Biology & AI2024 · Nature
Predicting how life's molecules fit together
AlphaFold 3 broadens the types of molecular complexes that can be modelled. Better predictions support experiments; they do not replace them.
Journal paper
The research desk
Research and
Development
Understand the discovery. Follow the evidence. Read the original.
Search current biomedical and life-science research, follow original publications and investigate the evidence behind new ideas. For other disciplines, use the external repository tools below. Source-backed results refresh here; historical explainers remain clearly dated reference material.
Source-backed discovery
Explore current research
Search biomedical and life-science literature supplied by Europe PMC—not a comprehensive search across all sciences. Results load when you visit, search or refresh and are checked every five minutes while this page is visible. Identical searches share a five-minute cache to respect the source. Dates, abstracts and citations come from source metadata, not SCIENCIBLE-generated summaries.
Fetching research from Europe PMC…
Metadata can be incomplete or delayed. A journal-article label does not establish peer review; preprints are not necessarily reviewed. Full-text access varies. Abstracts below are source text, not an independent appraisal of methods or findings.
Start with the idea
Historical discoveries, with the evidence attached
Selected milestones, not breaking news. “Breakthrough” is an editorial description of significance—not an extra level of scientific proof. Publication years stay visible so older work is never presented as new.
Genomics2023 · Nature
A human genome reference with more than one path
A pangenome represents genetic alternatives rather than forcing every genome onto a single linear map. The first draft demonstrated improvements in variant discovery.
Journal paper
Computing & AI2017 · Advances in Neural Information Processing Systems 30
Attention becomes the architecture
The Transformer showed that attention alone could power an effective sequence-to-sequence model. Its original evidence came from machine translation—not every task now associated with AI.
Conference paper
Physics & Space2016 · Physical Review Letters
Listening to a black-hole merger
GW150914 provided the first direct detection of gravitational waves. The signal was observed in 2015 and the discovery paper was published in 2016.
Journal paper
Go to the source
Historical papers & preprints
Search this curated collection by title, author, topic, year or identifier. Publication type and access are different things: a freely readable paper may be peer-reviewed, while an openly posted preprint may not be.
5 of 5 research records · ordered by original publication year
Journal paperBiology & AI · 2024
Accurate structure prediction of biomolecular interactions with AlphaFold 3
Josh Abramson, Jonas Adler, Jack Dunger et al. · Nature
8 May 2024 · first published online
AlphaFold 3 predicts the joint structure of complexes containing proteins, DNA, RNA, small molecules and other molecular components. It extends structural modelling beyond protein-only systems.
Publication status: Peer-reviewed journal article. The publisher also links an addendum dated 27 November 2024; consult it for updated availability information.
How the work was done
The authors use a diffusion-based model to generate atomic coordinates and evaluate predictions against experimentally determined structures. Comparisons include protein–ligand, protein–nucleic-acid and antibody–antigen benchmarks; performance depends on the target and evaluation setup.
What the source reports
The study reports improved accuracy over the specialist methods used in several of its benchmark comparisons. It demonstrates that one modelling framework can address multiple types of biomolecular interaction, rather than requiring a separate predictor for each type.
Why it matters
A structural prediction can help researchers develop hypotheses about how molecules interact and choose follow-up experiments. That is useful for basic biology and drug-discovery research, but a predicted binding structure is not evidence that a treatment is safe or effective.
Limits & what not to conclude
The paper documents stereochemical errors, plausible-looking invented structure, and limited coverage of molecular conformations. Predictions are generally static structures, not a complete simulation of molecular motion. Experimental validation is still necessary.
Access: Open-access publisher article; software and model terms are separate.
Identifier: 10.1038/s41586-024-07487-w
Citation
Abramson, J., Adler, J., Dunger, J. et al. (2024). Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature, 630, 493–500. https://doi.org/10.1038/s41586-024-07487-w
Journal paperGenomics · 2023
A draft human pangenome reference
Wen-Wei Liao, Mobin Asri, Jana Ebler et al. · Nature
10 May 2023 · published online
Instead of treating one linear reference sequence as the only map of human DNA, the Human Pangenome Reference Consortium combined assemblies from genetically diverse individuals into a reference that can represent alternative sequences.
Publication status: Peer-reviewed journal article. This entry describes the 2023 draft, not the current extent of the consortium's reference resources.
How the work was done
The team assembled 47 phased, diploid genomes, assessed assembly quality and built graph-based representations. They compared variant-discovery workflows using the draft pangenome with workflows using the GRCh38 reference.
What the source reports
The paper reports the addition of 119 million base pairs of euchromatic polymorphic sequence relative to GRCh38. In its short-read analyses, the draft reduced small-variant discovery errors by 34% and increased the number of structural variants detected per haplotype by 104% compared with the GRCh38-based workflows studied.
Why it matters
A reference that represents more genetic variation can reduce the bias introduced by comparing everyone with a limited reference. It makes previously difficult regions and alternative alleles more accessible to analysis.
Limits & what not to conclude
This is a draft based on a finite cohort, not a catalogue of all human genetic diversity. The reported improvements apply to the tested workflows and datasets; they are not a universal improvement in diagnosis or a guarantee for every population.
Access: Open-access publisher article; data and analysis links are in the paper.
Identifier: 10.1038/s41586-023-05896-x
Citation
Liao, W.-W., Asri, M., Ebler, J. et al. (2023). A draft human pangenome reference. Nature, 617, 312–324. https://doi.org/10.1038/s41586-023-05896-x
Preprint / technical reportComputing & AI · 2023
GPT-4 Technical Report
OpenAI · arXiv · cs.CL / cs.AI
15 March 2023 · first submitted; version 6 dated 4 March 2024
OpenAI describes a multimodal model that accepts images and text and produces text. The report presents results on professional and academic benchmarks, alongside discussions of limitations and safety evaluations.
Publication status: Author-posted technical report. No peer-reviewed journal or conference publication is established by the arXiv record checked for this entry. Treat it as the developer's report, not independent validation.
How the work was done
The authors describe next-token pretraining and post-training alignment, benchmark evaluations, and experiments on predicting aspects of performance from smaller models. The full training recipe, model size and dataset construction are not disclosed.
What the source reports
The authors report strong performance on several benchmarks while explicitly noting that the model is less capable than humans in many real-world scenarios. These are reported evaluations of a particular model, not a guarantee of accuracy for a user's question.
Why it matters
The report is useful both as a historical account of a language-model release and as an example of how to read developer-provided evidence critically: examine the test conditions, missing information and acknowledged failure modes.
Limits & what not to conclude
The report acknowledges hallucinated facts and reasoning errors. Limited disclosure makes independent reproduction of the full system difficult. A benchmark score is not evidence of safe autonomous use or reliable professional advice.
Access: Freely readable technical report; arXiv's distribution licence is not an open-source model licence.
Identifier: arXiv:2303.08774v6
Citation
OpenAI (2023; version 6, 2024). GPT-4 Technical Report. arXiv:2303.08774v6. https://arxiv.org/abs/2303.08774v6
Conference paperComputing & AI · 2017
Attention Is All You Need
Ashish Vaswani, Noam Shazeer, Niki Parmar et al. · Advances in Neural Information Processing Systems 30
2017 · conference publication; arXiv first submitted 12 June 2017
The paper introduces the Transformer, an encoder–decoder architecture built around attention rather than recurrent or convolutional layers. Attention lets the model combine information from different positions in an input sequence.
Publication status: Peer-reviewed conference paper with an arXiv manuscript. Hosting on arXiv does not make the conference publication an unreviewed preprint. Manuscript versions differ; the arXiv record shows revision history.
How the work was done
The authors evaluate the architecture on English-to-German and English-to-French machine translation tasks. They compare translation quality and training requirements with earlier approaches, using positional information and multi-head attention to process sequences.
What the source reports
The authors report strong translation results alongside more parallelizable training. The important architectural contribution is a sequence model that uses attention without recurrence or convolution in its encoder–decoder design.
Why it matters
This provides a concrete starting point for understanding the architecture behind many subsequent language models. Read the original task, training setup and evaluation before generalising its claims to present-day AI systems.
Limits & what not to conclude
Translation benchmark performance does not establish reliable reasoning, factual accuracy or human-level understanding. The work is a specific architecture and evaluation, not evidence that every Transformer model will perform equally well. Numerical results also vary between the proceedings abstract and revised arXiv text; they are not mixed here.
Access: Freely readable proceedings and an author manuscript on arXiv; check each version's licence.
Identifier: arXiv:1706.03762
Citation
Vaswani, A., Shazeer, N., Parmar, N. et al. (2017). Attention Is All You Need. Advances in Neural Information Processing Systems, 30. Author manuscript: https://arxiv.org/abs/1706.03762
Journal paperPhysics & Space · 2016
Observation of Gravitational Waves from a Binary Black Hole Merger
B. P. Abbott et al. · LIGO Scientific Collaboration and Virgo Collaboration · Physical Review Letters
11 February 2016 · publication; signal observed 14 September 2015
Two LIGO detectors observed the same transient signal, designated GW150914. Its changing frequency and shape matched the predicted gravitational waves from two black holes spiralling together, merging and settling into a single remnant.
Publication status: Peer-reviewed journal paper. The arXiv record explicitly identifies Physical Review Letters 116, 061102 (2016) as the journal reference.
How the work was done
The collaboration analysed strain measurements from both detectors, compared the signal with waveforms predicted by general relativity, and estimated how often detector noise could produce a similar event.
What the source reports
The paper reports the first direct detection of gravitational waves and the first observation of a binary black-hole merger. It connects an observed signal with predictions for the inspiral, merger and ringdown phases of the system.
Why it matters
The observation established a way to study astrophysical systems through gravitational waves rather than light alone. It is a historical milestone, not a newly announced discovery.
Limits & what not to conclude
Masses and distance are inferred using waveform models and have reported uncertainties. One event does not determine how common all black-hole systems are, nor does agreement with a prediction prove every aspect of general relativity.
Access: Free author manuscript on arXiv; the record links the journal publication.
Identifier: 10.1103/PhysRevLett.116.061102
Citation
Abbott, B. P. et al. (LIGO Scientific Collaboration and Virgo Collaboration) (2016). Observation of Gravitational Waves from a Binary Black Hole Merger. Physical Review Letters, 116, 061102. https://doi.org/10.1103/PhysRevLett.116.061102
No subscription barrier
Find open-access research
Open-access research makes scholarly work available to read online without a subscription barrier. It helps students, independent researchers and curious readers reach the evidence behind a claim.
Access is not a quality rating. A paper being free does not prove its conclusions, and a paywall does not make a paper more reliable. Likewise, “free to read” does not automatically give permission to redistribute, adapt or commercially reuse it.
Read PMC's explanation of access and reuse (opens in a new tab)PubMed Central
Biomedical full text
NIH/NLM's free full-text archive for biomedical and life-sciences literature. Unlike PubMed's broader citation index, PMC holds full-text articles.
Free to read does not mean unrestricted reuse. Check the licence for the individual article.
arXiv
Preprints & author manuscripts
Research manuscripts in fields including physics, mathematics and computer science. Version histories and journal references help trace a paper from preprint to publication.
arXiv moderation is not peer review. Some records have later journal or conference publications; inspect the record.
DOAJ
Open-access journals & articles
The Directory of Open Access Journals indexes open-access scholarly journals and article records across disciplines. It is a useful starting point for finding open journals.
Index inclusion is not a guarantee that every finding is correct. Read the methods and check the journal's policies.
Keep your critical lens
How to read the evidence
01
Identify the document
Is this a research article, review, conference paper, preprint or press release? A headline is not a substitute for the original study. Check authors, venue and date.
02
Check publication status
A preprint is shared before formal peer review. An arXiv record may also hold a version of a published paper. Follow the journal reference and version history rather than guessing from the repository name.
03
Read methods before conclusions
What was actually studied: people, cells, a simulation or a benchmark? Look at the comparison group, measurements, sample selection and whether the design supports the claim.
04
Keep uncertainty visible
An association does not by itself establish a cause. An AI prediction is not an experiment, and a laboratory result is not a clinical treatment. Compare effect sizes and uncertainty, not just dramatic wording.
05
Follow versions and corrections
Check the publisher or repository for revisions, corrections, addenda or retractions. Peer review is a useful checkpoint, not a guarantee of correctness or independent replication.
06
Cite the version you read
Record the title, authors, year, venue and DOI or repository identifier. Check licences before reusing figures, text, code or data; their permissions can differ within the same project.
What this page does—and does not—claim
SCIENCIBLE's explanations are editorial summaries of the linked work, not a new peer review or an independent replication. Selection is illustrative, not a ranking. Current biomedical and life-science results are requested from Europe PMC on opening, searching and refreshing, and every five minutes while the page is visible. Identical searches share a five-minute cache; automatic refresh stops on source errors. Historical explainers retain their original source-check dates; they do not update automatically. Neither section offers citation-count ranking or personalised alerts. For current versions and notices, follow the original record.
Looking for a field, not a paper?
Explore Domains maps the subjects. This hub helps you examine their evidence.
Practical guide
Read Research with Context
Use publication metadata to find evidence, then examine the original work before drawing conclusions or applying a finding.
Scope & availability
Europe PMC retrieval is active for biomedical and life-science literature. Historical discoveries are a separately curated collection, not live search results.
01
A repeatable search strategy
Start with a short topic or author query. Select a document type and a date range, then decide whether relevance or publication date best matches your purpose. Broaden the range if a specialized search produces no matches.
Identical searches can use a five-minute cache. Refreshing does not force the provider to publish new records or establish that the database is exhaustive.
02
Interpret the record correctly
A title and abstract provide a compressed description. Check the population or system studied, comparison, measurement and limitations. Publication types are source labels; a journal-article label alone does not establish peer review.
Preprints can change and may not have been independently reviewed. Metadata can be delayed or incomplete; follow the original record for the current version.
03
Access and reuse are different
An abstract being visible does not establish permission to reuse full text, figures or data. Check the publisher or repository record for license terms and any access restrictions.
A copied citation helps attribution, but it is not a permission grant. Retain the source URL and relevant version when using a record in a project or explanation.
04
Move from reading to appraisal
Compare methods, sample size, uncertainty and competing explanations. A single result may motivate another question rather than settle it. Avoid converting a finding in a limited setting into general medical or technical advice.
When sources disagree, state the disagreement and look for differences in methods or scope instead of selecting only the preferred result.
Put it into practice
Choose a useful next step
Your next step
Prepare a focused inquiry
Choose what you want to discuss and add context. Continue to the contact form to review and send. Nothing is submitted here, and these selections are not saved as account progress.
Only include information suitable for a general contact inquiry.
Review inquiry