6 ways AI reshaped scientific software in 2025

While the dawn of ChatGPT in 2022 helped kickstart the current AI wave, an AI moment that arguably mattered more for science was AlphaFold. December 2018, first place at CASP13, the Critical Assessment of Structure Prediction competition. By 2020, protein structure prediction was close enough to “solved” that the real question shifted from whether a fold was possible to what you could build on top of it. By 2024, it was Nobel Prize territory.

AlphaFold spawned an ecosystem: open-source competitors, specialized forks, commercial platforms built on its architecture. But it also helped legitimize a broader shift. Astronomers started using neural networks to classify galaxies faster than graduate students ever could. Climate modelers trained AI on decades of satellite data to sharpen precipitation forecasts. Epidemiologists built models that could track disease variants in near real-time. Drug developers stopped asking whether AI belonged in the pipeline and started asking which parts of the pipeline it could own.

Now in 2025, across disciplines, software is increasingly shaping which experiments get run. In structure prediction alone, the year delivered a wave of tools that built directly on AlphaFold’s foundation.

1. Protein structure prediction moves from folding to binding

The 2025 generation of structure prediction tools solved a problem AlphaFold left open: predicting how tightly a drug candidate will bind to its target. Boltz-2, released in June by MIT CSAIL and Recursion, predicts structure and binding affinity jointly, running 1,000 times faster than physics-based free energy perturbation methods. Genesis Molecular AI’s Pearl, backed by NVIDIA, claimed 40% improvement over AlphaFold 3 on drug discovery benchmarks and provided the first evidence of scaling laws in molecular AI. Meanwhile, the OpenFold Consortium released OpenFold3 in October.

The David Baker Lab at University of Washington pushed the design frontier. RFdiffusion2 (April) designs enzymes from functional group geometries. RFdiffusion3 (December) runs 10 times faster with atom-level precision. MIT’s BoltzGen, released in November, generates protein binders from scratch, achieving nanomolar binding for 66% of novel targets tested.

Major releases in structure prediction

Tool	Date	Developer	What it does	Why it matters	Sources
Boltz-2	June 2025	MIT CSAIL / Recursion	Predicts structure + binding affinity jointly	1000x faster than physics-based FEP; open-source under MIT license; already in Recursion’s pipeline	Recursion PR ∙ GEN ∙ C&EN
BoltzGen	November 2025	MIT	Generative AI for designing protein binders from scratch	Targets hard-to-treat disease targets; de novo molecule creation; nanomolar binders for 66% of novel targets	MIT News ∙ bioRxiv ∙ GEN
OpenFold3	October 2025	OpenFold Consortium	Open-source AlphaFold 3 competitor	Approaching AF3 parity; community-driven development	Nature
Pearl	October 2025	Genesis Molecular AI	Structure prediction optimized for drug discovery	Claims up to 40% improvement over AlphaFold 3 on key benchmarks; first evidence of scaling laws in molecular AI; NVIDIA collaboration	Business Wire ∙ Genesis ∙ MIT Tech Review
RoseTTAFold Diffusion 2/3	2025	David Baker Lab / UW	Protein design via diffusion models	RFdiffusion2 (April 2025): designs enzymes from functional group geometries; RFdiffusion3 (Dec 2025): 10x faster, atom-level precision	Nature Methods ∙ Chemistry World ∙ IPD
Schrödinger FEP+ Protocol Builder	2023 (active 2025)	Schrödinger	Active learning for free energy perturbation protocol optimization	Automates what was a manual, expertise-driven process; generates optimized models 4x faster than manual	Schrödinger ∙ J Chem Inf Model

 

Structure prediction has clean edges: sequence in, coordinates out, benchmark against experiment. The 2025 bet from frontier labs was whether the same logic could scale to messier problems—literature synthesis, experimental design, the part of research that lives in docs and Slack threads before it reaches a protocol.

2. Frontier labs build AI research assistants with real integrations

One of the most influential players in science AI is Google, whose researchers helped create both AlphaFold and the transformer architecture underlying most modern language models. In February 2025, Google unveiled AI Co-Scientist, a multi-agent system built on Gemini 2.0 (now a full generation behind the company’s latest LLM model) that generates hypotheses, designs experiments, and drafts research proposals. In one validation test, the system independently arrived at a bacterial gene transfer mechanism that Imperial College researchers had spent a decade confirming, in 48 hours.

Google is not alone in targeting the research workflow. OpenAI launched Deep Research the same month, a tool that synthesizes hundreds of scientific papers into cited reports in under an hour, essentially automating the literature review that consumes weeks of a PhD student’s life. Anthropic followed in October with Claude for Life Sciences, the company’s first formal scientific product, featuring direct integrations with Benchling, PubMed, 10x Genomics, and BioRender. At Novo Nordisk, the platform cut clinical study documentation from over 10 weeks to 10 minutes.

Major releases

Tool	Date	Developer	What it does	Why it matters	Sources
Google AI Co-Scientist	February 2025	Google DeepMind	Multi-agent system on Gemini 2.0; generates hypotheses, designs experiments, drafts proposals	Virtual research collaborator, not just an assistant; independently replicated decade-long research in 48 hours	Google Research ∙ arXiv ∙ HPCwire
OpenAI Deep Research	February 2025	OpenAI	Synthesizes hundreds of papers in <1 hour with audit trails	“AI PhD student” concept becomes tangible; powered by o3 model optimized for browsing/analysis	OpenAI ∙ Nature
OpenAI FrontierScience Benchmark	December 2025	OpenAI	Tests AI on expert-level reasoning in physics, biology, chemistry	700+ questions; Olympiad and Research tracks; GPT-5.2 scores 77% Olympiad, 25% Research	OpenAI ∙ TIME
Claude for Life Sciences	October 2025	Anthropic	Connectors to Benchling, PubMed, 10x Genomics, BioRender; skills for scRNA-seq QC	First formal life sciences product from a frontier lab	Anthropic ∙ CNBC
Anthropic AI for Science Program	May 2025	Anthropic	Up to $20K API credits for qualified researchers	Democratizes access for academic labs	TechCrunch ∙ Anthropic Support

 

Anthropic’s Benchling connector only works if your Benchling is populated. Google’s AI Co-Scientist can draft a proposal, but someone still has to know where the samples are. The bottleneck isn’t the model—it’s the metadata.

3. Lab informatics platforms add AI to core workflows

Laboratory information management systems (LIMS), electronic lab notebooks (ELNs), and inventory systems received significant upgrades in 2025. The difference from previous years: AI features moved from marketing bullet points to functionality that changes daily workflows. LabVantage 8.9 introduced voice command support through its Lottie/Open Talk system, unveiled at PittCon 2025. Sapio Sciences launched what it calls an “AI Lab Notebook” with molecular docking and ad hoc analytics built into the interface, demonstrated at SLAS 2025. Scispot positioned its LabOS as an API-first “Lab Operating System” connecting to 7,000+ applications.

Major releases

Tool	Date	Developer	What it does	Why it matters	Sources
LabVantage 8.9	March 2025	LabVantage Solutions	Voice command support (Lottie/Open Talk), AI-driven efficiency, database partitioning for speed	“SaaS 2.0” positioning; agentic AI integrations; unveiled at PittCon 2025	LabVantage PR ∙ Business Wire
LabWare Assure	April 2025	LabWare	New SaaS portfolio addition (joins GROW and QAQC)	Food safety and quality LIMS; expands cloud options for regulated environments; announced at PittCon 2025	PR Newswire ∙ LabWare
InstantGMP LIMS	October 2025	InstantGMP	Integrated QC module built into manufacturing platform	Purpose-built for GMP/GLP labs; unifies QC with production/inventory; Part 11 compliant	PR Newswire ∙ InstantGMP
Sapio ELaiN	September 2024 (GA); September 2025 (“3rd-gen ELN”)	Sapio Sciences	AI-native lab assistant; natural language data queries; “agentic” AI	First AI Lab Notebook (AILN); demonstrated at SLAS 2025; molecular docking, ad hoc analytics built-in	Sapio Sciences ∙ SLAS Blog
Scispot LabOS	Active 2025	Scispot	Unified ELN + LIMS + SDMS with AI, Jupyter integration	API-first “Lab Operating System” for biotech; no-code configuration; connects 7,000+ apps	Scispot ∙ LabOS blog

 

BayBE can pick your next experiment. It cannot pick your next experiment if it doesn’t know what you just ran, what failed, and what’s left in the freezer. That’s why LIMS upgrades and autonomous labs show up in the same year: one enables the other.

4. Self-driving labs move from demos to deployment

In 2025, autonomous labs moved from academic proof-of-concept to scaled deployment. A few products followed the footsteps of systems like Argonne National Laboratory’ Polybot, which launched in 2023. For instance, Lawrence Berkeley National Lab’s Distiller, released in April, streams electron microscope data to the Perlmutter supercomputer for real-time analysis. NC State’s Abolhasani lab published techniques in July that enable 10 times more data collection per experimental run through flow-driven data intensification.

The enabling software is maturing alongside the hardware. BayBE, an open-source Bayesian optimization framework from Merck KGaA and the University of Toronto Acceleration Consortium, has become core infrastructure for closed-loop experimentation. Sakana AI’s “AI Scientist,” released in August 2024 with collaborators from Oxford and UBC, demonstrated a complete autonomous research loop from idea generation through code, experiments, and paper drafting.

McKinsey estimated in 2023 that comprehensive automation could cut pharma R&D costs by approximately 25%. The pattern across these deployments: closed-loop systems connecting AI planners to robotic execution to automated analysis, with emphasis on resource savings and iteration speed.

Major developments

Tool/Initiative	Date	Developer	What it does	Why it matters	Sources
BayBE	December 2023 (active 2025)	Merck KGaA / U of Toronto Acceleration Consortium	Open-source Bayesian optimization for closed-loop experimentation	“Brain” for automated equipment; Apache 2.0 license; powers dozens of Merck use cases	Merck PR ∙ U of Toronto ∙ GitHub
The AI Scientist	August 2024	Sakana AI / U of Oxford / UBC	Autonomous research loop: idea → code → experiment → paper	Proof of concept for end-to-end AI research; ~$15/paper; open source	Sakana AI ∙ arXiv ∙ GitHub
NC State SDL techniques	July 2025	NC State (Abolhasani lab)	Flow-driven data intensification for 10x data collection at record speeds	Slashes costs, waste; Nature Chem Eng paper; “Rainbow” multi-robot SDL (Aug 2025)	NC State News ∙ ScienceDaily
Argonne Polybot / Autonomous Discovery	February 2025	Argonne National Lab	AI-driven automated materials laboratory; autonomous polymer thin-film processing	Federal lab investment in SDLs; Nature Communications paper; part of DOE Genesis Mission	Argonne ∙ UChicago News
Berkeley Lab Distiller	April 2025	Lawrence Berkeley National Lab	Web-based platform streaming electron microscope data to Perlmutter for real-time analysis	Enables experiment refinement while still running; DOE Superfacility concept	Berkeley Lab ∙ NERSC

 

NC State’s SDL collects 10x more data per run than previous setups. Berkeley’s 4D Camera generates 480 gigabits per second. Autonomy solves the throughput problem and creates the analysis problem. In genomics, that lands on bioinformatics.

5. Genomics data infrastructure scales for multiomics

The pipes that move genomic data from instruments to insights received critical upgrades in 2025. Illumina’s DRAGEN v4.4, released in May, delivered a 30% improvement in structural variant calling alongside oncology applications, multiomics support, and AWS F2 instance availability. The platform’s built-in AI/ML for variant calling has made it the industry standard for NGS secondary analysis.

On the corporate side, QIAGEN acquired Parse Biosciences in November for $225 million upfront plus $55 million in milestones, expanding its single-cell sequencing and bioinformatics capabilities. Parse’s Evercode platform serves more than 3,000 labs. The acquisition reflects a broader consolidation trend as vendors build end-to-end workflows from sample prep through analysis.

Major releases

Tool	Date	Developer	What it does	Why it matters	Sources
Illumina DRAGEN v4.4	May 2025	Illumina	30% improvement in SV calling; oncology apps; multiomics support; AWS F2 instances	Industry-leading NGS secondary analysis; built-in AI/ML for variant calling	PR Newswire ∙ Illumina
QIAGEN acquires Parse Biosciences	November 2025	QIAGEN	Expands single-cell sequencing + bioinformatics	$225M upfront + $55M milestones; Parse’s Evercode platform serves 3,000+ labs	QIAGEN PR ∙ Business Wire

---

6. Open-source scientific tools gain institutional backing

Press releases announce features. GitHub commits show who’s maintaining them. For a field where vendor claims outpace validation, public repositories are one of the few places you can check the work.

The Python scientific ecosystem continued maturing in 2025, with several packages crossing from “researcher tools” to “infrastructure.” The Boltz repository (jwohlwend/boltz) grew to more than 1,300 Slack community members and 200+ biotech adopters. OpenFold3 is approaching AlphaFold 3 parity with community-driven development and backing. Merck’s BayBE has become core to the company’s self-driving lab strategy while remaining fully open-source (also on GitHub).

Longer-standing tools maintained their central roles: scanpy remains the a common tool for single-cell RNA-seq analysis, and RDKit continues as the backbone of computational chemistry. AlphaFold itself has accumulated more than 43,000 citations, with weights released in November 2024.

Key repositories

Repo	What it is	2025 status	Link
jwohlwend/boltz	Boltz-1/2 protein structure + affinity	>1,300 Slack community; 200+ biotech adopters	GitHub
aqlaboratory/openfold	Open-source AlphaFold	OpenFold3 approaching AF3 parity	GitHub
Merck/BayBE	Bayesian experiment optimization	Core to Merck’s self-driving lab strategy	GitHub
scverse/scanpy	Single-cell analysis	Workhorse of scRNA-seq field	GitHub
rdkit/rdkit	Cheminformatics toolkit	Backbone of computational chemistry	GitHub
google-deepmind/alphafold	AlphaFold 2/3	43K+ citations; weights released Nov 2024	GitHub