Практические workflows
Каждый workflow показывает место generator, integrated scorer и external validation.
| Workflow | Сценарий | Evidence |
|---|---|---|
| Antibody inverse folding / masked CDR redesign | Есть Fv/VHH structure; нужно перебрать CDR/framework sequences, сохранив backbone. | 2 retrospective antibody evaluation; no prospective wet-lab in primary paper. |
| Antibody sequence infilling and CDR diversification | Нужно разнообразить CDR или infill span без antigen structure. | Retrospective naturalness/diversity. No antigen conditioning. |
| Antigen-conditioned CDR co-design | Есть antibody–antigen complex context; нужно co-design CDR sequence и geometry. | DiffAb/dyMEAN: 1–2 in silico. FlowDesign/AbDiffuser: narrow prospective, with caveats. |
| Antigen-conditioned CDR-H3 generation (PALM-H3) | Есть antigen/epitope sequence и antibody framework; нужно сгенерировать CDR-H3. | 3–4 prospective on SARS-CoV-2 family; predicted affinity ≠ experimental result. |
| Target-conditioned de novo antibody (RFantibody) | Есть target structure и epitope hotspots; нужен de novo VHH или scFv на фиксированном framework. | 3–4 prospective binding/function. Target success 0–2% reported. Initial affinities can be modest. |
| De novo backbone → inverse folding | Нужна новая topology или motif scaffold, затем sequence. | RFdiffusion: prospective general protein evidence. Vanilla model fails on antibody loops. |
| Epitope-specific mAb CDR generation (tFold) | Есть antigen sequence и epitope; нужен paired mAb на human framework с virtual screening. | 3–5 prospective, но generation, ranking и display intertwined. |
| Few-shot / labelled mutation optimization | Есть parent и возможность собирать experimental labels; нужна итеративная оптимизация, иногда multi-objective. | EVOLVEpro and Makowski: prospective. Not transferable pretrained generators. |
| Fixed-backbone inverse folding | Есть backbone (экспериментальный или predicted); нужно подобрать sequence. | ProteinMPNN/LigandMPNN: prospective structural/functional. Likelihood ≠ affinity. |
| Framework humanization | Есть non-human antibody или CDRs; нужно human-like framework. | Humatch retrospective. HuDiff limited prospective preprint. Humanness ≠ immunogenicity. |
| Joint / multimodal sequence–structure generation | Нужно coupled generation sequence и structure, возможно с function prompts. | ESM3/Chroma/ProteinGenerator have prospective general-protein examples, not antibody-specific. |
| Target-conditioned miniprotein binder (BindCraft) | Есть target structure и hotspots; нужен non-antibody binder. | 3–4 prospective binding/function; nanomolar examples. Not antibody. |
| Sequence-first de novo generation | Нет parent-структуры; нужен sequence prior для семейства или unconstrained exploration. | Зависит от инструмента: ProGen2/ESM3 prospective function; EvoDiff in silico. |
| All-atom universal binder generation (BoltzGen) | Frontier all-atom binder design across proteins, peptides, nanobodies. Preprint. | 3–4 prospective claimed in preprint; evidence differs by modality. Not full IgG evidence. |
| Zero-shot mutation proposal | Есть parent antibody/protein; labels почти нет; нужно предложить single substitutions. | Both highlighted workflows have prospective antibody-relevant evidence. |