이미지: X — 프론티어랩
Summary
- From a single prompt written by a human expert, Claude autonomously designed binding molecules for 14 of 15 target proteins
- Adaptyv Bio and Twist Bioscience independently synthesized and tested the designs, finding 22-35% actually bound successfully, surpassing the industry average of 10-15%
- Anthropic is using this as a springboard to continue automating the entire drug discovery process, from antibodies to small molecules
- 설계 성공 표적 수
- 15개 표적 중 14개에 결합 단백질 설계
- 검증 파트너
- Adaptyv Bio, Twist Bioscience (독립 제작·검증)
- 업계 평균 성공률
- 10~15%
- 클로드 설계 성공률
- 구성에 따라 22~35%
- 사용 모델
- Opus 4.8, Mythos Preview
- 대표 결합력 사례
- EGFR 1.7pM, VEGF-A 1.6pM, TREM2 1.1pM (Mythos Preview 설계)
- 생명과학 연구용 최상위 모델
- Opus 5
Claude's protein designs put to the test in the lab
Anthropic disclosed via an X post that it had its model Claude take on "protein binder design," the first gateway task in drug discovery. Using a single protein-design prompt written by a human, Claude designed brand-new proteins from scratch (de novo) capable of binding to 14 of 15 target proteins. Anthropic did not verify these designs itself. Third-party labs Adaptyv Bio and Twist Bioscience independently synthesized the proteins and tested whether they actually bound.
Why protein binder design is the first gateway in drug discovery
Most drugs work by attaching to a specific target in the body and blocking or altering its function. Designing a molecule that binds precisely to that target is the starting point of drug discovery, and until now, experts have had to spend weeks to months per target sifting through countless candidates. Protein binder design is an easier task than actual drug design, but it serves as a useful proving ground for gauging how well AI performs at this stage. Currently in this field, the success rate for human-designed binders that actually achieve binding is generally known to be around 10-15%.
Claude's design capability, measured by success rate
According to figures Anthropic disclosed, Claude's designs showed success rates ranging from 22% to 35% depending on configuration. Across all 15 targets combined, Opus 4.8 achieved 88 out of 390 (about 22.6%) using a multi-target approach, while the preview model Mythos Preview achieved 104 out of 390 (about 26.7%) with the same approach. In the single-target approach, which focused on one target at a time, Mythos Preview pushed the success rate up to 158 out of 450, or about 35.1%.
| Configuration | Success rate | Bar |
|---|---|---|
| Industry average (conventional methods) | 10-15% | 12 |
| Opus 4.8 (multi-target) | 22.6% | 23 |
| Mythos Preview (multi-target) | 26.7% | 27 |
| Mythos Preview (single-target) | 35.1% | 35 |
Results varied widely by target. TREM2 showed high success rates of 76-83% across all three configurations, while targets such as 15-PGDH and MBP stayed at just 0-1%. In terms of Kd values (a measure of binding strength, where lower is stronger), Mythos Preview's designed binders recorded 1.7pM for EGFR, 1.6pM for VEGF-A, and 1.1pM for TREM2, with some cases binding several times more strongly than previously published top-performing de novo binders.
Hurdles that remain
Anthropic itself drew a clear line around these results. Protein binders are not drugs. Creating a molecule that binds strongly to a target is merely the first step in developing a drug candidate, and there are far more steps remaining before that candidate can be proven safe and effective for humans. Anthropic said it is using this result as a springboard to train Claude to carry out the entire drug discovery process end-to-end, from antibodies to small molecules. The company also said it would soon unveil an access program letting scientists use its most capable model, noting that Opus 5 is currently its best model for life science research. Anthropic also open-sourced the experimental prompts and data.
Editor's view
The most notable point in this announcement isn't the success-rate figures themselves but the "independent verification" process behind them. Had Anthropic measured the binding rates itself, the results would have been hard to trust — but because third-party labs Adaptyv Bio and Twist Bioscience directly synthesized and tested the proteins, these figures have effectively passed verification outside the lab that produced them. At a time when AI model companies routinely tout their superiority through self-run benchmarks, this kind of external verification is likely to become the standard for credibility in fields like biology where physical experimentation is unavoidable.
There's another interesting angle when comparing generations. A model called "Mythos Preview" appearing in the table showed higher success rates and stronger binding than the existing Opus 4.8 across multiple targets. It appears to be an unreleased preview model, offering a clue as to how the next generation of Claude might perform in life sciences. As AI models continue to be attached to lab workflows beyond coding or writing, the same narrative keeps repeating — tasks that once took humans weeks are cut down to hours. This case is no different from that pattern.
That said, a dose of practical caution is warranted. It would be premature for Korea's bio-pharma industry to rush toward using AI to pick drug candidates based on this result alone. Protein binder design is just the first of dozens of steps in drug discovery, and verifying safety and efficacy requires time and cost on a scale incomparable to this experiment. At this point, the most practitioners can reasonably do is consider such AI design pipelines as an auxiliary tool for early-stage candidate screening — not as a stage ready to replace the overall process.
In the coming months, Anthropic's promised scientist access program is likely to take shape, and follow-up results expanding the experimental scope to other drug types such as antibodies and small molecules are expected to emerge. This announcement is likely to accelerate competition among AI models in life sciences to a level rivaling that seen in coding and image generation.
