De novo design of high-affinity protein binders with AlphaProteo (Zambaldi et al., arXiv 2024)

What this source contributed

Records in this release that cite this source. A record can cite several sources.

• 5 configurations
• 1 pipeline
• 4 protocols
• 8 datasets
• 81 evaluations
• 81 results
• 9 evidence claims
• 1 use case

Results

Values as the source printed them. Compare results only under compatible protocols, datasets and metrics.

Claims

Statements this source supports, each with where in the source it was checked.

• AlphaProteo Table 1 'Other design methods' rows are copied from other publications: Compilation rows not extracted Table 1, row 'Other design methods', Experimental success rate (%): BHRF1 18 a (17); SC2RBD 1.6 b (63); IL-7RA 0.15 c (14,912); PD-L1 13 b (16); TrkA 0.07 c (14,982); IL-17A 0.02 d (15,000); VEGF-A and TNF-alpha print an en dash. Table 1, row 'Other design methods', Binding KD (nM): BHRF1 58 a (17); SC2RBD 100 e (100,000); IL-7RA 3 c (14,912); PD-L1 0.9 b (16); TrkA 3000 c (14,982); IL-17A 47 d (15,000). Table 1, row 'Other design methods, optimized', Binding KD (nM): BHRF1 16*,a; SC2RBD 16*,e; IL-7RA 0.31 c; PD-L1 0.65 f; TrkA 1.4 c; IL-17A 0.01 d. Footnotes: a Procko et al. [35]; b Gainza et al. [11]; c Cao et al. [6]; d Berger et al. [3]; e Cao et al. [5]; f Yang et al. [45]. Table S2, row 'RFdiffusion (published)', Success (%): IL-7RA 33.7 (95); PD-L1 12.6 (95); TrkA 6.3 (95). Table S2, row 'RFdiffusion (published)', Binding KD (nM): IL-7RA 30 (95); PD-L1 1400 (95); TrkA 328 (95). Not recorded as results because the values are copied from publications this pass did not retrieve, and because Section S4.2 states that the asterisked values were measured by the AlphaProteo authors themselves by HTRF and that for PD-L1 they substituted a different design (RFD_PDL1_76, KD 1.6 nM) for the published best design (RFD_PDL1_77, KD 1.4 micromolar). That is a third party's reanalysis of another group's result. (Table 1, rows 'Other design methods' and 'Other design methods, optimized'; Table S2, row 'RFdiffusion (published)'; Sections S4.1 and S4.2)
• AlphaProteo Table 1 pools three different design systems: Aggregation scope Section S2, verbatim: 'We experimentally tested a design system containing the v1 model against all 7 targets, a v1-based system with an improved filter against SC2RBD and PD-L1, and a v2-based system on PD-L1 and TrkA (Table S2). Both improving the filter and the model resulted in increased experimental success rates. For simplicity, the results in Table 1 are pooled over all tested designs for each target.' The Table S2 denominators sum to the Table 1 denominators: SC2RBD 94+47+31=172, PD-L1 94+31+34=159, TrkA 94+37=131, TNF-alpha 30+24=54. Changing the filter alone moved the PD-L1 rate from 9.6% over 94 designs to 16.1% over 31 designs. (Section S2; Table 1; Table S2)
• Baseline for the de novo binder design use case: Baseline Directed evolution, display selection or grafting (Adaptyv preprint Sections 2.3 and 3.2 (EGF and cetuximab as positive controls; cetuximab at 9.94 nM); Bennett 2025 Discussion (contrast with yeast display selection from random libraries); AlphaProteo Table 1 caption ('Binders derived from selection-based methods, such as antibodies and nanobodies, are not considered here'))
• Relevance of uc-binder-20261022-protocol-alphaproteo-ecoli-htrf-funnel to "Which computational design workflow produces binders that work when they are made and tested?": Assessed by AlphaProteo expression and HTRF confirmation funnel for yeast-display hits (uc-binder-20261022-source-alphaproteo: Table S3)
• Relevance of uc-binder-20261022-protocol-alphaproteo-htrf-kd to "Which computational design workflow produces binders that work when they are made and tested?": Assessed by AlphaProteo best per-target binder affinity by HTRF (uc-binder-20261022-source-alphaproteo: Table 1, 'Binding KD (nM)' block; Table S2; Section S4.2)
• Relevance of uc-binder-20261022-protocol-alphaproteo-yeast-display-signal-0-2 to "Which computational design workflow produces binders that work when they are made and tested?": Assessed by AlphaProteo yeast-display binding screen, hit at binding signal above 0.2 (uc-binder-20261022-source-alphaproteo: Section S1.2.1; Table 1; Table S2)
• Relevance of uc-binder-20261022-protocol-alphaproteo-yeast-display-signal-1-3 to "Which computational design workflow produces binders that work when they are made and tested?": Assessed by AlphaProteo yeast-display binding screen, hit at binding signal above 1.3 (IL-17A) (uc-binder-20261022-source-alphaproteo: Section S1.2.1; Table 1, column 'IL-17A')
• Section S4.1 states that the RFdiffusion arm is the published designs re-tested, not a rerun: Evaluation actor Section S4.1, verbatim: 'Published experimental success rates of RFdiffusion were based on 96-well biolayer interferometry (BLI) measurements [S28], which are impractical for us to perform on our own designs. Therefore, to quantitatively compare success rates, we downloaded sequences of the published RFdiffusion designs and synthesized and tested them by yeast display (Section S1.2) alongside our own designs. Our results matched the published success rates for PD-L1 but were 2-fold lower for IL-7RA (16% versus 33%) and 0% for TrkA (versus 6%) (Table S2). This is potentially due to differences between the yeast surface environment versus purified proteins, as well as our use of 10- to 20-fold lower target concentrations (0.5-1 uM versus 10 uM in [S28]), which may exclude lower-affinity hits.' Section S4.1 therefore states a shared assay and does not state a shared selection funnel: the AlphaProteo candidates were chosen by the authors' own filter from a larger in silico set, while the RFdiffusion candidates are the 95 designs the RFdiffusion authors published. (Section S4.1)
• Summary of the de novo binder design evidence: Summary Four studies, pinned here as seven artifacts, carry measured outcomes for de novo binder design, and none of them answers which workflow to use, because no paper runs two design methods and a conventional comparator on the same targets with one assay and one hit definition. The literature pass behind this evidence found nine candidate sources with measured outcomes; the five that were not retrieved are recorded as gaps rather than as evidence of what they would show. What the measured evidence shows is that the target matters more than the method. AlphaProteo's yeast-display success rate across eight targets runs from 0% over 54 designs on TNF-alpha to 88% over 94 designs on BHRF1, at a binding-signal threshold of 0.2 (1.3 on IL-17A). BindCraft reports an average experimental success rate of 46.3% over twelve targets with a range of 10% to 100%, built from per-target counts as small as 2 binders from 7 designs on Bet v1 and as large as 6 from 6 on the SpCas9 REC1 domain. The same family of methods applied to antibody variable domains is two orders of magnitude worse: fine-tuned RFdiffusion gives 2% over 96 designs on TcdB, 0% over 96 on IL-7 receptor alpha, and 0.01% to 0.2% over 9,000 designs per target by yeast display. On influenza H1 stem the same design campaign gives 1% over 96 designs by SPR and 0.2% over 9,000 by display, so the screening method moves the rate as much as the target does. Nearly all of this is author-reported by the group that built the method, which also chose the targets, ran the assay and chose which designs to test. The exception is the Adaptyv EGFR competition, where a third-party laboratory expressed and measured designs from many entrants on one target and one assay: 2.5% of 201 round-1 designs bound (printed elsewhere in the same paper as 3%), and 53 of 400 round-2 designs bound, printed as 13%. That source cannot be turned into a per-method rate: design methods are the entrants' own declarations, the organisers hand-picked 101 of 201 round-1 and 300 of 400 round-2 designs for diversity, and the three binding thresholds behind the figure counts are never defined. Affinities reached are better than the hit rates suggest. AlphaProteo's best measured binder was 0.082 nM on IL-7 receptor alpha by HTRF; BindCraft's best ranged from 5 nM on CbAgo to 5.7 micromolar on CrSAS-6; the best de novo design in the Adaptyv competition reached 82 nM and the winning optimised cetuximab fragment 1.21 nM, against 9.94 nM for the cetuximab positive control measured on the same instrument. Three things stop these numbers from ranking methods. The hit definitions are different endpoints, not different thresholds on one endpoint. The denominators are small and chosen, and the designs tested are the top picks of each method's own filter, so design and selection are coupled: changing only AlphaProteo's filter moved its PD-L1 rate from 9.6% over 94 designs to 16.1% over 31, and the Bennett authors attribute their low antibody rates partly to a filter they describe as weak. And the spread within one method across targets is wider than any plausible difference between methods. No conventional comparator is run anywhere in this evidence, so there is no measurement of what a designed binder is being compared against. (AlphaProteo Table 1, Table S2 and Sections S1.2.1, S2 and S4.1; BindCraft Introduction and Results prose; Bennett Supplementary Methods Table 6; Adaptyv Sections 2.3, 2.4, 3.1, 3.2, 4.1, 4.2 and 4.3)

Models, benchmarks and other records

Evidence

Source checking verifies the cited claim or transcription. It does not establish independent reproduction.