515× T7 immunogenicity reduction is from a single cell-based assay

Measured by IFNB1 qPCR in BJ fibroblasts at 24 hours post-transfection. IFNB1 is one interferon-stimulated gene. BJ fibroblasts respond to dsRNA via RIG-I/MDA5, but are not a standard innate immune model — primary PBMCs or monocyte-derived dendritic cells would be more relevant. The 24-hour window may not capture the full kinetics of the response. The figure is consistent with the mechanistic expectation but is a single assay readout in one cell type.

Multi-objective T7 optimisation: the trade-off surface is not reported

EVOLVEpro optimised T7 RNAP for translation output, dsRNA contamination, and mRNA quality score simultaneously. E643G improved all three, so no trade-off arose in this case. The paper does not report whether other variants in the optimisation trajectory showed trade-offs between objectives — for example, variants with higher translation but more dsRNA. For future applications where engineering objectives conflict, the behaviour of the framework under those conditions is not characterised.

Assay noise tolerance is not reported

EVOLVEpro is parameterised around ~16 measurements per round. The paper does not assess how performance degrades when assay noise is high. Before applying the framework to a new protein, it is worth establishing the coefficient of variation (CV) of the intended assay and whether that level of noise is compatible with learning a meaningful activity signal from 16 variants per round.

Key terms