• Blog
  • /
  • 2025 AI & Biotech Annual Insight Report: A Curated Digest of arXiv and Leading Scientific Research
Biotech Healthcare

2025 AI & Biotech Annual Insight Report: A Curated Digest of arXiv and Leading Scientific Research

2025 overwhelmed the industry with giga-scale breakthroughs – from foundation models that understand cellular biology to fully robotic, AI-driven laboratories. This report synthesizes key findings from top journals and arXiv to separate the noise from the signal. Discover the specific technologies that are moving AI-designed drugs from digital prototypes to clinical reality.

by Gary Martin
13 min read Dec 15, 2025
226 views

Table of Content

2025 overwhelmed biotech with giga-scale breakthroughs: foundation models that understand cellular biology, fully robotic AI-driven laboratories, and the first real AI-designed drugs approaching clinical phases.

Yet the information noise was so intense that even R&D directors struggled to keep up.

As your technology partner, we prepared a 2025 Annual Scientific Report – drawing from arXiv, Nature, Cell, Science, bioRxiv, medRxiv, and leading industry sources.

Foundation Models & Multi-Omics

Towards Multimodal Foundation Models in Molecular Cell Biology

Key facts:

  • Multimodal biological datasets are growing at >35% CAGR, with single-cell atlases now exceeding 100M profiled cells worldwide (Human Cell Atlas, Chan Zuckerberg Initiative).
  • The ability to jointly embed omics modalities reduces batch-effect variability by up to 40-60%, according to benchmarking in related integrative models.
  • Cross-modality prediction accuracy (e.g., predicting chromatin accessibility from RNA) improved by ~20-30% compared to unimodal baselines.

Implication:
Multimodal foundation models are becoming an operating system for modern biology, supporting downstream tasks from target identification to patient stratification and mechanism-of-action inference.

Source: Cui, H. – Nature – 2025

CellFM: A 100-Million-Cell Foundation Model for Single-Cell Omics

Key facts:

  • Rare cell populations (<0.1%) often go undetected with classical clustering; CellFM improves recall of rare cell types by up to 2.3×.
  • Gene-signature prediction accuracy improves by 15-25% over existing single-cell integration frameworks.
  • The global single-cell sequencing market surpassed 2 billion cells profiled annually in 2024, making such FM-scale models technically and economically necessary.

Implication:
Large single-cell foundation models deepen our understanding of micro-niches and rare cell states, enabling more precise therapeutic targeting and biomarker discovery.

Comparison of gene function prediction performance in a zero-shot setting:  (a) Accuracy (ACC) and (b) Macro-F1 scores for CellFM vs. other single-cell foundation models on binary classification tasks.  (c) UMAP visualizations of gene embeddings generated by CellFM, scGPT, and Geneformer.  (d) AUPR values for predicting multiple Gene Ontology (GO) functions (MF, CC, BP), showing CellFM’s performance.
Comparison of gene function prediction performance in a zero-shot setting:  (a) Accuracy (ACC) and (b) Macro-F1 scores for CellFM vs. other single-cell foundation models on binary classification tasks.  (c) UMAP visualizations of gene embeddings generated by CellFM, scGPT, and Geneformer.  (d) AUPR values for predicting multiple Gene Ontology (GO) functions (MF, CC, BP), showing CellFM’s performance.

Source: Zeng, Y. – Nature Communications – 2025

Visual-Omics Foundation Model Unifying Histopathology and Omics

Key facts:

  • Histopathology is a 4-petabyte/year data domain globally; integrating it with omics allows computational pathology to achieve up to 90% accuracy for certain tissue-level predictions.
  • Image-to-omics mapping reduces experimental sequencing needs by ~30-50% for some workflows (e.g., tumor microenvironment profiling).
  • Spatial transcriptomics costs dropped by >5× since 2020, making multimodal FM training increasingly feasible.

Implication:
Triple-modal FMs open the door to automated drug repurposing, treatment-response prediction, and spatial-target discovery, connecting morphology, molecular states, and text-based biological knowledge.

Source: Chen, W. – Nature Methods – 2025

Related Case Study
Pharma & Biotech
Explore the Multi-Omics AI Platform Case

Our team built a multi-omics AI platform that generates drug candidates with optimized physicochemical and biological properties. See how AI accelerates drug, target, and indication discovery.

Explore the Full Case

Nicheformer: Foundation Model for Spatial + Single-Cell Omics

Key facts:

  • Spatial omics adoption is growing at ~28% annually, with >500 published datasets as of 2025.
  • Niche-level interactions explain up to 60% of variance in immune response in certain inflammatory conditions (based on benchmarking vs. single-cell-only models).
  • Spatial proximity modelling improves prediction of ligand-receptor interactions by ~1.5×.

Implication:
Therapeutic targeting is shifting from a “one cell type” paradigm to a “one niche” approach, which is crucial for oncology, fibrosis, immunology, and tissue regeneration.

Overview of the SpatialCorpus-110M collected for training Nicheformer. (a) The dissociated single-cell dataset includes 57.06M human and mouse cells across 17 organs, 18 cell lines, blood, bone elements, and additional tissues, grouped by major organ systems for visualization. (b) The spatial transcriptomics dataset includes 53.83M spatially resolved cells from humans and mice, collected using four profiling technologies across 15 solid organs. (c) All 110M cells in SpatialCorpus-110M were annotated using a harmonized metadata schema (Nicheformer), unified at both the gene and cell levels depending on modality.
Overview of the SpatialCorpus-110M collected for training Nicheformer. (a) The dissociated single-cell dataset includes 57.06M human and mouse cells across 17 organs, 18 cell lines, blood, bone elements, and additional tissues, grouped by major organ systems for visualization. (b) The spatial transcriptomics dataset includes 53.83M spatially resolved cells from humans and mice, collected using four profiling technologies across 15 solid organs. (c) All 110M cells in SpatialCorpus-110M were annotated using a harmonized metadata schema (Nicheformer), unified at both the gene and cell levels depending on modality.

Source: Tejada-Lapuerta et al. – Nature Methods, Nicheformer – 2025

Review: Leading AI-Driven Drug Discovery Platforms

Key facts:

  • As of 2025, the global landscape includes >40 AI-designed drug candidates in clinical development (Phase I-III).
  • Generative chemistry platforms claim 10-100× faster molecule design cycles compared to classical methods.
  • AI-assisted hit-identification improves hit-rate by 20-70%, depending on assay type.

Implication:
AI-driven drug discovery is no longer anecdotal; it is a portfolio-level capability, influencing timelines, probability of success, and pipeline economics.

Source: Dharmasivam, M. – Pharmacological Reviews – 2025

AI for Science Strategy: Government-Level Adoption of AI-Biology

Key facts:

  • Government programmes (UK, US, EU) allocated over £1.2B globally in 2024-2025 to AI-for-science infrastructure.
  • ML-accelerated molecular design pipelines have reduced early discovery timelines from 2-3 years to 6-12 months in documented public-private collaborations.
  • Image-based high-content screening with AI yields 20-40% higher hit identification rates compared to classical pipelines.

Implication:
At a policy level, AI is being formalized as a standard tool for national-scale R&D, no longer positioned as an experimental add-on but as core scientific infrastructure.

Source: Department for Science, Innovation & Technology – GOV.UK – 2025

Ivan Izonin, PhD Sr Scientific Advisor, Artificial Intelligence
Get Expert Guidance on Implementing AI in Your Biotech Project

Book a consultation with Ivan Izonin to explore how advanced AI methods can be applied to your biotech or biomedical challenges.

Book a Consultation

LLM Agents and Multi-Agent Systems for Biomedicine

Survey: LLM-Based Multi-Agent Systems in Medicine

Key statistics and facts:

  • The number of publications on multi-agent medical AI has increased by >300% between 2021 and 2025, indicating rapid adoption.
  • LLM diagnostic agents achieve 60-80% accuracy on benchmark clinical reasoning datasets (e.g., MedQA, PubMedQA), comparable to junior clinician performance.
  • Workflow-level simulation using agent teams reduces manual review load by 25-40%, according to studies included in the survey.

Implication:
Multi-agent systems are shifting from theoretical constructs to a practical orchestration layer in biomedical workflows.

Source: Lin, Y. – TechRxiv – 2025

Large Language Model Agents for Biomedicine

Key statistics and facts:

  • Studies demonstrate that enabling tool-use can improve LLM task success rates by 30-50% over text-only models.
  • Multi-agent collaboration improves solution robustness by 15-25%, particularly on diagnostic and literature-review tasks.
  • The agentic architecture described aligns with the broader trend of LLM operations frameworks adopted by major AI labs in 2024-2025.

Implication:
Biomedical teams can now design agents using standardized architectural templates, eliminating the need to reinvent complex coordination logic.

Source: Xu, X. – Information – 2025

Key challenges and mitigation strategies for biomedical LLM agents.
Key challenges and mitigation strategies for biomedical LLM agents.

STELLA: A Self-Evolving LLM Agent for Biomedical Research

Key statistics and facts:

  • STELLA demonstrates progressive performance gains of up to 20% on successive biomedical tasks due to self-training.
  • Memory-augmented architectures reduce information loss across sessions by ~35%, improving long-term task continuity.
  • In literature triage, STELLA automates up to 70% of screening decisions, comparable to semi-automated systematic review tools.

Implication:
Self-evolving agents introduce the realistic possibility of a “lab co-PI” AI system that accumulates expertise alongside human teams.

Source: Jin, R. – arXiv – 2025

AI Agents in Drug Discovery

Key statistics and facts:

  • Closed-loop AI-robotics pipelines have demonstrated 2-5× faster experimentation cycles in synthetic biology and medicinal chemistry.
  • Integrated agent workflows reduce handoff latency between pipeline stages by 30-60%.
  • Multi-step reasoning accuracy improves by ~25% compared to single-model baselines.

Implication:
The paradigm is shifting from “one model per task” to “one agent per workflow,” reshaping AI stack design in pharmaceutical R&D.

Drug discovery workflow enabled by Kiin Bio’s Virtual Scientist Platform. Each block represents a distinct research plan within a larger experimental workflow. Plans are tagged with the Virtual Scientist agent responsible, illustrating how specialized agents collaborate to execute and coordinate complex end-to-end discovery processes.
Drug discovery workflow enabled by Kiin Bio’s Virtual Scientist Platform. Each block represents a distinct research plan within a larger experimental workflow. Plans are tagged with the Virtual Scientist agent responsible, illustrating how specialized agents collaborate to execute and coordinate complex end-to-end discovery processes.

Source: Seal, S. – arXiv – 2025

DrugAgent: A Multi-Agent LLM Framework for Drug Discovery

Key statistics and facts:

  • Multi-agent DTI prediction improves mean predictive accuracy by 12-18% over single-model approaches.
  • Automated experiment-planning agents reduce human intervention by up to 40% for routine in-silico tasks.
  • Cross-agent consensus improves error detection rates by ~20% in benchmarking datasets.

Implication:
The multi-agent paradigm covers the entire workflow from hypothesis generation to in-silico evaluation, not just isolated tasks.

Source: Liu, S. – arXiv – 2025

Coated-LLM: Multi-Agent Framework for Alzheimer’s Combination Therapy

Key statistics and facts:

  • Alzheimer’s combination therapy research suffers from <10% availability of high-quality paired datasets; agent-based inference bridges part of that gap.
  • In experiments, Coated-LLM generated novel therapy combinations in 65% of runs that were not present in training data.
  • Graph-based reasoning improved gene-disease link prediction by ~15% over baseline LLMs.

Implication:
AI begins to function where classical data-driven methods fail-specifically, combination therapy design under sparse data conditions.

Inhibitory effects of therapeutic agents on amyloid beta aggregation (A) The aggregation profiles of amyloid beta, both in the absence and presence of various combinations of compounds, are depicted. Error bars represent the standard error of the mean (SEM). (B) The percentage of aggregation is presented to better illustrate the effect of the different compounds.
Inhibitory effects of therapeutic agents on amyloid beta aggregation (A) The aggregation profiles of amyloid beta, both in the absence and presence of various combinations of compounds, are depicted. Error bars represent the standard error of the mean (SEM). (B) The percentage of aggregation is presented to better illustrate the effect of the different compounds.

Source: Xu, Q. – iScience – 2025

Multi-Agent Drug Discovery & Clinical Simulation Pipeline

Key statistics and facts:

  • Portfolio optimization accuracy improves by up to 30% when integrating clinical simulation agents.
  • ADMET prediction agents achieve 10-25% higher accuracy than single-model baselines depending on endpoint.
  • In silico trial simulations reduce the number of necessary physical experiments by ~20-40%.

Implication:
Future R&D workflows increasingly resemble a simulation-driven portfolio management environment, where agents coordinate long-horizon decisions.

Multi-LLM Collaboration for Screening Prioritization

Key statistics and facts:

  • Multi-LLM committees reduce false-negative screening errors by 15-20%.
  • Cost of early-stage candidate selection decreases by 25-35% when automated agents are incorporated into review pipelines.
  • Agreement rates between LLM committees and expert panels exceed 80% on well-defined tasks.

Implication:
Multi-LLM committees are emerging as a credible alternative to the first round of human screening for large hypothesis pools.

Source: Zhao, Y. – medRxiv – 2025

Autonomous Labs and “Self-Driving” R&D Infrastructure

The Rise of Autonomous Labs in Life Sciences

Key statistics and facts:

  • Robotic liquid-handling systems already reduce manual experimental error by up to 70% (Nature Reviews Methods Primers, 2023).
  • Automated experiment-planning with AI can shorten iterative optimization cycles by 3-10×, depending on assay type (MIT/IBM “Bayesian Optimization in Materials & Biology”).
  • The autonomous-lab market is projected to grow at ~25% CAGR through 2030, driven by pharma and synthetic-biology sectors (Grand View Research, 2024).

Implication:
Competition is shifting from “who has the best model” to who has the infrastructure capable of running full R&D cycles autonomously, without human micromanagement.

Source: Sartorius Stedim Spain. (2025, November 13). The rise of autonomous labs in life science. Biotech Spain

Ginkgo’s Autonomous Lab: “Order Experiments by Asking”

Key statistics and facts:

  • Ginkgo’s automated foundry reportedly executes >50,000 experimental workflows per month, one of the highest throughputs globally.
  • Natural-language experiment generation reduces protocol setup time by up to 80%, compared to manual workflow design.
  • Robotics + AI integration increases reproducibility, with batch-to-batch variability reduced by 20-40% across common assays.

Implication:
This system represents a prototype of a “GitHub Actions for biology”: push a hypothesis → the platform automatically runs the corresponding experimental pipeline.

Categories of research tasks that experimentalists do not want to automate (top, 54 total responses), and the reasons why researchers may feel negativity around automation (bottom, 91 total responses, multiple responses could be chosen), partitioned by whether the respondent works primarily in computation or theory ("computa-tional"), and for experimentalists, their self-reported experience with lab automation.
Categories of research tasks that experimentalists do not want to automate (top, 54 total responses), and the reasons why researchers may feel negativity around automation (bottom, 91 total responses, multiple responses could be chosen), partitioned by whether the respondent works primarily in computation or theory (“computa-tional”), and for experimentalists, their self-reported experience with lab automation.

Source: Ginkgo Bioworks – Autonomous Lab – 2025

DoE-Backed Autonomous Platform for Microbial Biotech

Key statistics and facts:

  • The DoE Biological and Environmental Research (BER) program invested over $300M in autonomous biology initiatives between 2022-2025.
  • Autonomous microbial strain-engineering workflows achieve 5-15× faster design-build-test cycles (JBEI/LBNL reports).
  • Microbial screening throughput in autonomous platforms exceeds 100,000 variants per week, far outpacing manual laboratory capacity.

Implication:
Autonomous labs are evolving into national bioeconomy infrastructure, not just pharma R&D tools-shifting innovation capacity to state-level strategic assets.

Source: Cozier, M. – SCI – 2025

ChemLex: Robot-Run AI Drug Discovery Lab in Singapore

Key statistics and facts:

  • Automated medicinal-chemistry labs can accelerate compound synthesis by 3-5× and reduce reagent waste by up to 50% (ACS Central Science, 2024).
  • Robot-assisted drug-screening workflows reduce operational costs by 30-40% compared to manual setups.
  • Singapore’s R&D investment in autonomous laboratories has grown >20% YoY, reinforcing APAC leadership in smart-lab infrastructure.

Implication:
A “self-driving lab” is becoming a company-level competitive advantage, analogous to owning a dedicated supercomputing cluster in the early deep-learning era.

(From left) Economic Development Board managing director Jermaine Loy, Minister of State for Trade and Industry Gan Seow Huang, ChemLex founder and chief executive Sean Lin and A*STAR CEO at a launch ceremony on Dec 5.
(From left) Economic Development Board managing director Jermaine Loy, Minister of State for Trade and Industry Gan Seow Huang, ChemLex founder and chief executive Sean Lin and A*STAR CEO at a launch ceremony on Dec 5.

Source: Subhani, O. – The Straits Times – 2025

Top 100 Labs 2025: The Infrastructure Era

Key statistics and facts:

  • Labs implementing digital-twin simulation report 10-30% reductions in experimental iteration cycles.
  • AI-guided robotics increase throughput in synthetic biology and high-content screening by 2-8×, depending on assay complexity.
  • Over 60% of labs in the Top-100 report integrating at least one autonomous or semi-autonomous workflow component.

Implication:
Organizations must now think not only about their model portfolios but about building an “AI-ready physical laboratory architecture” designed for automation, robotics, and machine-driven experimentation.

Source: R&D World – Top 100 Labs – 2025

AI in Clinical Trials, Protocol Design, and Risk Management

AI in Clinical Trials: The Edge of Tech

Key statistics and facts:

  • Approximately 70% of clinical trial costs stem from patient operations, data collection, and site management (Tufts CSDD). AI-enabled automation directly targets these components.
  • AI-driven eligibility screening can reduce initial protocol deviations by up to 25%, improving trial integrity.
  • Real-time AI monitoring decreases adverse-event detection time from weeks to hours-days, according to multiple digital-health deployments.

Implication:
Clinical trials are becoming data-native environments, where AI informs decisions from dose selection to adaptive protocol modifications.

AI enhances clinical trials through smarter design, monitoring, and data use.
AI enhances clinical trials through smarter design, monitoring, and data use.

Source: Clinical Trial Risk – AI in clinical trials – 2025

What Are AI Clinical Trials

Key statistics and facts:

  • Poor recruitment is responsible for >80% of trial delays, and ~30% of trials fail outright due to recruitment challenges (FDA/NIH data).
  • AI-assisted patient matching can accelerate recruitment by 3-10×, depending on the therapeutic area.
  • Predictive modeling of trial success has reached 70-85% accuracy on retrospective datasets used for design calibration.

Implication:
Recruitment and design are no longer the primary bottlenecks-provided a company has high-quality data and a functional AI stack to leverage it.

Real-Time Data Monitoring scheme.
Real-Time Data Monitoring.

Source: NWAI – AI in clinical trials – 2025

Lifebit: AI-Driven Drug Discovery and AI for Clinical Trials

Key statistics and facts:

  • Integrated AI pipelines reduce discovery-to-clinical transition times by 20-40%, especially in immunology and oncology programs.
  • Cross-stage data harmonization reduces data-cleaning labor by up to 60%, according to internal case studies published by several platform providers.
  • Companies using unified data/AI infrastructures report 30-50% fewer duplicated experiments due to consistent metadata and traceability.

Implication:
A unified data and AI platform across the full drug lifecycle offers a competitive edge compared to fragmented, task-specific tool stacks.

Source: Lifebit – AI-driven drug discovery – 2025

IQVIA: Revolutionizing Clinical Study Design with AI

Key statistics and facts:

  • Adaptive designs supported by AI can reduce required sample sizes by 15-30%, depending on statistical power assumptions.
  • Simulation-assisted protocol design can reduce protocol amendments-one of the costliest trial disruptions-by 20-40%.
  • Each major protocol amendment costs sponsors $500,000 to $2 million, making AI-driven prevention financially significant (Tufts CSDD).

Implication:
Simulated clinical trials are becoming a standard preparatory step before real-world trial execution.

Digitalized protocols enable automated scoring and scenario planning.
Digitalized protocols enable automated scoring and scenario planning.

Source: IQVIA – Revolutionizing clinical study design – 2025

Balancing Innovation and Data Integrity

Key statistics and facts:

  • Over 50% of clinical trial data issues originate from inconsistent site documentation; AI audit tools reduce inconsistency rates by 15-35%.
  • Automated risk prediction models can identify high-risk subjects with 70-90% precision, enabling proactive intervention.
  • Regulatory bodies (FDA, EMA) published five+ AI governance frameworks between 2023-2025, underscoring the need for transparency.

Implication:
Model governance and validation are becoming as essential as performance metrics, particularly as AI transitions from support roles to regulatory-relevant decision systems.

Source: ACRP – Artificial intelligence in clinical trials – 2025

    Interesting? Latest Biotech topics directly to your inbox!

    Latest Biotech topics directly to your inbox!

    Written by

    Gary Martin Head of Biotech AI Partnerships

    at Blackthorn AI

    Discover More
    Related Articles

    All Articles