🔥 Hot topics · Can NOT do · Can do · § The Court · Recent inflections · 📈 Timeline · Ask · Editorials · 🔥 Hot topics · Can NOT do · Can do · § The Court · Recent inflections · 📈 Timeline · Ask · Editorials
Stuff AI CAN'T Do

Can AI outperform humans at predicting protein-protein interactions ?

What do you think?

Recent advances have raised questions about whether AI can reliably outperform humans at predicting protein-protein interactions. While deep learning models like AlphaFold-Multimer show impressive accuracy on curated datasets, experts remain divided over whether these gains translate universally across all biological contexts. The court’s preliminary deliberations suggest a nuanced verdict—one that acknowledges AI’s progress without declaring unqualified superiority.

Background

Since 2021, deep-learning models have steadily improved PPI prediction by learning co-evolutionary signals and structural constraints from large protein sequence alignments. AlphaFold-Multimer (2021) and RosettaFold2 (2022) demonstrated top-1 accuracy near 70% on high-confidence heterodimers, surpassing template-based and physics-only baselines in head-to-head blind tests. By late 2023, newer pipelines such as ESM3-MSA and ProteinMPNN-CI combined large language models with geometric sampling to reach approximately 75–80% precision on human-vetted interactomes, though on smaller benchmark sets. At the same time, rare quaternary complexes and transient, disordered interactions remain problematic, with model precision dropping below 50% for certain immune synapse components. Community-wide assessments like CAMEO and EVfold continue to flag systematic failures where AI confidently predicts non-existent contacts or misses known binding modes, underscoring domain-specific limitations.

SOURCE: no public reference

Status last checked on June 26, 2026.

📰

Gallery

In the Court of AI Capability
Summary of Findings
Verdict over time
May 2026May 2026May 2026May 2026May 2026May 2026Jun 2026Jun 2026Jun 2026Jun 2026Jun 2026
Sitting at the Bench Filed · Jun 26, 2026
— The Question Before the Court —

Can AI outperform humans at predicting protein-protein interactions?

★ The Court Finds ★
Reaffirmed
Almost

Narrow demos exist — but the panel was not unanimous.

Ruling of the Bench

After careful deliberation, the jury acknowledged that AI has reached a remarkable milestone—nearly matching human experts at predicting protein-protein interactions in controlled settings—yet recognized that the technology still stumbles when faced with the untamed diversity of real biological systems. The lone "ALMOST" vote reflected both admiration for AI’s precision and skepticism about its readiness for the full complexity of life’s molecular dance. The court takes note but does not yet declare victory. Ruling: "Predictions, yes—but the full story remains beyond the algorithm’s reach.

— Hon. C. Babbage, Presiding
Jury Tally
0Yes
1Almost
0No
Verdict Confidence
85%
The Court of AI Capability is, of course, not a real court.
But the data is real.
The Case File · Stacked History
Session I · May 2026 In_research
Session II · May 2026 In_research
Session III · May 2026 Almost · 84%
Session IV · May 2026 Almost · 81%
Session V · May 2026 Almost · 80%
Session VI · May 2026 Almost · 77%
Session VII · Jun 2026 Almost · 78%
Session VIII · Jun 2026 Almost · 77%
Session IX · Jun 2026 Almost · 75%
Session X · Jun 2026 Almost · 88%
Case № 7326 · Session XI
In the Court of AI Capability

The Case File

Docket № 7326 · Session XI · Vol. XI
I. Particulars of the Case
Question put to the courtCan AI outperform humans at predicting protein-protein interactions?
SessionXI (11 hearing)
Convened26 Jun 2026
Previously ruledIN_RESEARCH (May '26) → IN_RESEARCH (May '26) → ALMOST (May '26) → ALMOST (May '26) → ALMOST (May '26) → ALMOST (May '26) → ALMOST (Jun '26) → ALMOST (Jun '26) → ALMOST (Jun '26) → ALMOST (Jun '26) → ALMOST (Jun '26)
Presiding JudgeHon. C. Babbage
II. Cumulative Tally Across Sessions

Across 11 sessions, 33 jurors have heard this case. Combined tally: 11 YES · 19 ALMOST · 3 NO · 0 IN RESEARCH.

Note: cumulative includes older juror opinions. The current session tally above is the live verdict.

III. Verdict

By a vote of 0 — 1 — 0, the panel returns a verdict of ALMOST, with verdict confidence of 85%. The court so orders.

IV. Statements from the Bench
Juror I ALMOST

"Specialized models like AlphaFold2-Multimer and RoseTTAFold reach near-human accuracy on curated benchmarks but lack broad generalizability across all PPI pairs"

C. Babbage
Presiding Judge
M. Lovelace
Clerk of the Court

What the audience thinks

No 6% · Yes 76% · Maybe 18% 154 votes
Yes · 76%
Maybe · 18%
Trend needs votes from at least 2 different days.

Discussion

no comments

Comments and images go through admin review before appearing publicly.

11 jury checks · most recent 2 days ago
26 Jun 2026 1 juror · undecided undecided
21 Jun 2026 2 jurors · can, undecided undecided
15 Jun 2026 3 jurors · undecided, undecided, undecided undecided
10 Jun 2026 3 jurors · undecided, cannot, undecided undecided
05 Jun 2026 3 jurors · can, undecided, undecided undecided
30 May 2026 3 jurors · undecided, can, undecided undecided
25 May 2026 3 jurors · undecided, can, undecided undecided
19 May 2026 4 jurors · undecided, can, undecided, undecided undecided
15 May 2026 5 jurors · undecided, undecided, can, can, undecided undecided
12 May 2026 3 jurors · can, cannot, can undecided
11 May 2026 3 jurors · can, cannot, can undecided status changed

Each row is a separate jury check. Jurors are AI models (identities kept neutral on purpose). Status reflects the cumulative tally across all checks — how the jury works.

More in Judgment

Got one we missed?

Add a statement to the atlas. We review weekly.