https://link.springer.com/article/10.1038/s44320-026-00189-7

Predicting the protein interaction landscape of a free-living bacterium with pooled-AlphaFold3  

Key figures

• Figure 2: Establishes the two core technical claims of the paper, namely the size bias in ipTM and the predictive value of size-corrected pooled scores for known PPIs. 
  
• Figure 3: Directly shows that larger pooled jobs outperform pairwise jobs for recovering known interactions, which is the paper’s main methodological advance. 
  
• Figure 5: Shows that pairwise interactions extracted from pooled jobs can still recover structurally accurate interfaces within a large macromolecular complex, not just binary hits. 

1) Thesis (one sentence)

To address the gap of computationally scalable and accurate genome-scale PPI discovery, in the Mycoplasma genitalium proteome, pooled-AlphaFold3 prediction causes improved recovery of biologically relevant protein interactions by co-folding many mostly noninteracting proteins per job to suppress false positives and enable size-corrected interface scoring, supported by genome-scale benchmarking, structural recapitulation, and orthogonal cross-linking evidence. 

2) Evidence card (three bullets only)

• Strongest result: (Fig. 2D, Fig. 3) Size-corrected pooled ipTM recovered known STRING experimental interactions with AUROC 0.81 for strong interactions and 0.88 for the strongest interactions, whereas individually folded pairs performed much worse, and predictive performance improved as pool size increased up to 5000 aa. 
  
• Method enabler: (Fig. 1, Fig. 2, Methods; computational structural biology + genome-scale screening with AlphaFold3, robust regression, STRING benchmarking, Bio3D, PyXlinkViewer) The enabling workflow was greedy packing of 476 proteins into 2027 pools near the 5000 aa limit, extraction and averaging of raw ipTM over 5 models, robust size correction using expected ipTM = -0.036255571 + 0.004470512 \cdot \sqrt{\text{aa}_1+\text{aa}_2}, and benchmarking against STRING plus structural and XL-MS comparisons. 
  
• Critical limitation: (Fig. 2F, Fig. EV7, Methods) AlphaFold3 interface scores were intrinsically variable across identical runs with different random seeds, and full performance gains depended on very large pools near the 5000 aa ceiling, so smaller pools or single-seed pair runs can yield unstable or inflated interaction calls. 

Optional

Quote bank (2–4 short excerpts)

• Quote 1: “pooled-PPI prediction improves the accuracy of genome-scale screens” (Abstract, page 1)
  
• Quote 2: “pooled-AlphaFold3 predictions recapitulate experimentally validated PPIs” (page 2)
  
• Quote 3: “the full benefit of pooling requires at least 5000 aa pools” (Results, page 5)
  
• Quote 4: “variability in ipTM scores is intrinsic to AlphaFold3” (Results, page 6)

Key comparisons (1–3 lines)

• Compared to: Standard pairwise AlphaFold3 co-folding for genome-scale PPI discovery.
  
• Win: Higher specificity, fewer false positives, fewer jobs, and lower total inference burden while still recovering known interfaces in large assemblies.
  
• Tradeoff: Requires careful pool design, size-bias correction, and sufficiently large pool sizes near hardware or server token limits. 

Methods I might copy (protocol hooks)

• Construct design / Models: All-by-all proteome screen of 476 proteins from M. genitalium; greedy pooling algorithm to ensure every pair is sampled at least once while keeping jobs below 5000 tokens; additional random paired, 2k, 3k, 4k, and 5k pool benchmarks. 
  
• Conditions / Instruments: AlphaFold3 server or local AlphaFold 3.0.1 with default settings; local runs used 1 seed, 5 diffusion samples, 10 recycles, and template cutoff date September 30, 2021; 2027 comprehensive pools contained 4 to 23 proteins with median n=13; total screen took 68 person-days using the free web allotment. 
  
• Readout / Analysis: Mean raw ipTM extracted from summary_confidences files across 5 models; robust linear regression for size correction; AUROC benchmarking against STRING experimental scores; structural RMSD analysis with Bio3D; XL-MS mapping with PyXlinkViewer using a 30 Å satisfaction threshold. 

Open questions / Theoretical implications (2–5 bullets)

• Can pooled competitive folding be generalized from bacterial proteomes to larger eukaryotic target subsets without unacceptable combinatorial dilution or context effects? 
  
• Could pooled multimer prediction be adapted from unbiased proteome screening to focused interface-remodeling benchmarks where redesigned proteins compete for one shared binding partner? 
  
• How much of pooled-AF3’s gain comes from true physical competition versus generic spatial decoy effects within the generative model? 
  
• Might repeated pooled inference across designed variant libraries provide a practical computational filter for selecting interface redesigns before explicit ternary-complex modeling?