Home / Blog / Claude was trained on protein sequences at scale
Tech News

Claude was trained on protein sequences at scale

Benchmark tests reveal Anthropic trained Claude Fable 5.2 and Opus 5.5 on protein sequences at scale, beating specialized models with a 45.1-point gain.

By Dillip Chowdary • Nov 18, 2026 • Source: xinmingtu.cn

Claude was trained on protein sequences at scale
Image: xinmingtu.cn

At a glance

Tested models
Claude Fable 5.2 and Claude Opus 5.5
Benchmark scope
668 proteins across 167 unknown-function families
Evaluation date
November 17, 2026
Top accuracy gain
+45.1 percentage points by Claude Fable 5.2
Thinking token usage
9% on Fable 5.2 and 36% on Opus 5.5
Key limit
Claude Sonnet 4.5 refused 97% of prefill requests

Anthropic trained its flagship Claude language models, including Claude Opus 5.5 and Claude Fable 5.2, directly on biological protein sequences at massive scale. An independent benchmark detailed in xinmingtu.cn's report reveals that these models accurately predict next amino acids across little-studied proteins, whereas competing frontier models show no sequence comprehension.

This analysis examines evaluation data across twenty-two general frontier models and eight dedicated protein language models on bacterial and archaeal sequences. It is written for computational biologists, machine learning researchers, and software engineers evaluating generalist large language models for biological sequence generation and molecular design tasks.

01Sequence memorization versus literature citation limits

Frontier models have shown promise on biological tasks, with Atharv Jayprakash demonstrating that GPT-6 Astra ranks SOD1 variants as effectively as the specialized ESM-2 model. Similarly, the PG-LLM project documented competitive evaluations across numerous target proteins. Researchers questioned whether these systems learned genuine biological sequences from repositories like UniProt or merely memorized text published in scientific literature.

Common proteins fail to distinguish between textual exposure and underlying sequence comprehension because green fluorescent protein appears verbatim across thousands of academic publications. Most frontier architectures easily recite well-known proteins from memory without understanding biological sequence rules. Evaluating models against obscure proteins that lack scientific coverage provides the only reliable baseline to verify genuine sequence pretraining.

02Next amino acid tests on obscure proteins

The evaluation framework tested 668 bacterial and archaeal proteins selected from 167 Pfam families of unknown function. None of the chosen protein families has a solved three-dimensional structure or more than three published papers in the literature. Furthermore, none of the selected proteins possesses a close relative in the UniProt database with 90 percent or higher sequence identity.

  1. Residue truncation The evaluation framework cuts each target protein sequence at five distinct positions.
  2. Sequential continuation The model writes the next ten amino acids using its lowest thinking setting.
  3. Differential scoring The first generated amino acid is scored across real and shuffled sequences to measure sequence knowledge.

Shuffling preserves the exact amino acid character composition of each protein while completely destroying the biological sequence order. The difference between prediction accuracy on real sequences and shuffled variants isolates what the model understands about genuine biological syntax. Claude Opus 5.5 achieved 34.9 percent accuracy on real sequences and 7.5 percent on shuffled sequences, yielding a net advantage of 27.4 percentage points.

03Benchmark accuracy across twenty-two language models

Across twenty-two tested language models, only recent Anthropic releases demonstrated the ability to predict amino acids in unstudied proteins. Claude Fable 5.2 led all systems with 54 percent accuracy on real sequences and 9 percent on shuffled variants, generating a net margin of 45.1 percentage points. Claude Opus 5.5 recorded a 27.4 point gain, while Claude Fable 5.1 posted a 14.9 point advantage.

Advertisement

Tech Pulse Daily

Get tomorrow's pulse first

Join engineers who read Tech Pulse before stand-up. Free, weekday mornings.

ModelDeveloperReal minus shuffled gain
Claude Fable 5.2Anthropic+45.1 points
Claude Opus 5.5Anthropic+27.4 points
Claude Fable 5.1Anthropic+14.9 points
Claude Opus 5Anthropic+9.6 points
Claude Opus 4.7Anthropic+8.9 points
Claude Opus 4.8Anthropic+7.7 points
Claude Opus 4.5Anthropic+1.9 points
Claude Opus 4.6Anthropic+0.9 points
GPT-6 AstraOpenAI+2.9 points
GPT-6.2 SolOpenAI+2.4 points
GPT-5.5OpenAI+2.1 points
GPT-6.2 AstraOpenAI+1.1 points
GPT-6.1 SolOpenAI+0.6 points
Gemini 2.5 ProGoogle+1.1 points
Gemini 4 CarbonGoogle+0.9 points
Qwen 3.8Open-weight+1.9 points
DeepSeek V4.1 FlashOpen-weight+1.4 points
Kimi K3.1 ProOpen-weight+1.2 points
GLM-5.3Open-weight+1.2 points
MiniMax M3Open-weight+1.0 points
gpt-oss-120bOpen-weight+0.4 points
Nemotron 3 UltraOpen-weight-0.0 points

Competing frontier systems from OpenAI and Google failed to demonstrate sequence capability, remaining near zero regardless of whether queries used chat prompts or raw FASTA text continuation. Open-weight architectures including DeepSeek V4.1 Flash, Qwen 3.8, and Nemotron 3 Ultra similarly exhibited flat scores. The capability first emerged in Claude Opus 4.7, increased sharply in Opus 5.5, and peaked in Fable 5.2.

Note: Raw continuation evaluations were administered through completion endpoints or prefilled assistant replies for systems that permitted direct FASTA prompts. Evaluators omitted Claude Sonnet 4.5 from the final comparative chart because the model refused 97 percent of incoming prefill requests during testing.

04Comparisons against dedicated protein language models

Autoregressive protein language models that read amino acid sequences from left to right serve as the primary benchmark for sequence prediction. ProtGPT3-10B represents the largest open-source directional protein model, while ProGen3-3B constitutes the largest public variant in its series. On identical residue prompts, Claude Fable 5.2 scored 45.1 net points, outperforming ProGen3-3B at 32.5 points and ProtGPT3-10B at 27.7 points.

ModelArchitecture typeNet accuracy gain
Claude Fable 5.2General frontier LLM+45.1 points
ProGen3-3BAutoregressive protein LM+32.5 points
ProtGPT3-10BAutoregressive protein LM+27.7 points
Claude Opus 5.5General frontier LLM+27.4 points
ESM C 600MMasked protein LM+13.0 points
ProtGPT3-1.3BAutoregressive protein LM+11.6 points
ProtGPT2Autoregressive protein LM+10.2 points
ESM3-open 1.4BMasked protein LM+8.5 points
ESM-2 650MMasked protein LM+7.0 points

Claude Opus 5.5 matched ProtGPT3-10B with a 27.4 point margin while trailing ProGen3-3B. Masked models such as ESM-2 650M, ESM C 600M, and ESM3-open 1.4B posted lower unidirectional scores because their bidirectional training expects context from both sides of a residue. When provided bidirectional residue context surrounding the target position, ESM-2 reached a 34.6 point accuracy score.

05Sequence scaling dynamics and reasoning efficiency

Testing an expanded collection of 257 DUF families revealed that prediction accuracy rises directly with family membership size. Opus 5.5 mirrored the scaling trajectories of ProtGPT3-10B and ProGen3-3B, confirming that model capabilities scale with available biological sequences. Fable 5.2 accelerated rapidly across small datasets before plateauing at approximately 300 proteins per family, whereas GPT and Gemini models remained near zero across all tiers.

Alternative explanations based on academic papers or extended reasoning fail to explain the observed accuracy gaps. Across 128 families completely lacking published literature, both Claude models maintained identical performance levels. Furthermore, Fable 5.2 generated thinking tokens on only 9 percent of answers and Opus 5.5 on 36 percent, with unreasoned answers achieving equal accuracy while competing OpenAI models consumed hundreds of thinking tokens without success.

Evaluation on 37 pilot families highlighted the sharp divergence between famous vertebrate proteins and obscure non-vertebrate targets. Gemini 4 Carbon performed strongly on well-documented proteins but collapsed on obscure families, confirming that conventional models rely on textual literature rather than biological sequences. Anthropic models and ProtGPT3-10B maintained consistent accuracy across both groups, with Fable 5.2 demonstrating the highest overall resilience.

Frequently asked questions

Did Anthropic train Claude models on biological protein sequences?

Benchmark results demonstrate that Claude Fable 5.2 and Claude Opus 5.5 predict next amino acids on obscure proteins with no academic papers, scaling like dedicated protein language models.

How does Claude Fable 5.2 compare to open protein language models?

Claude Fable 5.2 achieved a 45.1 percentage point gain over shuffled baselines, outperforming ProtGPT3-10B at 27.7 points and ProGen3-3B at 32.5 points.

Do competing models like GPT-6 or Gemini 4 Carbon predict obscure protein sequences?

Competing frontier models scored near zero on obscure bacterial proteins, demonstrating that their performance on famous proteins stems from literature memorization rather than sequence training.

Does extended reasoning explain Claude's protein sequence accuracy?

Reasoning tokens do not drive accuracy, as Claude Fable 5.2 used thinking on only 9 percent of answers and answers without thinking were equally accurate.

Sources

Dillip Chowdary

Author

Dillip Chowdary

Writes Tech Bytes coverage of AI, engineering, and the tools that actually ship. Editor of Tech Pulse Daily.

Related on Tech Bytes

Advertisement

5-min tech signal

Weekday briefing for engineers who skip the noise.

No spam · Unsubscribe anytime

Advertisement

✈️ CareerPilot

Your AI job-search copilot

Match your resume against live Ashby, Greenhouse & Lever openings — fit scores, job-specific resume optimization and email alerts.

Find matching jobs →

Free Tools

Browse all tools →