Appendix B: Changelog

The Life Sciences AI Handbook changes as new research, models, and tools appear. Each day the handbook changed is listed here, newest first, with the chapters involved and the source behind each change.

11 updates in the last 30 days

November 2026

4 days with updates

A 2026 Nature Biotechnology analysis, mostly by Shift Bioscience employees, finds deep genetic-perturbation models can outperform uninformative baselines under calibrated metrics. On OpenAI's BioBasicsBench, GPT-6 Astra reportedly earned 66.7% of rubric credit. Anthropic's guidance adds grader validation checks, and federal policy makes DNA synthesis screening a purchasing condition for funded labs. A preprint finds fastVEP's ACMG/AMP classifier matched ClinVar's direction for 79.3% of two-star variants.

fastVEP ACMG/AMP classifier against ClinVar two-star variants (preprint) 1 update
New evidence

fastVEP's automated ACMG/AMP classifier benchmarked against ClinVar two-star variants

In a preprint, fastVEP's automated ACMG/AMP classifier matched ClinVar's direction for 79.3% of 673,660 two-star variants, leaving most disagreements as uncertain calls.

Why Automated classification at database scale leaves many expert-labeled variants uncertain, so case-specific expert review remains necessary.

Chapter Variant Effect Prediction

Anthropic guidance on building safety evaluations: stress-testing the grader; Federal DNA synthesis screening becomes a purchasing condition for funded labs 2 updates
Policy & regulation

Federal DNA synthesis screening becomes a purchasing condition for funded labs

Federally funded labs must buy synthetic DNA and benchtop synthesizers only from suppliers using NIST-verified screening, with deadlines in July and October 2027.

Why Order screening moves from voluntary practice to a funding condition, so synbio teams need adherent suppliers before the 2027 deadlines.

Chapter Synthetic Biology Design Tools

News & developments

Anthropic guidance on building safety evaluations: stress-testing the grader

Anthropic recommends stress-testing a model grader before trusting it and running safety evaluations in a sandbox with no internet access and real-time monitoring.

Why If a conclusion changes when the grader changes, the evaluation measured the grader rather than the research agent.

Chapter Agentic Science Workflows

OpenAI's BioBasicsBench: AI agents reading routine wet-lab data; RFdiffusion3 published in Science; Danaher announces an AI-powered autonomous antibody lab at Abcam; Connectome-constrained fly brain models 4 updates
New evidence

RFdiffusion3 published in Science

Protein Design and Engineering now cites the Science version of record for RFdiffusion3, an all-atom diffusion model that designs proteins around ligands, nucleic acids, and other non-protein atoms. Two zero-shot campaigns of 96 constructs each yielded active phosphotriesterases and transthyretin-specific zinc proteases; two enzyme classes do not establish general performance.

Why The all-atom design claim now rests on a peer-reviewed paper rather than a developer preprint.

Chapters Protein Design and Engineering · Model and Dataset Index · References

New evidence

Connectome-constrained fly brain models

Neuroscience adds two Nature studies. A connectome-built model of 64 cell types in the fly optic lobe, trained only to detect motion, agreed with neural activity reported across 26 earlier studies. A spiking model of the whole adult fly central brain matched 91% of 164 predictions tested in living flies. POYO-1 is now cited to its NeurIPS 2023 paper rather than the project page.

Why Wiring diagrams can now generate testable predictions about circuits, although simplified neuron models in one insect are hypotheses, not validated brain simulations.

Chapter Neuroscience AI and Brain Foundation Models

News & developments

OpenAI's BioBasicsBench: AI agents reading routine wet-lab data

Benchmarks adds OpenAI's 100-task benchmark, built with Ligo from raw gels, plates, microscopy, and bench videos. GPT-6 Astra reportedly earned 66.7% of rubric credit at the xhigh setting, against 43.8% for GPT-5.6 Sol; vendor-authored and not peer-reviewed.

Why Agents are starting to read raw lab data, and OpenAI's own failure cases show assumptions overriding controls, so their outputs still need a scientist's check.

Chapter Benchmarks for Bio AI

News & developments

Danaher announces an AI-powered autonomous antibody lab at Abcam

Self-Driving Labs adds Danaher's October 2026 announcement of its first AI-powered autonomous lab, based at Abcam, to design and test affinity reagents such as antibodies and feed each result back into the model. It is expected to operate at scale in early 2027; the speed and output figures are company design targets with no baseline or data.

Why The chapter classifies it as an announced design-make-test-learn platform, not demonstrated autonomous discovery, and names the data that would change that.

Chapter Self-Driving Laboratories

Calibrated metrics for single-cell perturbation models 1 update
Revisions

Calibrated metrics for single-cell perturbation models

Cells, Tissues, and Systems reports a 2026 Nature Biotechnology analysis across 14 datasets and 18 metrics arguing that metrics such as mean squared error are frequently miscalibrated and that deep genetic-perturbation models can outperform uninformative baselines under calibrated metrics. Five of its six authors are employees of Shift Bioscience.

Why Whether foundation models beat simple baselines depends on the metric, so the choice rests on evidence for each task rather than on model class.

Chapter Single-Cell Foundation Models

References, links, and wording were also kept current.

October 2026

13 days with updates

Dai et al. in Nature report Pinal, which designs proteins from plain-language descriptions, and RFdiffusion-designed metalloproteases and de novo cysteine proteases show measured catalysis. Miller et al. report calibrated perturbation-model gains, and Bits to Binders finds structure-confidence scores do not predict minibinder function in primary T cells. Rentosertib has Phase 2a safety evidence with a Phase 3 trial underway in China.

Teaching materials for graduate programs, fellowships, and course directors 1 update
New content

Teaching materials for graduate programs, fellowships, and course directors

Teaching with the Life Sciences AI Handbook offers four- and eight-session tracks, a map to the International Society for Computational Biology competency framework version 3 (13 competency statements, A3 to M3), journal club questions for every chapter, case discussions, and assessment items with answer points.

Why Course directors can teach AI for the life sciences in a few sessions against a recognized bioinformatics competency standard.

Chapters Teaching with the Life Sciences AI Handbook · Course Syllabus Template

Anthropic and five partners report agent-guided preclinical design results 1 update
News & developments

Anthropic and five partners report agent-guided preclinical design results

Agent-guided design shortened a CAR signaling domain from 154 to 45 amino acids and scored 34,873 antibody designs, in developer and partner reports.

Why These are partner-reported results, not peer-reviewed studies, and they mark how far language-model agents reach into bounded preclinical design today.

Chapters Agentic Science Workflows · Antibody and Biologic Design · Clinical Trial AI for Translational Research

Dai et al. in Nature: Pinal designs proteins from plain-language descriptions 1 update
New evidence

Dai et al. in Nature: Pinal designs proteins from plain-language descriptions

Pinal writes protein sequences from a plain-language description of the function wanted; the intended function appeared in all four tested design tasks.

Why Describing a protein in words produced working designs on individual tasks, which shows the approach can work but not yet that it generalizes.

Chapter Protein Design and Engineering

A 12-week graduate seminar on AI for biological discovery; AlphaFold 3 access terms 2 updates
New content

A 12-week graduate seminar on AI for biological discovery

A ready-to-teach 12-week graduate seminar with weekly readings, three written assignments, a grading rubric, and a learner AI-use policy.

Why The course teaches evaluation before capability: a model output is a hypothesis until an experiment, benchmark, or real deployment supports it.

Chapter Course Syllabus Template

Revisions

AlphaFold 3 access terms: noncommercial weights, commercial use through Google Cloud

AlphaFold 3's downloadable weights are for noncommercial use by noncommercial organizations; commercial users go through Google Cloud.

Why Companies and labs that need structure prediction on their own servers or for commercial work must check licences before choosing a tool.

Chapters Toolkit for AI-Augmented Bio Research · Protein Structure Prediction · AI for the Life Sciences · History of AI in the Life Sciences · Case Studies · Model and Dataset Index

De novo metalloproteases and cysteine proteases with measurable catalysis; EDEN designs large serine recombinases from attachment-site prompts 2 updates
New evidence

De novo metalloproteases and cysteine proteases with measurable catalysis

Designed metalloproteases cleaved cognate peptides in 36% of 135 designs, the best exceeding 100 million-fold rate enhancement; companion cysteine proteases succeeded in 13 of 69.

Why Reliable enzyme design has been the field's open problem; these are measured catalytic successes in a top journal, still limited to amide-bond hydrolysis.

Chapter Protein Design and Engineering

New evidence

EDEN designs large serine recombinases from attachment-site prompts

A DNA language model designed large serine recombinases active in 53.6% of 176 cell-free tests, 10 of 20 integrating a CAR cassette in primary T cells.

Why Generative design is reaching the genome-editing enzymes used in cell therapy, where delivery and specificity, not activity, decide clinical use.

Chapter Cell and Gene Therapy AI

Isomorphic Labs: pharma partnerships, preclinical status 1 update
Licensed research agents and Reuters' report of Anthropic's wet lab 2 updates
News & developments

Reuters reports Anthropic's Bay Area wet lab

Reuters reports Anthropic's Bay Area wet lab, without a protocol or result establishing a closed-loop self-driving laboratory.

Why Readers can distinguish a staffed laboratory with instrument access from a demonstrated self-driving laboratory.

Chapter Self-Driving Laboratories

Revisions

Evaluating licensed research agents, tools, and patient data

Evaluation of licensed research agents includes their tools and data, with provenance, usage rights, traceable analyses, and experimental validation.

Why Research teams need evidence of scientific performance before attributing better targets or faster development to a commercial license.

Chapter Agentic Science Workflows

Cao et al.: MolGenBench tests structure-based molecular generation and hit-to-lead optimization; Ingelfinger et al.: CytoVI, a generative model for antibody-panel single-cell data; Qu et al. preprint: AutoScreen agents for CRISPR screen design and hit prioritization; Li et al.: DePass integrates paired multi-omics in dissociated and spatial data; He et al.: gut microbiome disease models failed to transfer between districts; EMA workshop input on human oversight for draft EU GMP Annex 22 on AI; NIH approves a second stage of Bridge2AI; Owkin and Boehringer Ingelheim, and Novo Nordisk and Anthropic, announce AI agreements; Bessemer and Bain survey: executives report higher AI returns in preclinical discovery 9 updates
Policy & regulation

EMA workshop input on human oversight for draft EU GMP Annex 22 on AI

An EMA workshop report collects stakeholder proposals for evidencing human oversight of AI-supported GMP decisions under draft Annex 22.

Why The discussion concluded that human oversight cannot compensate for inadequate validation of an AI-supported manufacturing decision.

Chapter AI for Biomanufacturing

New evidence

Cao et al.: MolGenBench tests structure-based molecular generation and hit-to-lead optimization

Cao and colleagues' MolGenBench finds that 17 structure-based generative methods still struggle to produce active, target-aware molecules.

Why Better three-dimensional conformations did not reliably yield active molecules, so docking validity is not evidence of target engagement.

Chapter Small Molecule Generation and ADMET

New evidence

Ingelfinger et al.: CytoVI, a generative model for antibody-panel single-cell data

Ingelfinger and colleagues' CytoVI integrates antibody-panel single-cell data; a CLL-versus-control classifier reached AUC-ROC 0.74 ± 0.09 at one hospital.

Why Protein-panel cytometry is routine in clinical labs, and transcriptomic foundation models cover it only partly.

Chapter Single-Cell Foundation Models

New evidence

Qu et al. preprint: AutoScreen agents for CRISPR screen design and hit prioritization

A preprint reports that AutoScreen agents prioritized more validated CRISPR screen hits than the strongest agent baseline; results are author-reported.

Why Agentic prioritization can make screen follow-up auditable, but rankings still require experimental confirmation.

Chapters Target Identification and Prioritization · Cell and Gene Therapy AI

New evidence

Li et al.: DePass integrates paired multi-omics in dissociated and spatial data

Li and colleagues' DePass couples denoising with paired multi-omics integration; benchmark gains are author-reported.

Why Paired multimodal measurements are often noisier than single-modality profiles, and most methods handle either dissociated or spatial data, not both.

Chapters Spatial Omics and Tissue Models · Microbiome and Multi-Omics AI

New evidence

He et al.: gut microbiome disease models failed to transfer between districts

In He and colleagues' 14-district study, microbiome disease models built in one district failed in others.

Why Microbiome-based predictors need local validation before use in a new population.

Chapter Microbiome and Multi-Omics AI

News & developments

Bessemer and Bain survey: executives report higher AI returns in preclinical discovery

Executives surveyed by Bessemer and Bain reported higher AI returns in preclinical discovery than in clinical development; returns are self-reported.

Why Self-reported returns are not evidence that a method changed a program outcome, and bench-to-bedside timelines exceed annual measurement.

Chapter Translational Evidence and Failure Modes

Kosonocky et al. report functional results from Bits to Binders; Llompart et al. report MAGI outcomes tied to property-predictor accuracy; Google DeepMind and collaborators report AlphaProtein Novo enzyme designs; ARPA-H RAPID awards for rare-disease diagnosis and treatment 4 updates
Policy & regulation

ARPA-H RAPID awards for rare-disease diagnosis and treatment

ARPA-H RAPID awards fund AI-ready rare-disease data, diagnostic tools, and therapeutic-development support.

Why The awards support rare-disease research through data preparation, diagnostic tools, and therapeutic development.

Chapter AI for the Life Sciences

New evidence

Kosonocky et al. report functional results from Bits to Binders

Bits to Binders found that structure-confidence scores did not predict minibinder function in primary human T cells, within a single CAR scaffold and donor.

Why Functional measurements in therapeutic cells help researchers judge designs beyond structure-confidence scores.

Chapter Protein Design and Engineering

New evidence

Llompart et al. report MAGI outcomes tied to property-predictor accuracy

In a retrospective preprint, MAGI campaign outcomes tracked property-predictor accuracy rather than generation route; net acceleration over expert work remains preliminary.

Why Property-predictor accuracy helps explain whether generated chemistry meets a campaign objective.

Chapter Small Molecule Generation and ADMET

New evidence

Google DeepMind and collaborators report AlphaProtein Novo enzyme designs

AlphaProtein Novo preprints report enzyme designs for difficult chemistry, with catalytic activities still below natural or directed-evolution-optimized enzymes.

Why These results connect enzyme design to difficult chemical reactions while identifying activity and expertise limits that govern practical use.

Chapter Protein Design and Engineering

Yang et al. report MetagenomicsBench results for model and harness configurations 3 updates
New evidence

Yang et al. report MetagenomicsBench results for model and harness configurations

Benchmarks describes MetagenomicsBench, a preprint with 100 evaluations drawn from published metagenomic datasets and deterministic grading. Across 10,200 trajectories from 34 model-harness configurations, the strongest configuration passed 60.0%; greater cost, token usage, and tool use did not consistently track accuracy. Scores apply to configurations, not models alone.

Why Technically plausible analyses can still fail in problem interpretation, statistical reasoning, and biological interpretation.

Chapter Benchmarks for Bio AI

New evidence

Two preprints point to context diversity as a limit on virtual-cell models

Perturbation Prediction and Virtual Cells describes two 2026 preprints: Xaira Therapeutics reports little test-loss improvement beyond about 1.6 billion parameters in its X-Cell model-size sweep on a public subset. Biohub, University of Chicago, and Northwestern researchers report that recovery of differentially expressed genes rose with training contexts per perturbation and correlated weakly with cell count (r = 0.11).

Why New cell types, donors, and conditions may matter more for these models than more cells sequenced.

Chapter Perturbation Prediction and Virtual Cells

News & developments

Biohub announces a $1.8 billion Virtual Biology Initiative expansion

Biological Data Infrastructure describes Biohub's announced $1.8 billion coalition for predictive cell models, combining funding, data, computation, and measurement technology. Reuters and Axios report a one-year commercial embargo for company-funded data streams before public release; government-funded work is described as unrestricted. The program announcement does not establish a validated universal executable cell.

Why Funding composition and release conditions determine what biological data researchers can use and when.

Chapters Biological Data Infrastructure · Perturbation Prediction and Virtual Cells

Rentosertib Phase 2a safety evidence and September 2026 Phase 3 start 2 updates
News & developments

Xaira reports X-Design Vega antibody binding and inhibition results

Antibody and Biologic Design describes Xaira's reported 18 confirmed VHH binders from 182 XA-1 designs and a preclinical lead 7 weeks from DNA synthesis. For XA-4, Xaira reports 6 confirmed inhibitors; these internal campaign results await independent evaluation.

Why Campaign denominators and functional measurements help readers assess antibody design claims alongside the need for independent evaluation.

Chapter Antibody and Biologic Design

Revisions

Rentosertib Phase 2a safety evidence and September 2026 Phase 3 start

Small Molecule Generation and ADMET describes safety and tolerability as rentosertib's primary Phase 2a endpoint, with lung function (FVC) secondary. In September 2026, Insilico Medicine began dosing in GENESIS-IPF-3, a 52-week, placebo-controlled Phase 3 trial in China with annual FVC decline as its primary endpoint; Phase 2a did not establish clinical efficacy.

Why Trial endpoints determine whether evidence supports safety, an exploratory lung-function signal, or a clinical efficacy claim.

Chapters Small Molecule Generation and ADMET · Case Studies · Executive Summary

CytoVI for cytometry modeling and missing-marker imputation 1 update
New content

CytoVI for cytometry modeling and missing-marker imputation

Single-Cell Foundation Models describes CytoVI for generative modeling of flow cytometry, mass cytometry, and CITE-seq, including missing-marker imputation and annotation. Imputed markers remain estimates and are not a validated diagnostic.

Why Readers can distinguish estimated marker values from validated diagnostic measurements.

Chapter Single-Cell Foundation Models

Calibrated perturbation metrics; genomics AI policies; wet-lab autonomy terms 3 updates
New evidence

Calibrated metrics for deep genetic-perturbation models

Single-Cell Foundation Models states that low dynamic range in MSE and Pearson Δctrl can make deep perturbation models look weak against mean or PCA-linear baselines. Miller et al. (Nature Biotechnology, 2026) report that WMSE, weighted R²Δ, and normalized inverse rank let them beat those baselines on 14 datasets and 18 metrics. Ahlmann-Eltze et al. diagnose miscalibration.

Why Stops a common metric from being read as proof that deep perturbation models add nothing, and stops calibrated gains from being read as universal superiority.

Chapters Single-Cell Foundation Models · Perturbation Prediction and Virtual Cells · Evaluation Principles for Life Sciences AI

New evidence

AI-specific policies in national genomics initiatives

Biological Data Infrastructure reports that population genomics programs feeding AI training need explicit AI governance, not only generic data-use terms. Hazel et al. (npj Digital Public Health, 2026) found AI-specific public policies in only three of 90 national genomics initiatives surveyed.

Why Shows how rarely national genomics programs publish AI-specific rules for data that may train models.

Chapter Biological Data Infrastructure

News & developments

Cloud access treated as orthogonal to wet-lab autonomy

Self-Driving Laboratories treats cloud access as orthogonal to autonomy: a robot that carries out a preset protocol is automation, model-directed choice of the next step is autonomy, and remote facility access is cloud. Batalis's 30 Sep 2026 note is a short teaching companion to that distinction, an essay rather than a closed-loop methods paper.

Why Keeps remote facility access from being counted as evidence that a laboratory chooses its own next experiment.

Chapter Self-Driving Laboratories

References, links, and wording were also kept current.

September 2026

45 updates

19 days with updates

Anthropic opened a Life Sciences Verification Program for Mythos 5.1, and OpenAI introduced GPT-Rosalind trusted-access tools. Zhang et al. (Science) described a multi-agent Virtual Biotech that curated outcomes from 55,984 trials. SynthIDBio (Stutz et al., Nature) embeds a function-preserving watermark in ProteinMPNN-style sampling, with DeepMind's SynthID Bio note as the institutional companion. Anthropic reported agentic discovery of a novel reverse-transcriptase system.

SynthIDBio watermarks designed proteins; Nobel Turing Challenge sets an autonomy bar 2 updates
New evidence

SynthID Bio watermarking for protein-design provenance

Protein Design and Engineering describes SynthIDBio (Stutz et al., Nature, 2026) as a function-preserving watermark embedded in ProteinMPNN-style sampling: a proof of concept with redesign and resequencing limits, not a substitute for synthesis screening. DeepMind's Introducing SynthID Bio is the institutional companion: Swiss-cheese framing, DNA synthesis-provider triage, and an Evo 2 phage watermark watch item.

Why Readers can check a peer-reviewed provenance mark for designed proteins and the lab's account of how that mark would sit beside DNA synthesis screening.

Chapter Protein Design and Engineering

New evidence

Nobel Turing Challenge bar for autonomous scientific discovery

Self-Driving Laboratories states the Nobel Turing Challenge (Kitano, 2021) as an autonomous Nobel-level discovery, horizon around 2050, and distinguishes 2024 Nobels for building AI systems from a future prize for a discovery made by AI. A 2025 Nature News Feature records expert timelines from a decade to fifty years: challenge vision, not evidence a deployed system has met the bar.

Why Sets a public bar for autonomous discovery so readers do not treat prizes for building AI systems as prizes for discoveries made by AI.

Chapter Self-Driving Laboratories

SpaCEy links spatial tissue graphs to survival; phage-host ML predicts E. coli infection 2 updates
New evidence

SpaCEy links spatial tissue graphs to progression and survival

Spatial Omics and Tissue Models describes SpaCEy, an explainable graph model of tissue from molecular marker expression, without cell-type labels or anatomical regions as inputs. In a spatial proteomic lung-cancer cohort it surfaced progression-linked patterns, and across breast-cancer proteomic datasets it stratified survival within and across clinical subtypes. It is a research explainer, not a locked clinical prognostic device.

Why Gives readers a spatial-clinical model whose outputs stay hypotheses until a locked prognostic evaluation exists.

Chapters Spatial Omics and Tissue Models · Histopathology AI

New evidence

Phage-host machine learning predicts strain-level E. coli infection

Nucleic Acid and Genome Models describes phylogeny-agnostic machine learning that predicts strain-level phage-host infection from genomes. Experimental validation of 1,240 E. coli interactions reached AUROC 0.84, and model-guided cocktails of at most five phages covered up to 97.5% of held-out strains. The use is phage selection and cocktail design, not generative genome design, and cross-genus transfer remains limited.

Why Adds a measured infection-prediction result for phage selection without treating it as generative genome design.

Chapter Nucleic Acid and Genome Models

AGENT thermostabilizes solid-state mRNA-LNPs; generative NRPS T-domain redesign 2 updates
New evidence

AGENT thermostabilizes solid-state mRNA-LNP formulations

mRNA, RNA, and Vaccine Design describes AGENT: excipient screening plus Bayesian optimization for solid-state mRNA-LNP formulations. Six iterations in about one month yielded SM-102 and ALC-0315 compositions that kept full bioactivity after more than two months at 37 °C. Responses in rodents and non-human primates were non-inferior to fresh soluble vaccine. The evidence is particle storage, not sequence design.

Why Separates a storage-stability result for formulated particles from sequence design and from an approved-product claim.

Chapter mRNA, RNA, and Vaccine Design

New evidence

Generative redesign of NRPS thiolation domains

Protein Design and Engineering reports 76 de novo NRPS thiolation domains from ESM3, ProteinMPNN, and EvoDiff, and 578 variants tested in vivo. Designed domains supported product formation, kept hybrid junctions active, and raised yields by up to about three-fold versus the native domain. One design raised melting temperature by 12 °C. It is megasynthetase design, not autonomous invention of antibiotics.

Why Shows a wet-lab gate on generative carrier-domain design, short of a claim that models invent antibiotics on their own.

Chapters Protein Design and Engineering · Synthetic Biology Design Tools · Chemical Biology and Target Engagement

He et al. PASCode: phenotype-association scoring across single cells 1 update
New evidence

He et al. PASCode: phenotype-association scoring across single cells

Single-Cell Foundation Models adds PASCode, a graph-attention framework scoring how strongly individual cells associate with clinical phenotypes in single-nucleus RNA-seq; on the PsychAD prefrontal cohort (more than 6 million nuclei from 1,494 donors) it prioritized microglia and astrocytes for Alzheimer disease association. The chapter treats it as a tool for mechanistic and atlas work, not a diagnostic or treatment model.

Why Gives readers a released method, with code, pretrained models, and a web atlas, for ranking cells by phenotype association in population-scale single-cell data.

Chapter Single-Cell Foundation Models

Anthropic reports agentic discovery of a reverse-transcriptase-associated enzyme system 1 update
News & developments

Anthropic reports agentic genome-mining discovery of a novel enzyme system

Agentic Workflows adds Anthropic's report that Claude Mythos 5 agents surveyed about 200,000 reverse-transcriptase clusters across roughly 1.9 billion protein clusters in about 21.5 hours without human interruption, flagging a tandem-repeat array beside a jumbo-phage reverse transcriptase now named array-associated reverse transcriptases (ART); function remains unknown, and wet-lab work stays with human scientists under BSL-1 or BSL-2.

Why Reports this as company-published, not peer-reviewed, and not a basis for agents to handle biological materials without a human gate.

Chapter Agentic Science Workflows

Anthropic Life Sciences Verification Program and Mythos 5.1 access; CRISPR-GPT: LLM co-pilot for gene-editing experiment design; OpenAI GPT-Rosalind and Rosalind Workbench for life sciences; Saez-Rodriguez et al. Nature Methods perspective on foundation-model benchmarking; SAPP and DMX: standardized protein production and characterization; MutexaGPT: LLM agents for enzyme-engineering simulation workflows; Claude Opus 5.5 system card: life-sciences capability block 7 updates
New evidence

CRISPR-GPT: LLM co-pilot for gene-editing experiment design

Cell and Gene Therapy AI and Agentic Workflows add CRISPR-GPT, a 2025 Nature Biomedical Engineering LLM agent that plans and analyzes gene-editing experiments across knockout, base editing, prime editing, and CRISPRa/i; the authors report AI-guided wet-lab showcases including Cas12a knockout of four genes in a lung adenocarcinoma line and dCas9 activation of two genes in a melanoma line.

Why Documents a peer-reviewed gene-editing co-pilot as a human-AI tool with defined cell-line demos, not unsupervised clinical editing.

Chapters Cell and Gene Therapy AI · Agentic Science Workflows

New evidence

Saez-Rodriguez et al. Nature Methods perspective on foundation-model benchmarking

Evaluation Principles and Benchmarks add a Nature Methods Perspective arguing that biomedical foundation models need epistemological limitation-testing beyond leaderboards: asking whether a model can be refuted, verified, or judged by utility, avoiding a self-assessment trap where developers grade their own finals, and importing CASP/DREAM-style independent referees with withheld evaluation data.

Why Argues that leaderboard scores alone cannot settle whether a foundation model's outputs are trustworthy.

Chapters Evaluation Principles for Life Sciences AI · Benchmarks for Bio AI

New evidence

SAPP and DMX: standardized protein production and characterization

Protein Design and Engineering adds Semi-Automated Protein Production (SAPP), a modular expression, purification, and QC workflow that records yield, dispersity, and oligomeric state for hundreds of designs per day, plus DMX demultiplexing that cuts gene-synthesis cost roughly five-fold for thousand-design campaigns; reported showcases include redesigned fluorescent proteins and de novo binders that neutralize RSV in reported assays.

Why Adds a standardized QC pipeline for protein design libraries, short of function-specific potency or safety evidence.

Chapter Protein Design and Engineering

New evidence

MutexaGPT: LLM agents for enzyme-engineering simulation workflows

Protein Design and Engineering adds MutexaGPT, an open multi-agent platform in Nature Computational Science that translates plain-language enzyme-engineering intuition into physics-based simulation workflows and mutation libraries; a showcase cavity-engineering library reported about a 40% hit rate with roughly four-fold enrichment of successful designs.

Why Reports an LLM-agent design pipeline whose enrichment claim comes from one showcase campaign, not broad validation.

Chapter Protein Design and Engineering

News & developments

Anthropic Life Sciences Verification Program and Mythos 5.1 access

Foundations and Toolkit add Anthropic's September 2026 Life Sciences Verification Program beta, which grants verified organizations Standard Use or High-risk Use access to Mythos, Opus, and Sonnet with biology classifiers more permissive than generally available models, alongside Claude Mythos 5.1 trusted access. This is a vendor access and monitoring program, not an independent safety certification.

Why Distinguishes a vendor's own access-tier and monitoring program from independent validation of biological safety or performance.

Chapters AI for the Life Sciences · Toolkit for AI-Augmented Bio Research

News & developments

OpenAI GPT-Rosalind and Rosalind Workbench for life sciences

Foundations and Toolkit add OpenAI's GPT-Rosalind trusted-access life-sciences model series, launched April 2026, and Rosalind Workbench, a ChatGPT research preview shipped August 2026 with guided scientific tasks, molecular and sequence viewers, and an NGS analysis path; Explore and Research modes gate advanced workflows to verified organizations. The chapters read these as vendor announcements, not peer-reviewed biology outcomes.

Why Tracks a second frontier lab's trusted-access life-sciences program at parity with Claude Science, pending outcome evidence.

Chapters AI for the Life Sciences · Toolkit for AI-Augmented Bio Research

News & developments

Claude Opus 5.5 system card: life-sciences capability block

Benchmarks adds Opus 5.5 beside Mythos 5.1, given the same CB-1, not CB-2, classification with expanded bio safeguards, plus the system card's life-sciences capability block covering protein design, de novo binders, BioMysteryBench, and Protocols. The chapter treats this as a developer-reported classification and capability score, not an independent uplift trial.

Why Flags a frontier lab's own bio-capability classification as self-reported, not third-party validated uplift evidence.

Chapter Benchmarks for Bio AI

AI-disclosure trends in math preprints; MMBU biomedical VLM perception benchmark 2 updates
New evidence

Epoch AI math-arXiv tracker on AI research-assistance disclosure

Knowledge Graphs and Literature AI adds Epoch AI's math-arXiv tracker, which found research-assistance acknowledgments among established-author math preprints rose from about 1% in April 2026 to about 12% in August 2026, while any-purpose AI acknowledgments reached about 25% of all math preprints that month. The chapter frames this as voluntary disclosure in mathematics, not measured model use in biomedicine.

Why Shows disclosure norms shifting even outside biomedicine, without implying similar rates apply to biomedical research.

Chapter Knowledge Graphs and Literature AI

New evidence

MMBU preprint on biomedical vision-language model perception

Benchmarks adds MMBU, a preprint spanning 35 submodalities from 410 curated datasets, finding that medical fine-tuning gains on PathVQA, VQA-RAD, and SLAKE do not reliably survive a broader perception suite: closed-to-open gaps are large, object detection stays near or below random, and adapted and base models tie on most comparisons.

Why Signals a benchmark-culture problem in biomedical imaging VLMs, not a wet-lab or clinical finding.

Chapter Benchmarks for Bio AI

Virtual Biotech multi-agent system curates trial-outcome associations across 55,984 trials 1 update
New evidence

Zhang et al. Science paper on Virtual Biotech multi-agent system

Agentic Workflows and Clinical Trial AI for Translational Research add Virtual Biotech, a chief-scientific-officer agent coordinating domain scientist agents across genetics, omics, chemoinformatics, and clinical data; more than 37,000 clinical-trialist agents curated outcomes from 55,984 trials, finding drugs aimed at cell-type-specific, switch-like genes advanced with fewer adverse events. This is agent-curated association, not wet-lab proof of approval rates.

Why Reports agent-curated trial associations as observational, not evidence that agent companies raise approval rates.

Chapters Agentic Science Workflows · Clinical Trial AI for Translational Research

Cell perspectives on agentic science, virtual-cell world models, and aging-AI benchmarks 3 updates
New evidence

Guo et al. Cell perspective on agentic biomedical AI

Agentic Workflows and Self-Driving Labs add a Cell perspective mapping biomedical AI's shift from tools that answer human-posed questions to agents that pose their own, alongside risks of hallucination, bias, dual use, and cognitive deskilling; the chapters note its conclusion centers responsible human-AI collaboration, not unsupervised wet-lab agency.

Why Frames agentic-science risk in named terms while flagging that autonomy has not replaced scientific judgment.

Chapters Agentic Science Workflows · Self-Driving Laboratories

New evidence

Virtual cells as biomedical world models

Perturbation Prediction and Virtual Cells adds two Cell perspectives: Xing and Song frame the virtual cell as a multi-modal, multi-scale, stateful world model for simulation-first intervention design, and Noori, Zitnik, and colleagues define biomedical world models as action-conditioned simulators for counterfactual reasoning, distinct from static predictors. Evaluation Principles adds the same action-conditioned evaluation requirement.

Why Sets an evidence bar so world-model claims are judged by named interventions, not by definitional language alone.

Chapters Perturbation Prediction and Virtual Cells · Evaluation Principles for Life Sciences AI

New evidence

LongevityBench for aging-AI readiness

Benchmarks and Aging and Longevity Biology AI add LongevityBench, an open evaluation stack with Longevity-LLMs and Longevity Claw testing whether current AI systems are ready to lead aging research rather than only score aging clocks; the chapters treat strong results as readiness diagnostics, not lifespan, healthspan, or rejuvenation evidence.

Why Adds a benchmark for aging-AI readiness while guarding against overreading scores as clinical outcomes.

Chapters Benchmarks for Bio AI · Aging and Longevity Biology AI

Paper2Agent turns papers into MCP agents; FDA CTD archive access essay added 2 updates
New evidence

Paper2Agent: converting research papers into MCP agent servers

Agentic Workflows and the Model and Dataset Index add Paper2Agent, which converts a paper plus its codebase into a Model Context Protocol server a chat agent can query; case studies on AlphaGenome, Scanpy, and TISSUE reproduced tutorial results and answered novel queries. The chapter treats this as dissemination infrastructure, not evidence that agent teams raise wet-lab discovery rates.

Why Gives researchers a way to query a paper's methods directly, while flagging that agentifying incomplete code fails.

Chapters Agentic Science Workflows · Model and Dataset Index

News & developments

Institute for Progress essay on FDA CTD archive access

Clinical Trial AI for Translational Research adds a 2025 Institute for Progress essay on FOIA Exemption 4 limits on Common Technical Document dossier access after Food Marketing Institute v. Argus Leader Media (2019), and its proposed AI Regulatory Transparency Fund to buy orphaned CTDs from bankrupt sponsors for a searchable corpus. The drafting-speedup example is vendor self-report, not independent evaluation.

Why Surfaces a data-access barrier that keeps small biotechs from training regulatory AI tools on precedent filings.

Chapter Clinical Trial AI for Translational Research

Google DeepMind AI-in-science survey on productivity and verification bottlenecks 1 update
News & developments

Google DeepMind AI in Science survey on productivity and verification

Agentic Workflows adds a September 2026 Google and Google DeepMind snapshot surveying 637 US and UK scientists alongside about 15 million Gemini interactions; scientists reported average time savings near 6.9 hours per week, but about 46% of those who saved time spent over a quarter of it verifying AI outputs. The chapter treats this as self-reported evidence, not causal proof.

Why Flags the productivity data as self-reported and platform-specific, not a demonstrated rise in discovery rates.

Chapter Agentic Science Workflows

JAMA+ AI Conversations interview on clinical-trial enrichment with machine learning 1 update
New evidence

Perlis-Geraci JAMA+ AI Conversations on clinical-trial enrichment

Clinical Trial AI for Translational Research adds Perlis's JAMA+ AI Conversations interview with Joseph Geraci on machine learning for clinical-trial enrichment, covering subpopulation identification and what regulators may expect when trials apply AI biomarkers; the chapter treats it as a teaching pointer, not methods validation of any vendor algorithm.

Why Gives readers a plain-language framing of AI-assisted trial enrichment and how FDA reviewers may view it.

Chapter Clinical Trial AI for Translational Research

Self-Driving Labs cites the AI-scientist verification-budget preprint 1 update
New evidence

Self-driving-labs chapter cites the AI-scientist verification-budget preprint

Self-Driving Labs adds a pointer to the 2026 preprint arguing that closing the loop in biomedical discovery is constrained by the verification budget as much as by hypothesis quality, noting hard wet-lab verifiers are not interchangeable with soft simulators, with fuller teaching in Agentic Workflows.

Why Extends the same verification-budget caution to autonomous-lab claims, since self-driving-lab and AI-scientist claims share the same evidentiary gap.

Chapter Self-Driving Laboratories

DeepMind releases AlphaGenome Atlas variant-effect catalogue 5 updates
New evidence

GPN-Star phylogeny-aware variant-effect model added

Variant Effect Prediction adds GPN-Star, a phylogeny-aware genomic language model using whole-genome alignments and species trees for coding and non-coding variant-effect prediction, treating its scores, like AlphaGenome and GPN-MSA, as ACMG/AMP triage evidence rather than standalone classifications.

Why Adds a third phylogeny-aware scoring method to the chapter's triage-evidence framework for variant classification.

Chapter Variant Effect Prediction

New evidence

Preprint on the verification bottleneck for AI scientists

Agentic Workflows adds a 2026 preprint arguing that AI-scientist progress is constrained as much by the verification budget as by hypothesis quality, and that evaluation should test experiment selection, revision under new evidence, and calibrated uncertainty across simulation and wet-lab verifiers.

Why Gives the chapter a preprint perspective on evaluation design, explicitly not peer-reviewed and not evidence that any deployed agent closes the loop.

Chapter Agentic Science Workflows

New evidence

Preprint outlines DeepMind's AI Control Roadmap for internal agents

Agentic Workflows adds DeepMind's GDM AI Control Roadmap (version 0.1), which proposes system-level detection and response ladders for imperfectly aligned internal agents, labeled a vendor roadmap.

Why Gives readers the developer's own proposed safeguards for internal agent misalignment, framed as a vendor roadmap rather than an independently evaluated control method.

Chapter Agentic Science Workflows

News & developments

DeepMind releases AlphaGenome Atlas variant-effect catalogue

Variant Effect Prediction adds DeepMind's AlphaGenome Atlas, a precomputed catalogue of predicted molecular effects for about 9 billion single-nucleotide changes with an AlphaGenome Variant Impact score for ranking, described as a research resource built on AlphaGenome predictions, not a licensed clinical diagnostic.

Why Gives readers DeepMind's own framing of the atlas as a research resource, not a diagnostic tool, ahead of independent clinical validation.

Chapter Variant Effect Prediction

News & developments

OpenAI reports rising internal agent use in research

Agentic Workflows adds OpenAI's September 2026 research-acceleration snapshot, which claims an automated-research-intern level inside its lab and rising agent-workday use under human priority-setting, labeled a vendor self-report with a pointer to a companion handbook for fuller safety context.

Why Gives readers the developer's own claim about internal agent use, explicitly framed as a vendor self-report rather than independently verified capability.

Chapter Agentic Science Workflows

Nature Biotechnology proteomic aging-clock analysis added for rentosertib 1 update
New evidence

Proteomic aging clocks added to rentosertib evidence row

Small Molecule Generation and ADMET adds a Nature Biotechnology analysis nesting six proteomic aging clocks in longitudinal serum from a Phase 2a subset of the TNIK inhibitor rentosertib, framing clock shifts toward younger predicted biological age as exploratory biomarker signals, not proof of clinical rejuvenation or a geroprotection claim.

Why Keeps the rentosertib evidence row bounded to exploratory biomarker findings rather than a registration-level anti-aging claim.

Chapter Small Molecule Generation and ADMET

Australia's Clinical Trial Notification scheme added as a regulatory comparison 1 update
Policy & regulation

Australia's Clinical Trial Notification scheme added as a regulatory comparison

Clinical Trial AI adds Australia's Clinical Trial Notification scheme as a comparison: the human research ethics committee assesses the trial while the Therapeutic Goods Administration does not evaluate trial data at submission, though ethics approval and institutional authorization remain required, and notes this does not establish a comparative reduction in trial-start time or equivalent safety outcomes.

Why Completes the chapter's regulatory-process comparison with a primary source, while cautioning against reading a review-responsibility difference as a demonstrated speed or safety advantage.

Chapter Clinical Trial AI for Translational Research

Anthropic's Model Hardware Standard preview; evidence on invented citations in RAG tools 2 updates
New evidence

Evidence on invented citations added to literature RAG guidance

Knowledge Graphs and Literature AI adds citations showing early ChatGPT evaluations found high rates of invented bibliographic references that looked real, and reports that retrieval-augmented clinical evidence tools can reduce the rate of invented citations relative to general chatbots, while noting citation integrity is not the same as clinical correctness.

Why Grounds the chapter's resolve-before-cite guidance in measured rates of invented citations rather than a general warning about AI-generated citations.

Chapter Knowledge Graphs and Literature AI

News & developments

Anthropic previews a Model Hardware Standard for lab instruments

Anthropic's research preview of the Model Hardware Standard describes a model-agnostic driver that lets agents operate programmable instruments; partner notes report a liquid-handler and plate-reader workflow and an agent-supervised qPCR run. Self-Driving Labs treats it as an instrument-interface preview rather than a peer-reviewed self-driving lab, and a companion handbook covers instrument-access governance.

Why Gives readers the developer's own framing that the spec is not open source and that safety evaluations are still being built, distinct from a demonstrated autonomous lab.

Chapter Self-Driving Laboratories

NYT opinion: trial latency as an AI-era translation bottleneck 1 update
News & developments

NYT opinion: AI-era trial latency as a translation bottleneck

Clinical Trial AI adds a section on a New York Times opinion essay arguing that patient trial latency, not discovery science, is often the bottleneck for new cancer therapies, reporting an asymmetric incentive against fast action and citing HHS's Operation TrialBlazer and an FDA proposal for expedited first-in-human IND review as an incremental policy response.

Why Gives readers a researcher-sourced argument for why translation may be the binding constraint on AI-era cancer therapies, while noting interview-based opinion is not measured registry evidence and does not replace FDA and EMA validation discipline.

Chapter Clinical Trial AI for Translational Research

29 evidence bounds added across molecular, cellular, and therapeutic chapters 9 updates
New evidence

Protein design and engineering: four new evidence bounds

Protein Design and Engineering adds four bounds: EvoMax for wet-lab nuclease engineering, ProteinDPO aligning ESM-IF1 inverse folding to experimental stability data, DeepSCan mapping sequence to cell-surface translocation for display design, and a mammalian-display screen of thousands of AI-designed minibinders against PD-L1, CD276, and VTCN1.

Why Gives the chapter concrete wet-lab-validated bounds on what current design and scoring methods demonstrate, distinct from computational predictions alone.

Chapter Protein Design and Engineering

New evidence

ESM-3 dynamics readouts: Dyna-1 and ESMDynamic bounds

Molecular Discovery and Design adds two ESM-3 dynamics readouts: Dyna-1, trained on missing NMR assignments to mark microsecond-to-millisecond motion, and ESMDynamic, a dynamic-contact-map bound, both framed as readouts on top of ESM-3 rather than new base models.

Why Shows readers how new dynamics-prediction claims relate to the existing ESM-3 base model rather than replacing it.

Chapter Protein Structure Prediction

New evidence

RNA pseudoknot design and methylation-scoring bounds

Nucleic Acid and Genome Models adds a structure-creation bound on specified RNA pseudoknots, distinguished from viral-function claims, and Melody, which scores locus-specific DNA methylation from 10-kb genomic sequence across 39 human tissues.

Why Gives the chapter bounded examples of RNA structure design and methylation scoring rather than broader genome-regulatory claims.

Chapter Nucleic Acid and Genome Models

New evidence

Antibody-discovery and rare-variant-effect benchmark bounds

Antibody and Biologic Design adds AIntibody, a blinded prospective benchmark testing 511 designed or predicted antibodies from 29 organizations; Variant Effect Prediction adds RareEffect, an empirical-Bayes method estimating gene-level and variant-level rare-variant effect sizes.

Why Gives both chapters blinded or statistically grounded benchmarks rather than self-reported design or scoring claims.

Chapters Antibody and Biologic Design · Variant Effect Prediction

New evidence

Histopathology AI: four new benchmark and model bounds

Histopathology AI adds nnMIL, a slide-level aggregator over existing pathology foundation-model embeddings; ALPaCA, a slide-level vision-language framework trained on 35,913 whole-slide images; TRICARE, which scores risk across 2D levels inside 3D pathology volumes; and ConceptCLIP, a concept-aligned vision-language model pretrained on 23 million image-text-concept triplets.

Why Extends the chapter's evidence base across slide aggregation, visual question answering, 3D triage, and concept-aligned representation learning.

Chapter Histopathology AI

New evidence

Cell imaging and spatial-omics: six new evidence bounds

Adds VirTues (spatial-proteomics cohort bounds), MOSS (an LLM-assisted toolkit existence proof), a cryo-ET phantom with ground-truth annotations for six molecular species, scE2TM (an interpretable topic-model embedding), Bonsai (tree-based visualization with improved nearest-neighbor recovery), and a Cell Painting assay pairing PMA activation with 8,387 compounds.

Why Extends chapters covering spatial, imaging, and single-cell methods with named, source-grounded bounds on what each approach demonstrates.

Chapters Spatial Omics and Tissue Models · Microscopy and Cryo-EM AI · Single-Cell Foundation Models · Cell Painting and Image-Based Phenotyping · Toolkit for AI-Augmented Bio Research

New evidence

Therapeutics: four new evidence bounds across three chapters

Chemical Biology and Target Engagement adds VITAL, which scores peptide-protein interactions and affinity from language-model embeddings, and a MaSIF-mimicry surface-query screen of CRBN binding; Cell and Gene Therapy AI adds OptiPrime for prime-editing efficiency scoring; Drug Repurposing and Combination Therapy adds an iPSC cortical-culture phenotypic screen.

Why Gives each chapter a specific, source-grounded example of what current scoring and screening methods demonstrate in that therapeutic area.

Chapters Chemical Biology and Target Engagement · Cell and Gene Therapy AI · Drug Repurposing and Combination Therapy

New evidence

Lab automation and literature-extraction bounds

Self-Driving Labs adds GOLLuM, a calibrated Gaussian-process optimizer for next-experiment selection; Synthetic Biology Tools adds PUREdrop, a cell-free droplet screen of designed protein libraries; Knowledge Graphs and Literature AI adds PubMind, which extracts variant-function-disease associations from biomedical text.

Why Gives three chapters a calibrated optimizer, a cell-free screening method, and a literature-extraction tool as concrete evidence bounds.

Chapters Self-Driving Laboratories · Synthetic Biology Design Tools · Knowledge Graphs and Literature AI

New evidence

Science sandboxes added as a closed-loop reasoning evaluation

Benchmarks adds science sandboxes (MPRAbox, CodonBox) as a closed-loop evaluation of whether agents infer underlying rules or only raise an oracle score, noting the proposal is a preprint using damp and dry oracles.

Why Gives the chapter an evaluation design aimed at distinguishing genuine rule inference from benchmark gaming, ahead of peer review.

Chapter Benchmarks for Bio AI

References, links, and wording were also kept current.

August 2026

29 updates

10 days with updates

The handbook published a DOI and archival PDF edition and added a new Evaluation Workflows chapter with downloadable worksheets. Protein Design and Engineering added Anthropic's reported Claude-designed de novo protein binder campaign.

New Evaluation Workflows chapter with downloadable worksheets 4 updates
New content

New Evaluation Workflows chapter with downloadable worksheets

Adds a new Evaluation Workflows chapter with a universal decision record and five downloadable CSV worksheets covering configured-system evaluation, tool-versus-process comparison, single-cell model evaluation, self-driving-lab audits, and clinical-trial AI validation.

Why Turns the handbook's evidence principles into repeatable, version-controllable worksheets teams can use without new software.

Chapter Practical Evaluation Workflows and Worksheets

New content

Decision-record frameworks added to four evaluation chapters

Self-Driving Labs, Cells/Tissues/Systems, Evaluation Principles, and Clinical Trial AI each add a decision-record subsection: classifying automation versus a closed autonomous loop, a model-adoption path for single-cell foundation models, a decision-ready evaluation protocol, and a decision chain mapping trial functions to immediate and downstream endpoints.

Why Pairs each chapter with a structured record so a claim, such as autonomy or adoption, can be checked against explicit fields rather than a single score.

Chapters Self-Driving Laboratories · Single-Cell Foundation Models · Evaluation Principles for Life Sciences AI · Clinical Trial AI for Translational Research

New content

Clinical trial AI models added to Model and Dataset Index

The Model and Dataset Index adds a Clinical Trial AI and Evidence Generation table listing TrialGPT, HINT, and the Clinical Trial Outcome dataset with their main use and evidence tier.

Why Gives readers a quick-reference entry point to clinical-trial AI tools and datasets already covered in the chapter.

Chapter Model and Dataset Index

Anthropic reports Claude-designed de novo protein binders 5 updates
New evidence

Trial matching and enrollment evidence across five studies

The new Matching evidence from retrieval to enrollment section adds five studies spanning candidate-trial matching, randomized prescreening and clinician-notification trials, and multisite eligibility classification, each showing evidence at one step, such as retrieval or screening, does not establish performance at another, such as enrollment.

Why Gives the chapter primary evidence that trial-matching gains at one step, such as faster screening, do not automatically translate into gains at enrollment or outcomes.

Chapter Clinical Trial AI for Translational Research

New evidence

Trial Outcome Prediction and Clinical Trial Outcome benchmark datasets

The new Trial outcome prediction and benchmark validity section adds the Trial Outcome Prediction benchmark (17,538 trials, retrospective phase-specific prediction) and the Clinical Trial Outcome dataset (about 125,000 drug and biologic trial records), noting the original benchmark authors flagged look-ahead bias as a validity threat.

Why Gives readers the primary benchmark datasets behind trial-outcome-prediction claims and flags a known validity threat for future evaluations.

Chapter Clinical Trial AI for Translational Research

New evidence

Nature Medicine Perspective proposes AIDO multiscale foundation models

Virtual Organisms and Digital Biology cites a Nature Medicine Perspective proposing AIDO, a modular system of multiscale foundation models spanning molecules to individuals, describing it as a research agenda in the same class as the virtual-cell roadmap, not evidence of an executable digital organism.

Why Gives the chapter a primary source for the AIDO proposal while flagging the authors' financial interest in GenBio AI and that the article does not demonstrate a working digital organism.

Chapter Virtual Organisms and Digital Biology

News & developments

Anthropic reports Claude-designed de novo protein binders

Protein Design and Engineering adds Anthropic's reported binder campaign: Claude (Opus 4.8 and Mythos Preview), with a human-expert prompt and public design and folding models, produced de novo binders against 14 of 15 targets, independently produced and tested by Adaptyv Bio and Twist Bioscience, with reported hit rates of 22 to 35% against a field-typical 10 to 15%.

Why Gives readers a primary company report of a large de novo binder campaign, while noting it is company-reported, not peer-reviewed, and the binders are not drugs.

Chapter Protein Design and Engineering

Revisions

Clinical trial AI: context of use and evidentiary standards

Clinical trial AI evaluation sets a context of use before selecting metrics, citing FDA's reported submission experience with AI components in drug development from 2016 to 2023, EMA's adopted reflection paper on AI across the medicinal product lifecycle, and a Nature Medicine review situating trial applications within the broader drug-development pipeline.

Why Grounds the chapter's evidentiary standards in FDA and EMA frameworks and situates trial AI within the broader drug-development pipeline before it evaluates specific tools.

Chapter Clinical Trial AI for Translational Research

Field Guide sections added to History, Emerging Frontiers, and Protein Structure 1 update
New content

Field Guide sections added to History, Emerging Frontiers, and Protein Structure

History, Emerging Frontiers, and Molecular Discovery and Design each add a five-question Field Guide (what the field solves, the core idea, current state, what remains open, why it matters); the structure chapter now frames AlphaFold 2's single-chain accuracy against AlphaFold 3, Boltz, and Chai's extension to complexes.

Why Applies one evidence-tiered framework across chapters so readers can gauge a claim's evidence level before acting on it.

Chapters History of AI in the Life Sciences · Emerging Frontiers in AI for the Life Sciences · Protein Structure Prediction

NIST agent-evaluation draft; post-training data lineage and model-collapse evidence 3 updates
Policy & regulation

NIST draft standard treats agent settings as part of evaluation protocol

Agentic Workflows adds a Configured-System Evaluation section describing the evaluated object as the full configured agent system, not the model alone, following a NIST initial public draft that calls for explicit objectives, reproducible runs, uncertainty analysis, and qualified reporting.

Why Gives readers a draft government standard for evaluating agent systems rather than model names alone.

Chapter Agentic Science Workflows

New evidence

Agent evaluation benchmarks: repeated-trial and judge-bias findings

The same section cites a study showing final-state and repeated-trial checks in customer-service benchmarks exposed failures that single conversational scores missed, and a study showing human and LLM judges are both susceptible to multiple forms of judgment bias, noting neither establishes biological performance.

Why Supports the chapter's call for outcome, process, and safety checks beyond a single benchmark score, while flagging that these findings come from non-biological domains.

Chapter Agentic Science Workflows

New evidence

Post-training data lineage: Tulu 3, RLAIF, and model collapse evidence

Biological Data Infrastructure adds a Post-Training Dataset Card and cites an open recipe exposing post-training data types and decontamination steps, a study showing AI-feedback reinforcement learning matched human-feedback results in summarization and dialogue, and a study showing indiscriminate recursive training on model-generated data can lose distribution tails.

Why Grounds the chapter's guidance that post-training data types, generators, and evaluators must be tracked as separate provenance, not treated as interchangeable or as independent ground truth.

Chapter Biological Data Infrastructure

Evidence-to-claim framework and reading path added; Nature Methods on open AI 4 updates
New evidence

Nature Methods Perspective on open and sustainable AI recommendations

Reproducibility cites a Nature Methods Perspective mapping open and sustainable AI recommendations across the life-sciences ecosystem, covering data and model documentation, archival infrastructure, and environmental considerations, and notes it is an implementation resource, not proof that a checklist ensures reproducibility.

Why Gives the chapter a primary source for openness and sustainability guidance spanning data, code, weights, and documentation.

Chapters Reproducibility and Open Science · References

New content

Evidence-to-claim matching table and decision-first reading path

Evaluation Principles adds a table matching claim types, such as prediction versus experimental benefit, to the evidence design each requires; How to Read adds a decision-first reading path; the Preface states the book distinguishes predictions, benchmark results, validated findings, and decision changes.

Why Gives readers explicit frameworks for judging what a result does and does not establish, and for navigating the handbook by decision rather than model name.

Chapters Evaluation Principles for Life Sciences AI · How to Read This Handbook · Preface

New content

Timeline of AI in biology added to History chapter

Adds a milestone timeline from 1970 Needleman-Wunsch sequence alignment through BLAST, profile-HMM methods, DeepBind and DeepSEA, and DeepVariant, describing later systems as changing scale and reach while retaining earlier databases and evaluation practices.

Why Shows readers that current systems built cumulatively on established methods rather than replacing the prior scientific record.

Chapter History of AI in the Life Sciences

New content

Tool-adoption record and comparative trial guidance

Adds a tool-adoption record template covering workload and owner, decision and endpoint, data boundaries, candidate tool, evidence basis, integration path, cost, and acceptance and exit tests, plus a guide for running a comparative tool trial.

Why Gives research teams a structured, durable basis for evaluating and reviewing tools instead of relying on feature lists.

Chapter Toolkit for AI-Augmented Bio Research

Self-driving labs: bounded optimization and experimental limits 2 updates
Revisions

Self-driving labs: bounded optimization and experimental limits

Self-Driving Labs describes how evidence spans closed-loop chemistry and materials systems while current biology examples remain lab-in-the-loop rather than autonomous robotic loops, citing Coscientist, Virtual Lab, Robin, and A-Lab alongside published critiques of A-Lab's novelty claims.

Why Distinguishes bounded optimization from general autonomous discovery.

Chapter Self-Driving Laboratories

Revisions

Protein structure predictions and their experimental limits

Molecular Discovery and Design describes AlphaFold 2's routine role since CASP14, notes the AlphaFold Protein Structure Database now covers over 214 million predicted structures, and frames AlphaFold 3, Boltz-1, and Chai-1 alongside the questions a prediction still leaves for experiments to answer.

Why Distinguishes predicted structures from questions that require experimental evidence.

Chapter Protein Structure Prediction

USG policy for stopping high-risk life sciences research 1 update
New medical-countermeasure readiness section in vaccine design chapter 1 update
New content

New section: vaccine design to medical-countermeasure readiness

mRNA, RNA, and Vaccine Design adds a section arguing that faster antigen or sequence design alone does not establish end-to-end medical-countermeasure readiness, citing CEPI's 100 Days Mission, a BARDA-authored review of COVID-19 countermeasure development, a 2026 commentary mapping AI chokepoints across discovery, manufacturing, trials, and supply, CEPI's planned Pandemic Preparedness Engine, and joint FDA-EMA good AI practice principles.

Why Separates faster computational design from the regulatory, manufacturing, and supply readiness a real pandemic response requires.

Chapter mRNA, RNA, and Vaccine Design

DeepDETAILS: cell-type signal reconstruction from bulk assays; HiCFoundation: pretrained 3D genome contact map representations; Measured pretraining scaling in single-cell foundation models; Experimentally tested antibody and small-molecule binder design; ContactSeek: structure-guided base editor specificity; AI-nominated Parkinson disease targets tested in mouse models 6 updates
New evidence

DeepDETAILS: cell-type signal reconstruction from bulk assays

Microbiome and Multi-Omics AI and Variant Effect Prediction add DeepDETAILS, which reconstructs cell-type-specific regulatory signals from bulk assays such as PRO-cap, PRO-seq, and ChIP-seq to prioritize regulatory hypotheses in primary sclerosing cholangitis; the chapters note reconstructed signals are model-derived estimates, not direct single-cell measurements.

Why Gives readers a new route to regulatory-variant hypotheses that still needs external functional validation.

Chapters Microbiome and Multi-Omics AI · Variant Effect Prediction

New evidence

HiCFoundation: pretrained 3D genome contact map representations

Nucleic Acid and Genome Models adds HiCFoundation, evaluated across reproducibility assessment, resolution enhancement, loop detection, cross-species transfer, single-cell Hi-C analysis, and prediction of epigenomic activity from chromatin structure.

Why Adds a pretrained representation for 3D genome structure whose regulatory predictions remain hypotheses pending Hi-C-context validation.

Chapters Nucleic Acid and Genome Models · Microbiome and Multi-Omics AI

New evidence

Measured pretraining scaling in single-cell foundation models

Cells, Tissues, and Systems adds a section reporting that pretraining 400 models on up to 22.2 million cells across 6,400 model-task combinations found no clear data scaling law, and describes SIGnature, which linked the MS1 inflammatory program to Kawasaki disease and was followed by an in vitro test showing patient serum induced the phenotype.

Why Tests whether scale alone improves single-cell models and shows one attribution-based hypothesis confirmed experimentally.

Chapter Single-Cell Foundation Models

New evidence

Experimentally tested antibody and small-molecule binder design

Protein Design and Engineering adds Germinal, which tested 43 to 101 antibody designs per antigen across four targets and reported functional binders including nanomolar-affinity designs, and Neural iterative selection-expansion, which reported binding for all four tested exatecan designs and five of six tested apixaban designs.

Why Reports binder design results with tested-design denominators, though neither establishes therapeutic efficacy or manufacturability.

Chapter Protein Design and Engineering

New evidence

ContactSeek: structure-guided base editor specificity

Cell and Gene Therapy AI adds ContactSeek, which combined AlphaFold 3 contact predictions with measured off-target data to prioritize residues for base-editor specificity, then tested selected changes in Cas9-TadA adenine base editors and a Cas12a-based cytosine editor.

Why Adds a structure-model-guided workflow that still needs empirical off-target and genome-integrity testing before use.

Chapter Cell and Gene Therapy AI

New evidence

AI-nominated Parkinson disease targets tested in mouse models

Therapeutic Discovery and Translation adds a preclinical study in which XunZi nominated CHK2 and IRAK4 as candidate targets in Parkinson disease models, and pharmacological or genetic Chk2 inhibition reduced dopaminergic neuron loss and motor deficits in mice.

Why Shows AI-assisted target hypothesis generation followed by experimental validation, short of human disease modification evidence.

Chapter Target Identification and Prioritization

NIST CAISI agent red-teaming results; International AI Safety Report on benchmark and real-world evidence 2 updates
Policy & regulation

NIST CAISI agent red-teaming results

Information Hazards adds a summary noting that NIST's CAISI large-scale agent red-teaming competition found tool-use and agent-hijacking risk under adversarial testing, and states the results concern agent security rather than biological misuse capability.

Why Distinguishes agent-hijacking findings from biological misuse capability so readers do not conflate the two kinds of risk.

Chapter Information Hazards in Capability Research

Policy & regulation

International AI Safety Report on benchmark and real-world evidence

Information Hazards now cites the 2026 International AI Safety Report to support its point that evaluation practice separates benchmark evidence from real-world biological outcomes.

Why Adds an intergovernmental reference for the claim that benchmark performance does not by itself demonstrate real-world biological risk.

Chapter Information Hazards in Capability Research

References, links, and wording were also kept current.

July 2026

15 updates

5 days with updates

Foundations, Workforce and Compute, and Toolkit added Anthropic's beta launch of Claude Science, an AI workbench for scientific research. Information Hazards updated DURC and PEPP oversight coverage following Executive Order 14292, and Small Molecule Generation and ADMET added antimicrobial discovery evidence, including SyntheMol-RL, ApexGO, and CAMPER compounds validated in mouse infection models. Benchmarks organized biosecurity evaluation into five layers, adding GPT-5.6 and Claude Opus 5 classifications.

GPT-5.6 and Claude Opus 5 biosecurity classifications; five-layer evaluation framework 7 updates
Policy & regulation

NIST AI 800-2 and 800-3 on benchmark evaluation validity

Benchmarks cites NIST AI 800-2, an initial public draft on the validity, transparency, and reproducibility of automated benchmark evaluations, and NIST AI 800-3, which provides statistical methods for distinguishing fixed-benchmark performance from generalized task performance.

Why Gives readers government standards for the statistical rigor benchmark claims should meet, separate from biosecurity-specific evaluation.

Chapter Benchmarks for Bio AI

New evidence

Uplift study evidence: substantial novice gains on bounded tasks, no full-workflow effect

Benchmarks adds a 2026 preprint reporting substantial novice uplift on bounded digital biology tasks, and a separate preregistered randomized trial using mid-2025 models that found no significant difference in full-workflow completion for novices, though secondary analyses remained compatible with a possible modest benefit.

Why Shows that uplift findings vary by task scope, bounded tasks versus full research workflows, and should not be read as a single capability ceiling.

Chapter Benchmarks for Bio AI

New evidence

RFdiffusion3: all-atom protein design model, developer preprint

Protein Design and Engineering and the Model and Dataset Index add RFdiffusion3, an all-atom diffusion model generating protein structures with ligands and nucleic acids; the developer preprint reports about tenfold faster generation than RFdiffusion2 with experimental validation limited to two selected design tasks, and releases inference code, training code, and model weights.

Why Flags a new protein-design model as reported, developer-preprint evidence, since general performance across target classes is not yet established.

Chapters Protein Design and Engineering · Model and Dataset Index

New evidence

Mutation-split choice inflates protein-language-model generalization claims

Evaluation Principles adds a 2026 preprint analysis of 41 viral and 33 cellular deep-mutational-scanning datasets finding that pooled mutation splits let supervised protein-language-model predictors exploit site effects, inflating apparent generalization, while site-aware splits reduced performance and a site-mean baseline matched or outperformed supervised models on many datasets.

Why Warns readers that mutational-effect benchmark results depend heavily on split design, though the finding is specific to single-substitution datasets.

Chapters Evaluation Principles for Life Sciences AI · Foundation Models for Biology

New evidence

Retrosynthesis benchmark tools: Syntheseus, ChemCensor, and URSA-expert-2026

Small Molecule Generation and ADMET adds Syntheseus, a controlled reevaluation that changed retrosynthesis model rankings under consistent settings, and a developer preprint proposing ChemCensor, a precedent-based plausibility score, and URSA-expert-2026, an expert-annotated out-of-distribution set of 100 novel targets; the chapter notes these test model comparison and plausibility, not route execution or experimental feasibility.

Why Gives readers tools for retrosynthesis benchmark validity beyond exact-match accuracy, while flagging that neither scoring nor theoretical routes establish real synthesis feasibility.

Chapter Small Molecule Generation and ADMET

New evidence

TargetBench 1.0: disease-specific evaluation of ranked drug targets

Therapeutic Discovery and Translation adds TargetBench 1.0, an Insilico Medicine-led benchmark evaluating ranked target lists against historical clinical-stage targets and development-related evidence, noting that retrospective labels and text features can encode historical bias or temporal leakage, and that its code is proprietary.

Why Gives readers a target-ranking benchmark while flagging risks specific to retrospective, text-based evaluation designs.

Chapter Target Identification and Prioritization

News & developments

Five-layer biosecurity evaluation framework; GPT-5.6 and Claude Opus 5 classifications

Benchmarks, Executive Summary, and Chapter Summaries organize biosecurity evaluation into five layers, capability, uplift, safeguards, operational consequence, and lifecycle governance, and note that OpenAI treats all three GPT-5.6 models as High biological and chemical capability while Anthropic treats Claude Opus 5 as CB-1 but not CB-2 under ASL-3 protections.

Why Distinguishes what a public benchmark can and cannot establish, and keeps developer capability classifications separate from independent uplift claims.

Chapters Benchmarks for Bio AI · Executive Summary · Quick Reference: All Chapter Summaries

PBPK and QSP added as mature regulatory digital twins; benchmark uncertainty guidance 3 updates
New evidence

Statistical uncertainty in benchmark scores added to Evaluation Principles

Evaluation Principles adds a section explaining that benchmark scores carry sampling uncertainty like any other estimate, recommending standard errors clustered by correlated test-case groups, paired comparison when models share a benchmark, and a power-analysis formula for the number of test cases needed to detect a claimed improvement.

Why Completes the chapter's discussion of statistical significance testing for evaluation scores with a concrete method.

Chapter Evaluation Principles for Life Sciences AI

New content

PBPK and QSP models: the field's most mature regulator-recognized digital twin

Virtual Organisms and Digital Biology adds a section on PBPK and QSP models, reporting that 65 of 245 FDA new drug and biologics applications from 2020 to 2024 (26.5%) submitted a PBPK model as pivotal evidence, and covering the ASME V&V 40 credibility framework, FDA's 2023 guidance, MIDD, and the finalized ICH M15 guideline.

Why Closes a coverage gap: PBPK, QSP, MIDD, and ASME V&V 40 had no prior mention despite PBPK's routine use as pivotal regulatory evidence.

Chapter Virtual Organisms and Digital Biology

New content

In silico clinical trial credibility: a test for AI model-matches-trial claims

Virtual Organisms and Digital Biology adds an FAQ subsection on evaluating claims that a computational model matches clinical trial results, describing a hierarchical credibility-assessment workflow for in silico trials and citing its use in device submissions, pediatric trials, and randomized-trial augmentation with digital twins.

Why Gives readers a generalizable test for vendor claims that a simulation reproduces trial outcomes, rather than a single accuracy figure.

Chapter Virtual Organisms and Digital Biology

AI-designed antibiotics validated in mouse infection models 1 update
New evidence

Antimicrobial discovery with experimental validation: SyntheMol-RL, ApexGO, CAMPER

Small Molecule Generation and ADMET adds a section on antimicrobial discovery with experimental validation: SyntheMol-RL generated MRSA-active compounds tested in a mouse wound-infection model, ApexGO generated peptides evaluated in two Acinetobacter baumannii mouse infection models, and CAMPER prioritized membrane-targeting peptides tested against MRSA persisters in vitro and in mice.

Why Adds experimentally validated evidence that generative and optimization methods can produce candidate antibiotics active against drug-resistant bacteria, not just in-silico benchmark performance.

Chapter Small Molecule Generation and ADMET

Executive Order 14292 and NIH's rescinded DURC/PEPP notice; AI-discovered drug pipeline currency: rentosertib Phase 2a, Recursion, Insitro; Virtual Cell Challenge results: no consistent outperformance of naive baselines 3 updates
Policy & regulation

Executive Order 14292 and NIH's rescinded DURC/PEPP notice

Executive Order 14292 directs agencies to revise or replace the 2024 U.S. government dual-use research policy, and NIH has rescinded its implementation notice NOT-OD-25-061. Information Hazards advises following current institutional and funder guidance for DURC and PEPP oversight.

Why Keeps the handbook's dual-use oversight coverage aligned with the current regulatory status rather than the superseded 2024 policy notice.

Chapters Information Hazards in Capability Research · Quick Reference: All Chapter Summaries

Revisions

AI-discovered drug pipeline currency: rentosertib Phase 2a, Recursion, Insitro

Case Studies, Executive Summary, and Small Molecule Generation and ADMET update Insilico Medicine's ISM001-055 (rentosertib) with a peer-reviewed Nature Medicine report on its randomized Phase 2a trial, alongside the earlier Nature Biotechnology discovery paper, and note Recursion's and Insitro's pipeline candidates; no AI-discovered small molecule has yet reached FDA or EMA approval.

Why Reflects the trial's progression from discovery and Phase I to a randomized Phase 2a readout, while keeping the caveat that this is not registration-grade evidence.

Chapters Case Studies · Executive Summary · Small Molecule Generation and ADMET

Revisions

Virtual Cell Challenge results: no consistent outperformance of naive baselines

Perturbation Prediction and Virtual Cells, Emerging Frontiers, and Benchmarks update to reflect the Virtual Cell Challenge's public results: models showed improvements on some perturbation-prediction metrics but did not consistently outperform naive baselines across all metrics, reframing the challenge as a shared benchmark rather than a paper-specific split.

Why Reports the challenge's results and their limits across perturbation-prediction metrics.

Chapters Perturbation Prediction and Virtual Cells · Emerging Frontiers in AI for the Life Sciences · Benchmarks for Bio AI

Anthropic launches Claude Science, a vendor-reported beta AI workbench 1 update
News & developments

Claude Science, Anthropic's beta AI workbench for scientific research

Foundations, Workforce and Compute, and Toolkit add Anthropic's June 2026 beta launch of Claude Science, an AI workbench linking literature search, data analysis, code execution, remote compute, and scientific software; the chapters label the evidence vendor-reported and give practical criteria (data boundary, exportable code, rerunnability) for evaluating it.

Why Anthropic's own product announcement is the only current source for Claude Science; the chapters treat industrial workbench adoption as a signal to watch, not evidence of scientific productivity.

Chapters AI for the Life Sciences · Workforce, Compute, and Institutional Readiness · Toolkit for AI-Augmented Bio Research

References, links, and wording were also kept current.

June 2026

5 updates

2 days with updates

Foundation Models for Biology added OpenAI's Preparedness Framework, Anthropic's Responsible Scaling Policy, and Google DeepMind's Frontier Safety Framework, and Benchmarks added a dangerous-capability evaluation section covering WMDP, LAB-Bench, and uplift studies. Agentic Workflows and Self-Driving Labs cited Google Co-Scientist and FutureHouse Robin, and Clinical Trial AI for Translational Research added an FDA request for information on AI-enabled early-phase trials plus ProJenX and Unlearn's digital twin comparator in an ALS trial.

FDA clinical-trials AI pilot RFI; ProJenX digital twin comparator in ALS trial 2 updates
Policy & regulation

FDA request for information on AI-enabled early-phase trials pilot

Clinical Trial AI for Translational Research cites FDA's April 2026 Federal Register request for information on an AI-enabled early-phase clinical-trials pilot program, which explores whether AI can improve trial efficiency, safety monitoring, dose-selection, and early go/no-go decisions while aligning with the NIST AI Risk Management Framework.

Why Gives readers the primary regulatory source for FDA's exploration of AI-assisted early-phase trial design.

Chapter Clinical Trial AI for Translational Research

News & developments

ProJenX and Unlearn digital twin comparator for ALS trial PRO-101

The chapter adds ProJenX's PRO-101 Phase 1 ALS trial (ClinicalTrials.gov NCT05279755) as a boundary case, noting ProJenX and Unlearn's announced ALS Digital Twin Generator would estimate ALSFRS-R, slow vital capacity, and plasma neurofilament light trajectories as a model-based comparator, not a substitute for randomized efficacy evidence.

Why Illustrates that the credibility of an AI-generated external comparator rests on prespecified context of use, endpoints, and uncertainty analysis rather than the digital twin label.

Chapter Clinical Trial AI for Translational Research

Dangerous capability evaluation guidance: frontier safety frameworks and WMDP added 3 updates
New evidence

Dangerous-capability evaluation: WMDP, LAB-Bench, and biological uplift studies

Benchmarks adds a section on safety and dangerous-capability evaluation, covering the WMDP hazardous-knowledge proxy benchmark, the LAB-Bench evaluation, and uplift studies; it cites RAND's red-team study finding no statistically significant difference in biological attack plan viability with or without LLM assistance, and OpenAI's early-warning study on LLM-aided biological threat creation.

Why Distinguishes dangerous-capability evaluation, whether a model lowers the barrier to biological misuse, from benchmarks that measure beneficial scientific capability.

Chapter Benchmarks for Bio AI

New evidence

Google Co-Scientist and FutureHouse Robin added to agentic science evidence

Agentic Workflows and Self-Driving Labs cite Google Co-Scientist for biomedical hypothesis generation and FutureHouse Robin for lab-in-the-loop candidate discovery in dry age-related macular degeneration alongside ChemCrow, Coscientist, Virtual Lab, and CellVoyager as demonstrated systems, and reference a Nature editorial cautioning that autonomous research still requires human authorization for consequential steps.

Why Grounds claims about autonomous research systems in the published record rather than vendor description, and marks the line between demonstrated and ungoverned autonomy.

Chapters Agentic Science Workflows · Self-Driving Laboratories

News & developments

OpenAI, Anthropic, and Google DeepMind biological capability safety frameworks

Foundation Models for Biology notes that OpenAI's Preparedness Framework, Anthropic's Responsible Scaling Policy (ASL-3), and Google DeepMind's Frontier Safety Framework tie biological and CBRN capability thresholds to deployment mitigations, access controls, and model-weight security, adding that a gated model is not necessarily scientifically weaker.

Why Gives readers primary sources for how frontier developers connect biological capability evaluation to release decisions.

Chapter Foundation Models for Biology

May 2026

17 updates

4 days with updates

The Life Sciences AI Handbook published its first edition, with chapters including Knowledge Graphs and Literature AI, Histopathology AI, Systems Biology and Multiscale Modeling, and Aging and Longevity Biology AI. A blinded NIST inter-tool analysis found baseline nucleic acid synthesis screening performance above 95%.

NIST biosecurity screening evaluation and pathology model findings; Glossary expanded with about 40 new terms 2 updates
New evidence

NIST biosecurity screening evaluation and pathology model findings

Adds a blinded NIST inter-tool analysis finding baseline nucleic acid synthesis screening performance above 95% (Ikonomova et al., 2025; Laird et al., 2025) to the Synthetic Biology Tools and Information Hazards chapters, and a pathology foundation model fine-tuned for EGFR mutation prediction (Campanella et al., 2025) to the Histopathology AI chapter.

Why Gives readers primary evidence for the safety-screening and pathology-translation claims these chapters make.

Chapters Synthetic Biology Design Tools · Information Hazards in Capability Research · Histopathology AI

New content

Glossary expanded with about 40 new terms

Adds roughly 40 new glossary entries, including A-Lab, AAV, analytical validity, antisense oligonucleotide, and base editing, each defining the term and noting where it applies in the handbook.

Why Gives readers a fuller reference for terminology used across the handbook's molecular, cell, and translational chapters.

Chapter Glossary

New chapters on organismal, environmental, and translational AI; Organismal and therapeutic methods, from genomic selection to BEELINE; Cell-imaging and protein-design tools added to the glossary; Glossary adds counterfactual and trial-emulation terms 4 updates
Milestone

New chapters on organismal, environmental, and translational AI

Adds nine chapters: Systems Biology and Multiscale Modeling, Aging and Longevity Biology AI, Plant, Crop, and Agricultural AI, Environmental and Ecological AI, a new neuroscience chapter, Virtual Organisms and Digital Biology, Cell and Gene Therapy AI, Chemical Biology and Target Engagement, and Diagnostics and Biomarker Translation.

Why Extends the handbook's scope from molecular and cellular AI into organismal, environmental, and translational domains that previously had no dedicated coverage.

Chapters Systems Biology and Multiscale Modeling · Aging and Longevity Biology AI · Plant, Crop, and Agricultural AI · Environmental and Ecological AI · Neuroscience AI and Brain Foundation Models · Virtual Organisms and Digital Biology

New evidence

Organismal and therapeutic methods, from genomic selection to BEELINE

Eight chapters on agriculture, environmental ecology, systems biology, aging, neuroscience, virtual organisms, cell and gene therapy, and chemical biology cite primary studies including genomic selection (Meuwissen et al., 2001), plant breeding prediction (Crossa et al., 2017), SCENIC (Aibar et al., 2017), and the BEELINE benchmark (Pratapa et al., 2020).

Why Lets readers trace AI methods in agriculture, ecology, systems biology, aging, neuroscience, and therapeutics to the primary studies behind them, from genomic selection to the BEELINE benchmark.

Chapters Plant, Crop, and Agricultural AI · Environmental and Ecological AI · Systems Biology and Multiscale Modeling · Aging and Longevity Biology AI · Neuroscience AI and Brain Foundation Models · Virtual Organisms and Digital Biology

New evidence

Cell-imaging and protein-design tools added to the glossary

Adds glossary entries for Cellpose3, a restoration-aware cell-segmentation release (Stringer et al., 2025); COMPSS, a computational enzyme-design workflow (Johnson et al., 2024); and EVOLVEpro, a few-shot active-learning method for protein evolution (Jiang et al., 2025).

Why Gives readers primary sources for tools referenced across the cell-imaging and protein-design chapters this update expanded.

Chapters Glossary · Model and Dataset Index · Microscopy and Cryo-EM AI · Protein Design and Engineering

New evidence

Glossary adds counterfactual and trial-emulation terms

Adds glossary definitions for counterfactual response, neural optimal transport, Perturb-CITE-seq, and target trial emulation, the concepts behind virtual-cell and perturbation-prediction claims.

Why Gives readers primary citations for the causal-inference concepts used across the perturbation and virtual-cell chapters.

Chapter Glossary

Glossary, Model Index, Case Studies, and References; Six new chapters added, spanning imaging, therapeutics, and biomanufacturing 2 updates
New content

Glossary, Model Index, Case Studies, and References

The Glossary, Model and Dataset Index, Case Studies, Chapter Summaries, and References pages carry term definitions, a canonical table of over 50 models and datasets, twelve applied case studies, and a summary of every chapter.

Why Gives readers topic-organized reference tools for terminology, models, evidence-based cases, and citations across the handbook.

Chapters References · Glossary · Model and Dataset Index · Case Studies · Quick Reference: All Chapter Summaries

New content

Six new chapters added, spanning imaging, therapeutics, and biomanufacturing

Adds six chapters: Knowledge Graphs and Literature AI, Histopathology AI, Microscopy and Cryo-EM AI, Drug Repurposing and Combination Therapy, Real-World Evidence and Biomarker AI, and Biomanufacturing, each with matching glossary terms, model index entries, and chapter summaries.

Why Extends the handbook's coverage into the literature and knowledge-graph evidence layer, imaging modalities, drug repurposing, real-world data, and biomanufacturing.

Chapters Knowledge Graphs and Literature AI · Histopathology AI · Microscopy and Cryo-EM AI · Drug Repurposing and Combination Therapy · Real-World Evidence and Biomarker AI · AI for Biomanufacturing

First edition of The Life Sciences AI Handbook published; Foundations and Protein Structure Prediction: full-depth expansion; History of AI in the Life Sciences; Evaluation Principles and Foundation Models for Biology deepened; Protein Design and Variant Effect Prediction deepened; Single-Cell Foundation Models deepened; Small Molecule Generation and ADMET deepened; Self-Driving Laboratories deepened; Toolkit and Emerging Frontiers chapters 9 updates
Milestone

First edition of The Life Sciences AI Handbook published

Publishes the initial handbook with chapters organized into foundations, molecular AI, cellular systems, therapeutics AI, engineering and automation, and practice and governance, plus appendices and front matter.

Why Opens a dedicated evidence-based reference on AI for biomedical discovery, biotechnology, and translational research, alongside the companion handbooks.

Chapters Welcome · Biological Data Infrastructure · Antibody and Biologic Design · Microbiome and Multi-Omics AI · mRNA, RNA, and Vaccine Design · Agentic Science Workflows

New content

Toolkit and Emerging Frontiers chapters

Adds a Practical Workflows part with a Toolkit chapter on selecting tools for structure prediction and sequence search, and a Future Directions part with an Emerging Frontiers chapter surveying developing capabilities such as virtual cells.

Why Completes the handbook's eight-part structure with practical tool guidance and a forward-looking survey of developing capabilities.

Chapters Toolkit for AI-Augmented Bio Research · Emerging Frontiers in AI for the Life Sciences

Revisions

Evaluation Principles and Foundation Models for Biology deepened

Expands Evaluation Principles with sections on blind benchmarks and biology-aware data splits, and expands Foundation Models for Biology with sections tracing the protein language model lineage from ESM-1b through ESM3.

Why Establishes the evaluation standards and the shared foundation-model pattern the rest of the handbook builds on.

Chapters Evaluation Principles for Life Sciences AI · Foundation Models for Biology

Revisions

Protein Design and Variant Effect Prediction deepened

Expands Protein Design with sections on backbone generation (RFdiffusion) and sequence design (ProteinMPNN, LigandMPNN), and expands Variant Effect Prediction with missense, splicing, and non-coding variant sections.

Why Documents the generative design and variant-interpretation methods that complement structure prediction in the molecular AI chapters.

Chapters Protein Design and Engineering · Variant Effect Prediction

Revisions

Single-Cell Foundation Models deepened

Single-Cell Foundation Models covers scGPT, Geneformer, scFoundation, scBERT, UCE, and tGPT, with sections on representation learning at scale, cell-atlas search, and integration and batch correction.

Why Surveys the wave of single-cell foundation models trained between 2022 and 2024 so readers can compare their scale and validation.

Chapter Single-Cell Foundation Models

Revisions

Self-Driving Laboratories deepened

Expands the chapter with sections on chemistry automation, biology automation, and the discipline needed to evaluate novelty claims from systems such as A-Lab, alongside Coscientist and the Virtual Lab.

Why Distinguishes the more mature chemistry-automation track from the earlier-stage biology track in closed-loop experimentation.

Chapter Self-Driving Laboratories

References, links, and wording were also kept current.