Ainnocence Achieves New Milestone with Protein Foundation Models

The baseline numbers are from the July 2025 ProteinGym zero-shot leaderboard; Ainnocence values are few-shot results reported below.

Proprietary sequence-first model rivals and in key areas surpasses the leading ProteinGym sequence-only baselines

SAN FRANCISCO, CA, UNITED STATES, August 23, 2025 /EINPresswire.com/ -- Ainnocence, a pioneer in generative AI for biotechnology, today announced the successful internal benchmarking of its protein language model, AINN-P1, trained on UniRef sequences with up to 167 million parameters.

AINN-P1 is Ainnocence’s protein language foundation model, built to understand protein sequence patterns to power next-generation drug discovery. The new model posts state-of-the-art Spearman correlation scores across four core protein-fitness tasks, reinforcing Ainnocence’s sequence-first drug-discovery platform and underscoring its competitiveness against much larger transformer models from Meta and BioMap. 

Highlights:
• Top overall sequence-only score: With an average ρ = 0.441, AINN-P1 edges past ProSST (0.438) while using orders-of-magnitude fewer parameters than 100 B-scale transformers.
• Best-in-class stability prediction: A ρ = 0.625 on stability is the highest of any published sequence-first model, providing critical accuracy for manufacturability screens.
• Closing the binding gap: AINN-P1’s binding ρ = 0.390 exceeds ESM2-150M by 0.064 points, narrowing the historic advantage of structure-aware models.

Detailed benchmark results
Recent in-house evaluations of the AINN-P1 performed in a few-shot embedding  regime for efficiency, produced the following Spearman correlations: Activity 0.3581, Binding 0.3901, Expression 0.3913, Stability 0.6251.  These metrics match or exceed many transformer-based benchmarks, demonstrating that smart architecture and curated data can beat brute-force scale.

Crucially, these benchmarks were achieved without expensive full fine-tuning. Instead, fixed AINN-P1 embeddings feed lightweight regressors – a strategy that slashes compute costs while preserving accuracy, and aligns with emerging best practices when only limited experimental data are available. 

Context within the protein-AI landscape
Meta’s ESM family and BioMap’s 100 B-parameter xTrimo model have advanced the field dramatically, yet their training budgets remain prohibitive for most labs. Ainnocence’s leaner AINN-P1 delivers comparable accuracy without billion-scale parameters or MSAs, validating a sequence-only paradigm for broad adoption. 

“Sequence is the new paradigm,” said Dr. Lurong Pan, CEO. “By learning the language of proteins, we predict complex properties without relying on structures or wet-lab screens. These results rival Big Tech’s best while remaining cost-efficient.” 

Integration & impact on Ainnocence’s platform
The upgraded model is now being rolled into SentinusAI^® (antibody engineering), CarbonAI^® (small-molecule & PROTAC design) and CellulaAI^® (cell-therapy optimization). Higher-fidelity predictions of stability and expression accelerate antibody lead-opt studies, while enhanced activity/binding scores improve small-molecule hit selection – compressing discovery timelines from years to weeks. 

Ainnocence invites research groups and biopharma partners to leverage its sequence-first AI platform for protein engineering, vaccine work and hard-to-drug targets. Contact service@ainnocence.com or visit www.ainnocence.com for collaboration details. 

About Ainnocence
Founded in 2021, Ainnocence is a next-generation biotech company whose self-evolving AI platform can virtually screen 10¹⁰ protein sequences or small-molecule candidates for multitarget and multi-objective optimization, optimizing multiple properties simultaneously to deliver high-probability leads with unprecedented speed and cost efficiency. 

Lurong Pan, PhD
Ainnocence
+1 205-249-7424
service@ainnocence.com
Visit us on social media:
LinkedIn
YouTube

Legal Disclaimer:

EIN Presswire provides this news content "as is" without warranty of any kind. We do not accept any responsibility or liability for the accuracy, content, images, videos, licenses, completeness, legality, or reliability of the information contained in this article. If you have any complaints or copyright issues related to this article, kindly contact the author above.