Bio AI Agent: Autonomous multi-agent intelligence — accelerating antibody therapeutics from bench to clinic

Yi Ni^*, Liwei Zhu, Qi Yan

Bio LIMS INC, Boston, Massachusetts, United States

^*Corresponding author: yi.ni@biolims.net

ChAbS 2026 Annual Conference — Chinese Antibody Society

May 9–10, 2026

Boston Marriott Cambridge, Cambridge, MA

Poster

Abstract

Antibody therapeutic development remains constrained by 10–15 year timelines, $2.6B+ per-approval costs, 90%+ candidate attrition, and fragmented workflows across target validation, antibody engineering, developability assessment, and regulatory planning. Existing AI tools address isolated stages but lack end-to-end integration with autonomous reasoning, cross-domain memory, and dynamic decision-making. We present Bio AI Agent, an LLM-powered agentic system with dynamically extensible Bio AI Agent Skills, a persistent four-layer memory architecture, and vector knowledge retrieval that orchestrates six specialized agents — Target Discovery, Antibody Engineering, Developability Assessment, Safety & Immunogenicity, Clinical Translation, and Decision Synthesis — across the entire antibody development pipeline. In a retrospective anti-HER2 antibody optimization case study, the system autonomously executed target selection, CDR grafting onto optimal human germline scaffolds (95% humanization with <2-fold affinity loss), in silico affinity maturation (K_D 8.2 nM → 0.4 nM), and developability remediation (T_m +8.5°C, viscosity <15 cP at 150 mg/mL, 65% immunogenicity reduction). Compared with prior multi-agent architectures and conventional workflows, Bio AI Agent compresses end-to-end cycles from 14–27 months to 3–5 weeks, delivers ~200× analysis acceleration, 87% cost reduction, and 250× token efficiency, while preserving FDA-aligned human-in-the-loop oversight and a full audit trail for regulatory inspection.

Six-agent architecture for antibody development

Target Discovery Agent

Multi-parametric target scoring across >5,000 surface antigens — druggability, genetic association, expression selectivity, pathway essentiality, competitive landscape. Epitope accessibility and tumor-vs-normal profiling via GTEx (54 tissues) and HPA.

Antibody Engineering Agent

CDR grafting and framework humanization with germline alignment scoring. VH/VL pairing optimization via AbLang embeddings and ESM fitness landscapes. In silico affinity maturation through structure-guided mutation scanning, AlphaFold modeling, and DiffDock binding pose validation with ΔG estimation.

Developability Assessment Agent

Multi-parameter profiling — aggregation (SAP/SCM), viscosity, thermal stability (T_m/T_agg), charge variants, polyreactivity. CMC-aware sequence optimization for isomerization hotspots, oxidation-prone residues, N-glycosylation, and PTM liability removal.

Safety & Immunogenicity Agent

MHC-II binding prediction via NetMHCIIpan for T-cell epitope ID; deimmunization through germline-proximal substitution without affinity loss. Tissue cross-reactivity via GTEx/HPA and FDA FAERS pharmacovigilance mining across >50M adverse event reports.

Clinical Translation Agent

FDA and EMA regulatory pathway alignment, IND-enabling study design, GMP manufacturability and CMC gap analysis, automated regulatory document drafting. FTO patent analysis across USPTO, EPO, and WIPO; competitive trial monitoring; biomarker-guided stratification.

Decision Synthesis Agent

Synthesizes all agent outputs into executive summaries, risk/benefit matrices, and phased action plans (1–3, 3–6, 6–12 months). Cross-domain conflict resolution balancing affinity, developability, and immunogenicity trade-offs, with confidence-scored Go/No-Go recommendations.

Case study — anti-HER2 antibody optimization

Target discovery

HER2 prioritized with selectivity score 0.92 across breast, gastric, and colorectal carcinomas via multi-omics analysis.

Humanization

CDR grafting onto VH3-23 / VK1-39 germline; 95% humanization achieved with <2-fold affinity loss.

Developability triage

Aggregation-prone motif in CDR-H3 flagged (SAP > threshold); three stabilizing variants auto-designed.

Affinity maturation

K_D improved from 8.2 nM to 0.4 nM via structure-guided scanning — 12 positions, 240 variants evaluated in silico.

Stability and formulation

T_m increased by 8.5°C after removing isomerization and oxidation hotspots; viscosity <15 cP at 150 mg/mL.

Immunogenicity reduction

T-cell epitope removal lowered immunogenicity risk by 65% with full binding activity preserved.

Headline performance

Analysis acceleration

~200×

vs. manual workflow

Development cycle

3–5 wk

from 14–27 months

Time saved

85%+

end-to-end per candidate

Cost reduction

87%

per antibody program

Token efficiency

250×

vs. prior multi-agent architecture

Surface antigens covered

>5,000

target discovery scope

Key contributions

• Autonomous multi-agent intelligence for integrated antibody therapeutic development — target discovery, antibody engineering, developability optimization, safety profiling, and regulatory planning in a single orchestrated workflow.

• Persistent four-layer memory architecture enabling cross-session learning — the system accumulates institutional expertise with each antibody program, progressively improving accuracy and speed on subsequent candidates.

• Dynamically extensible Bio AI Agent Skills that invoke enterprise system connectors via standardized APIs, allowing the platform to adapt to new analytical tools and data sources without rearchitecting the agents.

• Validated identification of developability liabilities — aggregation, immunogenicity, tissue cross-reactivity — with actionable mitigations that target late-stage clinical attrition before it happens.

• End-to-end cycle compression from 14–27 months to 3–5 weeks with 87% cost reduction and 250× token efficiency relative to prior multi-agent architectures, demonstrated on a complete anti-HER2 optimization case.

Regulatory posture

The framework is aligned with the FDA Discussion Paper (2023) and CDER Guidance (2025) on AI/ML in drug discovery, preclinical evaluation, and regulatory submission support. A human-in-the-loop (HITL) framework requires expert review on every AI recommendation before advancement to experimental or regulatory stages. Every agent decision, data source, and reasoning chain is logged for reproducibility and regulatory inspection.

Keywords

Antibody therapeuticsMulti-agent AILarge language modelsPersistent memoryAntibody engineeringDevelopabilityImmunogenicityHumanizationAffinity maturationHER2FDA AI/ML guidance

How to cite

Ni Y, Zhu L, Yan Q. Bio AI Agent: Autonomous multi-agent intelligence — accelerating antibody therapeutics from bench to clinic. Poster presented at: 2026 Annual Conference of the Chinese Antibody Society (ChAbS); May 9–10, 2026; Boston Marriott Cambridge, Cambridge, MA, USA