| # Supplement methods |
|
|
| This release uses the frozen cohorts in `evidence/freeze.json`, not every run in the live catalog. It includes current Figure 2 methods and the additional RFD4 High conditioning arms. Antibody, HalluDesign, full-pipeline BindCraft and Genie3 search are excluded. The per-run evidence is under `models/<model>/provenance/`; exact native inputs and timing receipts take precedence over any current upstream default. |
|
|
| ## Common accounting and timing |
|
|
| Each model × protocol × target has **200 original attempts**, including native failures and unmatched required motifs. We do not replace rejected originals or select outputs with the common AF3 evaluation. Protein-Hunter and RFD4 search have explicit intrinsic selection; those stages count toward their generation cost. Novo's two arms share original scaffolds and are not independent samples. |
|
|
| `generation_gpu_s` is the complete stored warm generation cost, including required sequence design, intrinsic folding/search, and native gaps inside that timed boundary. It excludes setup, loading, warmup and compilation where separately measured. It is not a diffusion-only timer. Batched costs are amortized across the batch's originals; batch shares are not individually measured latency. `timing_workers.csv` retains available setup and warm totals. `batch_size` is the measured actual batch, including tails—not a claim of maximal possible throughput. Unknown metadata remains blank. Exact-length/bucket compilation limitations are retained in executed receipts; the release does not assert that all models were equally optimized. |
|
|
| `af3_gpu_s` is separate common evaluation time. Default efficiency is the **arithmetic mean across targets of successes / (generation + AF3 GPU time)**, multiplied by 600 for per-10-GPU-minute units. Failed originals and their consumed time stay included. MPNN and native-fold columns are components already included in generation: **do not add them again**. A zero means the component did not run or used no GPU; blank means it was not separately measured. A missing AF3 cost excludes that configuration from an inclusive efficiency comparison, not its successes from rate comparisons. |
|
|
| Timing exceptions: ODesign retains the user-approved original-generation timing proxy, which includes its old MPNN step rather than separately adding corrected MPNN. Complexa fixed-small-molecule costs are registered batched remeasurements, not original batch-1 wall time. Historical RFD4 PPI all-atom uses the approved backbone timing proxy and has no measured AF3 cost. These are identified in `timing_basis`; they are not disguised as direct per-design measurements. |
|
|
| ## Model recipes and adaptations |
|
|
| | Model | Executed generation and input/output considerations | Author reference | |
| |---|---|---| |
| | RFD4-Proteina | Joint sequence/structure generation with stored Low/Medium/High schedules; exact schedule/checkpoint comes from each native config. Search is a separate intrinsic workflow: AF2-Multimer for PPI and OF3 for the other benchmarks, with all search work timed. Preserve native motif correspondence; unmatched required motifs fail. Historical PPI-AA validation is explicitly separate from the newer BB campaign. | [RFD4 repository](https://github.com/RosettaCommons/RFD4-Proteina) | |
| | RFdiffusion3 | Native 200 steps, CFG off, step size 1.5; one LigandMPNN sequence at temperature 0.1, no downstream pipeline screening. DNA uses the DNA-finetuned checkpoint. Native JSON/derivatives preserve shared lengths and correspondence. Approved PPI OXT omission and native M0110 bond/charge/preprocessing losses are retained. Ligands can follow native reference conformers; no stereo repair. | [Foundry](https://github.com/RosettaCommons/foundry) | |
| | BoltzGen | Design: 500 steps/3 recycles; inverse folding: 200 steps/3 recycles, one final sequence. Original indices 0–99 use diverse, 100–199 adherence, preserving the native quarter settings across shards. No terminal folding/filtering/ranking. Atom hotspots become residue hotspots; PPI is all-atom. Seven heme cases use disclosed native bond-type projections; difficult chemistry uses source-derived archives, never author benchmark molecules. | [BoltzGen](https://github.com/HannesStark/boltzgen) | |
| | Proteina-Complexa | One 400-step joint draw, no additional MPNN/selection. PPI uses V1; AME V2/LoRA32; fixed small molecules use the corrected dedicated V1/LoRA32 ligand route. Native matcher is retained: unmatched motifs fail downstream. PPI and ligand inputs use native atom37/template features and residue-projected hotspots. No DNA or fully unfixed ligand route. | [Proteina-Complexa](https://github.com/NVIDIA-BioNeMo/Proteina-Complexa) | |
| | DISCO | Native diverse/max joint draw, 200 diffusion steps/four recycles. No extra sequence designer. SDF/JSON derives from shared CIF chemistry, never checked-in benchmark SDFs. Same-target/exact-length batches; all batch members retained. Native hydrogen/bond-order losses are documented, not repaired. | [DISCO](https://github.com/DISCO-design/DISCO) | |
| | Genie3 | One 100-step generative draw and one CA-MPNN sequence at 0.1. **No search or ColabFold.** Force the original hotspot, using target CA and source hotspot sidechain geometry, not extended contacts. Native CA-only outputs stay CA-only and are evaluated with binder CA RMSD. Some executed input receipts reference an earlier source release; exact executed files remain preserved, not silently rebound to a newer queue. | [Genie3](https://github.com/aqlaboratory/genie3) | |
| | BindCraft (single backbone) | One 75/45/5/15 optimization trajectory, then one soluble ProteinMPNN v_48_020 sequence at 0.1. Three continuation gates bypassed to retain bad trajectories. No post-MPNN fold/relax/filter/rank. Native independent batch 1; exact lengths warmed. Native interface-fixed residues retained. Output sidechains incompatible with the new sequence become unobserved rather than being grafted. | [BindCraft tutorial](https://github.com/martinpacesa/BindCraft/wiki/De-novo-binder-design-with-BindCraft) | |
| | BindCraft2 (single backbone) | One 50/25/45/5/15 trajectory, then one native soluble-MPNN redesign at 0.1. Continuation gates disabled; no independent validation or relaxation. Native independent batch 1, unchunked; resident AF2/MPNN pair, separately warmed signatures. Preserve exact native redesign constraints and report omitted incompatible sidechains. | [BindCraft2](https://github.com/PacesaLab/BindCraft2) | |
| | Protein-Hunter | Seven approved native refinement cycles, 100%-X initialization, 200 steps/3 recycles, MPNN 0.1. SolubleMPNN for PPI; LigandMPNN for ligands/DNA. Retain highest native ipTM cycle among those with ≤20% alanine. All intrinsic cycles are timed. Native rejection and writer/chemistry defects are not rescued. PPI guidance is residue hotspots and CA/CB/frame guidance, not hard sidechain enforcement. | [Protein-Hunter](https://github.com/yehlincho/Protein-Hunter) | |
| | ESMFold2 inversion | 150 inversion steps, two pinned models/four intrinsic hero critics. Critic 0 is predesignated; no common-AF3 ranking. Target sequence/residue hotspots, not target coordinates. Multichain extension preserves all target copies and native whole-complex ipTM, including target–target contributions. Intrinsic critics are not separate benchmark attempts. | [ESM binder-design tutorial](https://github.com/Biohub/esm/blob/main/cookbook/tutorials/binder_design.ipynb) | |
| | RFdiffusion | 50-step Complex_base draw with example CA/frame noise scales 0; one ProteinMPNN v_48_020 sequence at 0.1. Resident native batch 1. No Rosetta/refold/selection. Explicit contig/chain maps retain target copies and gaps. Preserve native backbone-only output and atom omissions. | [RFdiffusion binder design](https://github.com/RosettaCommons/RFdiffusion#binder-design) | |
| | RFdiffusion2 | AME RFD_140 checkpoint, 100 native flow steps, one LigandMPNN 0.1 sequence with native packing/context. Resident native batch 1. M0110 uses the explicitly approved disconnected ligand projection: preserve source atoms/coordinates but omit two bonds in the native derivative. This is not equivalent physical topology. Native guidepost matching/idealization is retained. | [RFdiffusion2](https://github.com/RosettaCommons/RFdiffusion2) | |
| | ODesign | Native scaffold generation followed by the corrected one-LigandMPNN0.1, source-mask-fixed sequence route. No terminal AF3 selection. Native AME receives disclosed extra observed motif context and is not a strict tip-only baseline. Fixed-ligand and fixed-DNA membership is labeled explicitly. Original-generation timing is a proxy, as described above. | [ODesign](https://github.com/OTeam-AI4S/ODesign) | |
| | AlphaProtein-novo | Native 1000-step joint codesign. Paired LigandMPNN arm uses one sequence at 0.1, checkpoint 030, atom context on, sidechain context off, no packing. CPU MPNN adds zero GPU seconds; shared scaffold generation time is inherited, not omitted. Native APN1 composite-ligand handling of M0110 changes interface interpretation and is disclosed. | [AlphaProtein-novo](https://github.com/google-deepmind/alphaprotein-novo) | |
|
|
| Native source/current workflow snapshots and immutable repository citations are listed in `references.csv`. They explain the implementation; executed receipts—not the newest README—identify what each cohort actually ran. |
|
|
| ## Common AF3 scoring |
|
|
| The existing accelerated external-eval AF3 outputs are reused; this release launches no generation, folding or rescoring. Protein targets are templated from propagated coordinate-conditioning information using stored coordinates. DNA alignment includes the target DNA in both conditioning modes. Template sanity metrics are diagnostic, not substitutions for binder accuracy. Native biological-symmetry handling is retained in stored scores; do not derive alternative scores merely by relabeling chains. |
|
|
| | Benchmark | Current pass criterion | |
| |---|---| |
| | PPI, 24 targets | AF3 binder pTM ≥80 on the native 0–100 scale; minimum interface PAE ≤1.5 Å; target-backbone-aligned binder backbone RMSD ≤2.5 Å (CA-only for Genie3). Required target mapping/conditioning must be valid. Target mutations at a hotspot or within 5 Å of binder heavy atoms in design or refold fail; distal substitutions may pass. | |
| | Small molecule, 77 targets | AF3 minimum interface PAE ≤1.5 Å, chain-of-interest ipTM ≥80/100, protein-backbone RMSD ≤1.5 Å, and protein-backbone-aligned ligand RMSD ≤5 Å. This is `sm_binder`, not the looser historical `disco` criterion. | |
| | Original DISCO, same 77 targets | Protein-backbone RMSD ≤2 Å and maximum ligand-centroid displacement ≤2 Å after protein-backbone alignment. No AF3 PAE/ipTM cutoff. Stored verdicts and independently computed passing-cluster labels are retained separately from `sm_binder`. The criteria are not mathematically nested because the ligand-position tests differ. | |
| | Atomic Motif Enzyme (AME), 37 targets | Protein-backbone RMSD ≤2 Å, required motif all-atom RMSD ≤1.5 Å, and AF3 minimum interface PAE **<1.5 Å**. Motif RMSD aligns/scores mapped conditioned protein atoms against stored input references, respecting coordinate groups. Native-unmatched/sequence-failed motifs fail; missing metrics are not assigned zero. Clashes are diagnostic only in this current policy. | |
| | DNA, 3 targets | AF3 minimum interface PAE ≤1.5 Å and DNA-target-aligned binder backbone RMSD ≤8 Å. Alignment uses all target DNA atoms in both fixed/unfixed modes. | |
|
|
| Ordinary rate = passing originals / 200. Clustered rate = distinct clusters among passing originals / 200. Clusters are computed **within each model/protocol/target**, using generated binder/protein structures and Foldseek TM threshold 0.5, alignment type 1, minimum sequence identity 0; they are not pooled between targets or models. Dark bar plus light extension is the ordinary rate; dark alone is clustered. These are algorithmic structural clusters, not experimental successes or globally distinct proteins. |
|
|
| ## Supplement selection and conditioning comparisons |
|
|
| Per-target panels include RFD4 High, the best other generative model, and best non-generative model by mean clustered rate (ordinary rate breaks ties). The requested AME panel instead shows four: RFD4 High, AlphaProtein-novo Codesign, Proteina-Complexa and RFdiffusion3. Both small-molecule criteria get separate panels and comparator selection. Full selection and all model values remain in CSVs. |
|
|
| The High conditioning panels use all 200 originals per target and do not assume the desired direction. **PPI-AA and PPI-BB are different campaigns**; other settings/checkpoints can confound conditioning. AME compares *partially* to fully fixed ligand, not fully unfixed. Three NAD small-molecule cases use different input conformers. Therefore these plots show observed differences, not isolated causal effects of removing constraints. |
|
|
| Historical originals with unavailable separate native/raw/AF3 artifacts remain represented. `artifacts-audited.csv` and `input_coverage.csv` distinguish available files from unresolved or intentionally retired evidence; a missing file is not a new failed design. |
|
|
