ProteinDPO Flashcards (IPD reading)

Widatalla et al. Nature Methods 2026, DPO alignment of ESM-IF1 + the H5N1 HA campaign. TRAP cards = where the obvious summary is wrong. last 4 cards are figures to explain.

what is ProteinDPO in one sentence?

direct preference optimization, the LLM alignment trick of training on "this one is better than that one" instead of on labels, applied to the inverse folding model ESM-IF1 so it prefers thermostable sequences without losing what pretraining taught it

who wrote ProteinDPO and where was it published?

Widatalla, Borah, King, Driscoll, Rafailov and senior author Brian Hie. Stanford, Arc Institute and UCSF. Nature Methods 23(9):1805-1813, 2026. Rafael Rafailov co-authored the original DPO paper in LLM alignment

what is the alignment gap?

the paper's framing term: the pretraining objective (recover the native sequence given a backbone) is not the objective you actually want (make this protein survive heat). unsupervised models know a lot in general and lose to small supervised models on any one property

why not just fine tune on the stability data?

because supervised fine tuning overfits the training distribution and erodes general knowledge. in this paper SFT improved in distribution scores but got WORSE than untouched vanilla on external benchmarks, FireProt Pearson down 0.05 and Spearman down 0.04, S669 Pearson down 0.01

how does DPO map onto proteins here?

backbone is the prompt, sequence is the response, experimental fitness is the preference. measurements become preferences over variants of the same backbone, and DPO rewrites the RL objective as a classification loss on those pairs

what keeps ProteinDPO from forgetting?

a KL penalty with beta = 0.1 tethering the tuned model to the frozen reference, vanilla ESM-IF1. that plus preference learning is what generalizes where SFT did not

what is the base model and why that choice?

ESM-IF1, a structure conditioned inverse folding model: given a backbone it autoregressively writes a sequence, so it assigns a likelihood to any sequence given any structure, which is what lets you score mutations

what data trained ProteinDPO?

the Megascale dataset v2, ~660,000 variants across 403-405 small domains of 40-72 residues (the Results and Methods disagree on the count), split 90/5/5 by FoldSeek structural clustering so similar domains never straddle the split

what are the three DPO objectives compared?

paired (Bradley-Terry, K = 2), ranked (Plackett-Luce, K = 3), and a new weighted objective using the scalar stability values as Boltzmann style soft weights. they land within a narrow band of each other, so the novel objective is not clearly the reason it works

how gentle is the training?

learning rate 1e-7, Adam, batch 32, up to 30 epochs, 0.1 A structural noise matching ESM-IF1 pretraining, on a single 80 GB H100. no LoRA

held out Megascale performance

ProteinDPO R = 0.72-0.73, Spearman 0.69-0.72, AUROC 0.82-0.84, versus vanilla ESM-IF1 at 0.55 / 0.53 / 0.74 and SFT at 0.59 / 0.57 / 0.76

TRAP: does ProteinDPO beat the supervised specialist?

not uniformly. the Discussion says it "has not clearly surpassed all supervised models in scoring". it loses to ThermoMPNN on FireProt (R ~0.60 vs 0.65) and ties or slightly wins on S669 (0.44-0.47 vs 0.43). the sell is comparable ranking while staying generative

the out of domain transfer claim

trained only on small single chain domains, it improved on multichain complexes it never saw: SKEMPIv2 +0.02-0.04 R and rho, AB-Bind +0.04-0.08 R, and +0.08-0.12 in both R and rho on 483 monoclonal antibody melting temperatures. ThermoMPNN cannot even evaluate multi chain proteins

what does ProteinDPO show about epistasis?

on double mutants, scoring additively it loses to ThermoMPNN (R 0.36-0.40 vs 0.41), but scoring likelihood over the whole sequence it wins (0.44-0.48). its advantage comes from seeing the sequence as a whole, and epistasis is still not solved

the H5N1 single mutation results

top 30 model ranked substitutions were measured by DSF: 19 improved Tm by more than 1 C, 7 were within 1 C, and 6 were significantly stabilizing at the paper's 9 C threshold. it also recovered six known stabilizing mutations and flagged the pH switch residues H26, K51 and E103 zero shot

what does the abstract's "~80%" actually refer to?

the whole campaign: 27 stabilizing variants out of 45 total designs tested, with ~80% either increasing or maintaining stability. the 30 singles alone were 26 of 30

the stacking result

combinations of three, five and nine substitutions were built from top ranked singles. all nine substitution variants beat wild type, the best by up to 17 C (DPO-16), averaging 9.8 C better than the singles

the transfer across 20 years of flu evolution

DPO-16's mutation set transplanted onto 2024 strains gave +13 C on Texas dairy cattle (B.3.13) and +32 C on British Columbia (D1.1) HA

how did they check they had not wrecked the antigen?

antibody binding retained at nanomolar to subnanomolar affinity for the stem bnAb CR6261 and the head antibody 13D4, with AZD3152 as a dead control, plus circular dichroism showing preserved secondary structure

TRAP: what is missing before calling this a vaccine result?

the authors concede "further evidence would be needed to demonstrate antigenic properties required for vaccine efficacy". there is no immunogenicity, no serum neutralization, no cryo-EM conformational check, no expression yield and no animal protection

TRAP: the postfusion problem

nothing in the objective penalizes the postfusion state. for a class I fusion protein thermal stability can come from relaxing into the wrong conformation. the authors flag it: discouraging the postfusion state "was not considered in variant generation"

did generation improve or just scoring?

both. sampling 500 sequences per model across five temperatures on three backbones, ProteinDPO sequences had better Rosetta energies while keeping ESMFold pLDDT > 80 and high self consistency TM-score, despite lower identity to native, so it did not collapse into copying wild type

what is Tm and how was it measured here?

melting temperature, where half the protein has unfolded, measured by differential scanning fluorimetry. it is the practical readout for every HA number in this paper

why would a nanoparticle vaccine lab care about this paper?

a displayed antigen must stay prefusion, survive fusion, assembly, purification and an imperfect cold chain, and keep its neutralizing epitopes intact. the HA result speaks to all three. and a lab that already generates display, SEC, BLI and yield data is generating exactly the paired readouts DPO consumes as preferences

the obvious question this paper opens up

could a design model be aligned to several objectives at once, stable AND assembles into the particle AND keeps epitope X intact, rather than stability alone?

explain this figure: ProteinDPO fig 1 (dataset curation and model training)

a and b are the analogy: an LLM gets prompt, response and a correct/incorrect label, while ProteinDPO gets backbone, sequence and a stable/unstable label. c is the data pipeline, 479 natural and Rosetta designed domains to a 1.8M variant library, Foldseek clustered and split into 609,000 train / 22,000 validation / 25,000 test, with a stabilizing only subset reserved for SFT. d is the argument in one picture: vanilla ESM-IF1 maximizes native sequence likelihood, SFT sees only stabilizing variants, ProteinDPO sees stabilizing AND destabilizing as a preference pair

explain this figure: ProteinDPO fig 2 (benchmarks)

a is the experimental scheme, native monomers to site directed mutagenesis to a stability assay. b is the Megascale holdout, three bar panels (Spearman, Pearson, AUROC) where vanilla and SFT sit low and the three DPO objectives sit high. c is the double mutant holdout where ThermoMPNN is competitive. d ranks ~25 methods on S669 by Pearson, with the DPO variants mid pack near ThermoMPNN. e is FireProt, where ThermoMPNN is highest and DPO is close behind. the honest read of this figure is "competitive, not dominant"

explain this figure: ProteinDPO fig 4 (zero shot stabilization of H5 HA)

a is the concept cartoon: a metastable immunogen sits in a shallow energy well between prefusion and postfusion, and the designed sequence deepens the prefusion well. b ranks the 45 designs by measured Tm, coloured by number of mutations, with wild type ~45 C as a dashed line, so the 9 mutation variants sit at the top. c maps the mutated positions onto Vietnam 2004 HA. d is the raw evidence, DSF first derivative curves with wild type melting at 45 C and DPO-4 at 55 C plus a second transition at 77 C

explain this figure: ProteinDPO fig 5 (2024 strains and antigenicity)

a superimposes the antibody complexes on HA with mutated positions marked, 13D4 on the head and CR6261 on the stem. b shows DPO-16 transplanted onto Texas 2024 and BC 2024 HA, 62 to 74 C and 43 to 75 C. c is the KD heatmap, variants still binding 13D4 at ~1e-12 and CR6261 at ~6-8e-9 against a dead control at 1e-3. d is circular dichroism showing the fold is intact. what it does not show: cryo-EM, immunogenicity, protection