Pallatom

0. Link verification(visited2026-10-04)

Conventions: PMLR labels ICML 2025, OpenReview Spotlight. Conference dates were2025-07-13–19; volume publication2025-10-06. This report uses ICML 2025 Spotlight. Numbers come from the official PDF actually read; abstracts only verify metadata/links.

1. Candidate pool and random selection

Sunday rotation selects AI4Science and interdisciplinary research. First checked ml-report-index.mdand excluded recently covered MolJO, I²Mole, Circuit Insights, Among Us, and neighboring topics. Candidates primarily come from visited ICML 2025 PMLR Volume 267, with title/resource searches on OpenReview, arXiv, Semantic Scholar, Hugging Face Daily Papers, and Papers with Code. Aggregator coverage is incomplete, so official conference/arXiv entries determine eligibility.

30 candidates:

1. From Mechanistic Interpretability to Mechanistic Biology: Training, Evaluating, and Interpreting Sparse Autoencoders on Protein Language Models
2. A Variational Perspective on Generative Protein Fitness Optimization
3. Predicting Mutational Effects on Protein Binding from Folding Energy
4. Boosting Protein Graph Representations through Static-Dynamic Fusion
5. Elucidating the Design Space of Multimodal Protein Language Models
6. NMA-tune: Generating Highly Designable and Dynamics Aware Protein Backbones
7. Aligning Protein Conformation Ensemble Generation with Physical Feedback
8. Diffusion on Language Model Encodings for Protein Sequence Generation
9. P(all-atom) Is Unlocking New Path For Protein Design
10. PPDiff: Diffusing in Hybrid Sequence-Structure Space for Protein-Protein Complex Design
11. Reward-Guided Iterative Refinement in Diffusion Models at Test-Time with Applications to Protein and DNA Design
12. Flexibility-conditioned Protein Structure Design with Flow Matching
13. Protriever: End-to-End Differentiable Protein Homology Search for Fitness Prediction
14. Sidechain Conditioning and Modeling for Full-Atom Protein Sequence Design with FAMPNN
15. All-atom Inverse Protein Folding through Discrete Flow Matching
16. CFP-Gen: Combinatorial Functional Protein Generation via Diffusion Language Models
17. Protein Structure Tokenization: Benchmarking and New Recipe
18. ReQFlow: Rectified Quaternion Flow for Efficient and High-Quality Protein Backbone Generation
19. Bridging Protein Sequences and Microscopy Images with Unified Diffusion Models
20. OneForecast: A Universal Framework for Global and Regional Weather Forecasting
21. Fixing the Double Penalty in Data-Driven Weather Forecasting Through a Modified Spherical Harmonic Loss Function
22. Fast and Low-Cost Genomic Foundation Models via Outlier Removal
23. SPACE: Your Genomic Profile Predictor is a Powerful DNA Foundation Model
24. ViTally Consistent: Scaling Biological Representation Learning for Cell Microscopy
25. scSSL-Bench: Benchmarking Self-Supervised Learning for Single-Cell Data
26. Enforcing Latent Euclidean Geometry in Single-Cell VAEs for Manifold Interpolation
27. Code-Generated Graph Representations Using Multiple LLM Agents for Material Properties Prediction
28. Efficient Molecular Conformer Generation with SO(3)-Averaged Flow Matching and Reflow
29. Symmetry-Driven Discovery of Dynamical Variables in Molecular Simulations
30. Offline Model-based Optimization for Real-World Molecular Discovery

Executed as requested shuf -i 1-30 -n 1returned 9, selecting Pallatom. Its31-page full paper, code, main results, ablations, hyperparameter scans, and length extrapolation provide sufficient material;no redraw。

2. Motivation: beyond adding more atoms

Protein generators often first generate backbones, then add sequences with inverse folding such as ProteinMPNN, and verify with structure predictors. This divides joint p(sequence,all-atom coordinates) into conditional distributions—a mature pipeline with three specific gaps.

First,Side chains are not a negligible rendering layer. Active sites, hydrogen-bond networks, hydrophobic cores, and ligand contacts depend on side-chain atoms. Cα/backbone-only models can yield apparently foldable backbones without paying atom-clash or chemical-consistency costs during generation. Second,sequence and structure are strongly coupled. Backbone-first pipelines push early geometric errors into downstream repair; a backbone designable only by a narrow sequence set may still score well. Third,existing all-atom representations are costly or complex: Protpardelle’s atom73 superposes possible side chains, entangling discrete identity with continuous coordinates during reverse generation.

Main experiments sample250 candidates at each of 60,70,80,90,100,110,120 aa, totaling1,750. Strongest co-design baseline MultiFlow has CO-DESIGN DES-aa 62.74%versus Pallatom 85.03%, an absolute increase of 22.29 ppand relative increase of 35.53%. Foldseek structural clusters rise from 134 to 291, adding 157 clusters, or 117.16%; MMseq2 sequence clusters increase by 1,042 to 1,466, adding 424, or 40.69%(Table 1, PDF p.7). Joint all-atom modeling can expand designable diversity rather than necessarily sacrifice it.

This does not imply universal superiority over backbone methods. With ProteinMPNN postprocessing, Pallatom DES-bb is 89.89%below ProteinGenerator’s 93.14%, a difference of 3.25 pp(Table 1, p.7). Evidence favors the specified co-design protocol, not dominance over two-stage methods in every design setting.

3. Narrative, core assumptions, and tradeoffs

The narrative has four steps.

1. Unified representation: each residue becomes a fixed 14×3 coordinate block. Real atoms occupy slots; missing side-chain atoms sit temporarily at Cα. Initially only backbone element types are known; side-chain slots are UNK。
2. Geometry before identity, within one trained network: diffusion predicts denoised Atom14 geometry; SeqHead predicts20 amino-acid classes and removes virtual slots. Unlike external ProteinMPNN, sequence and geometry heads train jointly.
3. Separate global/local tracks: residue Transformers model long-range topology; atom modules handle nearby geometry. AtomDecoder converts denoised coordinates into new pair representations fed back into the backbone.
4. Differentiable structural proxies: atom regression, sequence CE, smooth-LDDT, residue/atom distograms, and intermediate losses train jointly; sampling uses 200 first-order Euler steps.

The core assumption is that chemically redundant fixed Atom14 slots make variable side chains batchable, with geometric-learning benefits outweighing ambiguity from UNK and virtual Cα points. Second,recycling matters more than a one-shot prediction from noisebecause identities, side chains, and neighborhoods require repeated reconciliation. Third, proxy designability predicts promising experiments. Ablations support the first two; no wet-lab evidence supports the third.

Fixed slots simplify parallelization while sacrificing explicit chemical identity. Larger η accelerates denoising/provides long-chain success but pushes toward helical attractors.200 Euler steps yield roughly 10–58s sampling without establishing a quality–cost frontier.

4. Equations and sampling

Inputs: length N,T=200,η=2.25,γ=0.2,t∈[0.01,1]
Initialize Atom14 x_T~N(0,σ_T²I); all side-chain identities UNK
for i=T...1:
    1. Encode x_i with global residue/local atom tracks
    2. MainTrunk denoises once, producing coordinates/pair features
    3. AtomDecoder rebuilds distograms and feeds them back
    4. MainTrunk denoises again; SeqHead predicts identities
    5. Compute Euler updates from EDM scores;η/γ adjust steps/noise
end
Remove virtual slots by predicted residue type; output sequence/all-atom coordinates

5. Key mechanisms and engineering details

Table 6(PDF p.20): batch 32, LR0.001, no warmup,35,156 samples/epoch, crop 128; α0–α4=0.25,1,0.5,0.5,1; polar-residue weight2, others 1; single/pair dimensions256/128,8 heads/decoder units.300,000 steps on 4 A6000s take about 10 days, roughly 960 A6000 GPU-hour. Exact parameters, peak memory, and energy are absent; fast sampling does not establish low lifecycle cost.

6. Datasets, benchmarks, and metrics

PDB filtering uses resolution below 3Å,95% sequence identity, and excludes structures over 50% loops, leaving7,459. AFDB-cluster begins from a nominal214-million database, retains582,652 with mean pLDDT≥80/length≤128, then filters neighborhood density≥20, core fraction≥30%, fragments≥4, gyration radius<25, maximum loop≤15, and Foldseek TM0.8/coverage0.9, removing about 30% near-duplicates and retaining27,697. Total35,156 per epoch(§4.1, Appendix B, pp.6,19–20).214 million is upstream database size, not training-set size.

7. Main-results matrices

8. Layered ablations, hyperparameter sensitivity, and transitions

9. Genuine contributions, reusable design, and transfer potential

Actual contribution 1: fixed slots bridge continuous diffusion and discrete chemistry. Geometry forms first, then SeqHead decides which slots exist, converting variable dimensionality into masking without directly diffusing amino-acid tokens.

Actual contribution 2: all-atom recycling is reusable feedback. Table 3’s−66 pp shows that coordinate→pair/distogram feedback matters more than one-shot prediction. Ligand complexes, RNA, and crystal cells could similarly recondition denoising on differentiable geometry summaries.

Actual contribution 3: explicit sampling transitions. Tables 2/7 exposeη as both an integrator parameter and morphological-prior control, more informative than only reporting optimal settings.

Transfer should retain local atoms/global entities/recycling without necessarily retaining Atom14. Complexes can use residue/chain/atom hierarchies; molecules explicit elements/bonds; long proteins partial latents as in La-Proteina. Functional tasks need active-site geometry, energy, or experimental labels, beyond fold self-consistency.

10. Related work: problem origins and technical lineage

RoseTTAFold/AlphaFold-style pairs + recycling
              │
              ├── RFdiffusion: backbone diffusion → ProteinMPNN sequences
              │                         └── MultiFlow: discrete sequence + continuous backbone
              │
Explicit atoms┼── Protpardelle: atom73 superset
              │            └── Pallatom: Atom14 + SeqHead + atom recycling
              │                         ├── APM: multichain complexes
              │                         └── FAMPNN: fixed-backbone side-chain conditioning
              │
Latents ──────└── La-Proteina: explicit Cα +8D/residue latents, up to 800 aa

11. Limitations acknowledged by the paper

1. Task scope: general architectures/complex systems remain future work(Discussion, PDF p.9). Unconditional monomer results do not establish complexes, membrane proteins, or ligand systems.
2. Scale limits: crop 128 training(Table 6,p.20) extrapolates to 400 aa, but Table 8 falls behind CarbonNovo after350, directly revealing a boundary.
3. Validation limits: only computational designability/diversity/novelty, without expression, stability, binding, or function experiments; claims remain in silico.

12. Independent critique (evidence → implications)

1. [Evaluation validity] Because DES-aa uses ESMFold pLDDT/aaRMSD while training relies heavily on high-pLDDT AFDB structures, generator/evaluator share predictor-ecosystem preferences;therefore 85.03% more directly means predictor-judged self-consistency than experimentally verified foldability.
2. [Long-range generalization] Because crop 128 training yields93.3% DES-aa at 200 aa but17.9% at 400,20.9 pp below CarbonNovo,therefore the global residue track lacks truly length-robust topology priors.
3. [Mode collapse] Becauseη3.0 at 400 aa gives92% HHH/0% EEE while smallerη yields only 4% DES-aa,therefore some gains likely come from helical shortcuts rather than free sampling acceleration.
4. [Incomplete attribution] Because removing recycling costs 66 pp without an equal-compute deeper/wider nonrecycling control,therefore the implementation works, but iterative feedback cannot be cleanly separated from extra forward compute.
5. [Resources/reproduction] Because data pipelines/sampling commands are available but author weights and the curated35,156-sample snapshot were not found,therefore third parties can rebuild the process but cannot precisely reproduce Table 1 or exclude database-version effects.
6. [Statistics] Because main/ablation/hyperparameter tables lack seed variance/CIs, with SD only for timing,therefore the 22.29 pp main gap is more credible than small2–3 pp smooth-LDDT/γ differences.
7. [Cost conventions] Because training uses about 960 A6000 GPU-hours without parameters, peak memory, emissions, or preprocessing-inclusive costs,therefore 10.2s sampling does not establish overall efficiency.
8. [Biological function] Because results concern unconditional monomers/structural proxies without active-site, binding-energy, or functional readouts,therefore protein design here means foldable sequence–structure co-design, not a complete functional-engineering workflow.

13. Open challenges: actionable head-to-head gaps

14. Recent-work radar

1. La-Proteina: Atomistic Protein Generation via Partially Latent Flow Matching(2025-07-13): explicit Cα plus8D latents reportedly scale to 800 aa, directly addressing long-chain collapse, though46-million-pair training complicates fairness.
2. An All-Atom Generative Model for Designing Protein Complexes(arXiv v1,2025-04-17; ICML 2025): APM extends all-atom generation to multichain complexes and releases Code, addressing monomer limits but requiring strict new-target splits against leakage.
3. PPDiff(ICML 2025): interleaved sequence/structure networks and causal attention co-design binders, reporting pretraining/mini-binder/antibody success50.00%,23.16%,16.89%. More functional, but not equal-protocol all-atom monomer comparison.
4. NMA-tune(ICML 2025): lowest-normal-mode conditioning plugs into RFdiffusion and tests dynamics with MD, advancing from folding to desired motion.
5. FAMPNN(ICML 2025): joint identities/side-chain conformations in fixed-backbone inverse folding improve zero-shot stability/binding prediction, complementing Pallatom’s generation-from-noise focus.
6. Consistent Synthetic Sequences Unlock Structural Diversity in Fully Atomistic De Novo Protein Design(2025-12-01): ProteinMPNN/predictor-generated consistent synthetic pairs improve La-Proteina structural diversity54% and co-designability27%, suggesting pairing quality may dominate representation.
7. Exploring the Alignment of Generation and Understanding in Protein Structure Modeling(2026-07-15): aligns generators to understanding encoders, raising Protpardelle-1c MotifBench39.2→47.1(relative 20%), adding functional representation beyond geometric consistency.

15. Conclusion

Pallatom establishes that simple Atom14 slots plus strong recycling can encode sequence–side-chain coupling without complex discrete diffusion. Its main-table +22.29 pp DES-aa with 2.17× structural clusters are substantial progress; Table 3’s −66 pp without recycling reveals the mechanism, while long-chain/helical transitions in Tables 7/8 define boundaries.

Treat Pallatom as a strong baseline/module library for explicit all-atom joint generation, not a wet-lab-validated platform. Equal-data/compute comparisons of explicit versus latent representations and blinded experiments calibrating proxies matter more than further optimizing ESMFold pass rates.

Sources/completeness footnote: core numbers use only the 31-page official PMLR PDF. Missing SDs, weight links, downloads, and wet-lab outcomes remain missing.ml-report-index.md Report entry appended. Charts use consistent Table 8 DES-aa; if local /usr/bin/chromium or /opt/pw-browsersis absent, build logs will mark unavailable Playwright visual regression rather than claim browser checks.