Pinned Tweet
Getting answers is the easy part. Knowing how to trust them is where the real work begins.
AI Mentor doesn't just give you a destination; it maps out the journey.
Ask in natural language, explore the underlying logic, and verify with integrated citations.
✅ Traceable reasoning
✅ Citation-backed insights
✅ Strategic next steps
From complex literature synthesis to workflow optimization, move from curiosity to discovery with confidence.
See it in action. 🚀
#AIforScience #ResearchTools #AcademicAI #ScienceNavigator #Bohrium
2
8
978
A new academic year is getting underway, and it’s time to turn bolder questions into better research. 🎓
With Bohrium’s Education Offer, eligible students can save 40%:
– Plus: $9/month, normally $15
– Pro: $24/month, normally $40
More photons. More expert searches. More room for deep research.
Start the semester with Bohrium 👇
bohrium.com/en
5
198
Introducing DPA4C: a new performance frontier for universal machine-learning potentials
Designed for GPU-native molecular dynamics, DPA4C brings near-quantum accuracy into the speed range of classical empirical potentials, with broad chemical coverage and a model as small as 29,809 parameters.
Key results:
• Near-MACE-Omat accuracy at approximately 100× higher measured throughput
• Broad chemical coverage with a model as small as 29,809 parameters
• 14.05M atoms on one H20 and 2.048B atoms across 1,024 V100 GPUs
• 83.3%–91.2% weak-scaling efficiency, with up to 7.67B atoms per second
• Compared with NEP89: 61.4%, 48.1%, and 34.3% lower energy, force, and stress errors, respectively, at 1.92× its saturated throughput
This combination of accuracy, speed, and scale opens new possibilities for simulating grain boundaries, multi-principal-element alloys, reactive interfaces, and defect evolution.
Read the DPA4C paper:
arxiv.org/abs/2608.19041
#AIforScience #MaterialsScience #MachineLearning #MolecularDynamics #DPA4C
3
137
Today’s #POTD asks:
Can bird flu (H5N1) reach offspring before—or after—birth?
A new mouse study reports trimester-dependent spread through placental and mammary routes.
Paper:
bohrium.com/en/paper-details…
Analysis:
bohrium.com/en/chat/share/01…
1
1
133
During second-trimester-equivalent infection, researchers detected infectious virus—not only viral RNA—in the uterus, placenta and fetus.
That supports in-utero transmission of H5N1 in this mouse model.
After third-trimester-equivalent infection, infectious virus was found in mammary tissue, neonatal milk rings and lungs.
Offspring were also smaller and showed neurodevelopmental delays and later behavioral changes.
1
71
⚠️ The boundary matters:
This study used pregnant mice.
It does not show that H5 Bird Flu passes from pregnant or breastfeeding women to infants, and it does not establish clinical guidance.
If a virus can use different routes before and after birth, where should protection begin—the mother, the placenta or the newborn?
I asked Bohrium to unpack the evidence and limitations. Links are in post 1.
#BirdFlu #H5N1 #H5BirdFlu #AvianInfluenza #Virology
19
The preliminary ECCV 2026 accepted-paper list is now public.
Taking place in Malmö from September 8–12, ECCV 2026 reportedly received 10,473 submissions and accepted 2,883 papers—an estimated 27.5% acceptance rate.
Bohrium’s guide highlights early research from MMLab@NTU, the University of Tokyo, and Xiamen University, alongside papers from NVIDIA, Adobe, ByteDance, and Tencent Hunyuan.
The emerging themes extend across multimodal reasoning, 3D and spatial intelligence, controllable visual generation, embodied AI, and trustworthy vision systems.
Bohrium helps researchers search papers by topic, method, or institution, compare related approaches, and build a focused reading list. With Chat with Paper, they can move from a headline result to the methods, experiments, figures, and limitations supporting it.
Explore the ECCV 2026 research guide:
bohrium.com/en/blog/eccv-202…
#ECCV2026 #ComputerVision #MultimodalAI #AIResearch
2
145
MICCAI 2026 reported 1,211 acceptance decisions and a 27.5% acceptance rate. The conference will bring the medical image computing and computer-assisted intervention community to Strasbourg from September 27 to October 1.
With the complete proceedings still pending, Bohrium’s early guide examines 12 publicly announced papers across multimodal clinical models, surgical intelligence, robustness, reproducibility, and scalable medical imaging.
The emerging direction is clear: models are incorporating richer clinical context, trustworthiness is becoming measurable, and procedural AI is connecting perception with timely action.
Bohrium Science Navigator helps researchers find papers by clinical task, modality, method, and evidence level. They can then compare datasets, validation settings, and limitations before tracing key claims to the original paper with Chat with Paper.
Explore the MICCAI 2026 guide:
bohrium.com/en/blog/miccai-2…
#MICCAI2026 #MedicalImaging #SurgicalAI
2
142
DeepSeek Harness takes an “Everything is a Plugin” approach to AI agents.
I asked Bohrium to unpack its architecture, the role of Cordis, and why the harness layer matters.
Learn more: bohrium.com/en/chat/share/01…
Source: github.com/deepseek-ai/deeps…
#DeepSeek #AIAgents #OpenSource #DeveloperTools
1
156
Which task should AI remove first from a doctor’s workload?
Vote below, then tell us why.
We compared 10 specialist AI tools, and full guide linked in the comments👇
#ClinicalAI #MedicalAI #HealthcareAI #EvidenceBasedMedicine
100%Mapping medical evidence
0%Drafting clinical notes
0%Differential support
0%Reviewing imaging alerts
2 votes • Final results 1
2
91
AI tools now support evidence research, documentation, differential thinking, imaging, and care coordination. But they remove different kinds of work—and clinical judgment must remain human-owned. Explore all 10 tools:
bohrium.com/en/blog/best-ai-…
2
46
🦠 Today’s #POTD asks:
What if a computer could grow more processors while it works?
A new study presents a microfluidic computer that uses motile bacteria as self-multiplying computational agents.
The bacteria do not perform arithmetic inside their cells. Instead, they explore a microscopic network encoding the Subset Sum Problem.
At each junction, a bacterium follows one of two paths:
A. ➕ Include a number
B. ⏭️ Skip it
Each route represents a different numerical combination. Reaching a target exit signals a valid sum.
Why use bacteria?
E. coli and Vibrio natriegens can explore many routes in parallel—and V. natriegens can divide every 9–12 minutes under suitable conditions.
The experimental system solved instances up to N=7, with measured Pass-junction error rates of 1.3% for E. coli and 0.3% for V. natriegens.
⚠️ This is still a proof of concept. Claims about outperforming electronic computers at scale come from modeling, not a head-to-head benchmark.
If a computer could grow more processors, what would become its hardest limit?
A. Computing speed
B. Error control
C. Physical space and resources
Comment A, B or C—and tell us why.
I asked Bohrium to unpack the evidence and limitations. Explore the paper and the analysis here:
Bohrium: bohrium.com/en/paper-details…
Bohrium analysis:
bohrium.com/en/chat/share/01…
#POTD #Biocomputing #Microfluidics #SyntheticBiology #AIForScience
1
159
Explore KDD 2026 with Bohrium’s research guide: bohrium.com/en/blog/kdd-2026…
KDD 2026’s first proceedings volume includes 256 accepted papers—a research navigation challenge, not just a reading list.
As the conference unfolds in Jeju, a clear shift is emerging: AI is moving beyond prediction toward systems that can plan, use tools, evaluate evidence, and support scientific discovery.
From the first AI for Sciences Track and scientific replication agents to agentic data workflows and evidence-based analysis, this shift runs throughout the program.
For researchers, the key question is not only what was accepted, but which papers matter to their work—and what evidence supports the results.
Bohrium’s guide maps key research trends, keynotes, Big Tech papers, and KDD Cup challenges. Once researchers identify a relevant paper, they can explore it further with Chat with Paper by asking:
You can ask:
→ How does the method differ from its baselines?
→ Which dataset supports the improvement?
→ Which table or ablation provides the strongest evidence?
→ Where are the limits of the claim?
Bohrium grounds each conversation in the paper, with its figures, experiments, and sources available for verification.
From finding relevant work to examining its evidence, Bohrium turns a fast-moving conference into usable research knowledge.
#KDD2026 #AIforScience #ResearchAI #DataAgents #MachineLearning
1
192
💧 Today’s #POTD starts with the first water molecule.
How can a MOF harvest water from dry air?
A new Nature Communications paper examines bnn-1-Ni, a hydrolytically stable metal–organic framework designed to adsorb water under low-humidity conditions.
3 findings worth saving:
1️⃣ It starts capturing water below 20% RH.
That low uptake threshold matters for atmospheric water harvesting in dry environments.
2️⃣ It releases water at 60°C or below.
Low-temperature regeneration could reduce the heat required to run repeated adsorption–desorption cycles.
3️⃣ The first water molecule acts as an anchor.
Using single-crystal X-ray diffraction and DFT calculations, the researchers found a strongly bound initial water site that helps subsequent water layers form inside the pores.
Beyond this molecular mechanism, bnn-1-Ni combines fast loading, little hysteresis, low-temperature regeneration and stability over more than 100 cycles.
Under simulated temperature-swing conditions, its projected material-level water productivity reached 4.38 kg of water per kg of MOF per day.
⚠️ That figure is projected material productivity—not measured output from a complete atmospheric water harvesting device.
A useful water-harvesting material needs more than high uptake.
It must capture water at low humidity, release it with limited heat, cycle quickly and remain stable.
But when harvesting water from dry air, which becomes the real bottleneck first?
A. How much water it captures
B. How little heat it needs
C. How fast it can cycle
D. How long it remains stable
What would you prioritize—and why?
Reply with A, B, C or D in the comments.
I asked Bohrium to explore the same trade-off. Read the paper and its answer here 👇
Bohrium analysis:
bohrium.com/en/chat/share/01…
Bohrium:
l.dp-lib.com/e1LxLWo8
Nature:
nature.com/articles/s41467-0…
#POTD #MOFs #AtmosphericWaterHarvesting #MaterialsScience #WaterSecurity
3
3
209
🧬 Today’s #POTD is not just another liver-cancer target.
It is a molecular “mute button” for chemotherapy stress
How do tumor-initiating cells stay stem-like—and survive cisplatin—when damaged proteins should activate the unfolded protein response?
A new @NatureComms paper points to the p38β–BiP axis.
Three takeaways:
1️⃣ p38β marks a resistant cell state.
An integrated analysis of chemotherapy-enriched HCC spheroids and DepMap data prioritized p38β/MAPK11 as a kinase linked to stemness and chemoresistance.
Across 136 paired HCC samples, high p38β was associated with more aggressive disease and poorer survival.
2️⃣ It suppresses the cell’s stress alarm.
p38β phosphorylates the ER chaperone BiP at Thr648.
This strengthens BiP’s association with the UPR sensors PERK and IRE1α, reducing UPR activation and unfolded-protein accumulation under chemotherapy stress.
The result: tumor-initiating cells preserve stemness and resist cisplatin.
3️⃣ The pathway may be druggable.
Inhibiting BiP with HA15 restored UPR signaling and sensitized patient-derived xenograft and organoid models to cisplatin.
⚠️ Important caveat: this is preclinical evidence—not a new treatment for patients. HA15 was evaluated in HCC cells, mice, PDXs, and organoids.
The interesting shift may be this:
Chemoresistance is not only about repairing drug-induced damage.
It may also be about preventing the cell’s stress system from sounding the alarm.
I asked Bohrium to unpack the evidence and limitations. Links below 👇
🔬 Bohrium analysis:
bohrium.com/en/chat/share/01…
📘 Explore the paper on Bohrium:
l.dp-lib.com/qqTJf4G7
📄 Original paper in Nature Communications: nature.com/articles/s41467-0…
#LiverCancer #HCC #CancerStemCells #DrugResistance #ERStress
1
2
265
What if machine-learning interatomic potentials could deliver state-of-the-art accuracy—without the usual computational cost? ⚛️
Meet DPA4, a new equivariant interatomic-potential architecture built around EMFA SO(2) convolution. It pushes the accuracy–cost frontier across both inorganic crystals and organic molecules.
Explore its architecture, benchmark performance, and what DPA4 could mean for atomistic simulation 👇 bohrium.com/en/blog/dpa4-mac…
#AI4Science #MachineLearning #MaterialsScience #MolecularDynamics #DPA4 #Bohrium
3
158
🧪Today's #POTD asks a deceptively simple question:
If coding agents can rewrite scientific software, what makes the result scientifically trustworthy?
A new field report examines 8 real projects, from maintenance patches to full Rust and GPU rewrites.
Three takeaways:
1⃣ The gains can be substantial.
RustQC cut a reported 15h 34m sequence of QC tasks to 14m 54s on a 186-million-read dataset. A Rust rewrite of STAR reached 99.815% single-end and 99.883% paired-end agreement on the evaluated yeast reads.
2⃣ "It runs" is weak evidence.
Several agent-built systems produced plausible outputs while hiding changed defaults, skipped cases, or statistical defects. Tests had to compare released predictions, read-level behavior, posterior calibration, known answers, and real-world datasets.
3⃣ Human work moves up the stack.
In 7 of 8 cases, people remained the final judges: defining success, building validation harnesses, tracing discrepancies, and deciding whether a rewrite should merge upstream, live in a community project, or remain experimental.
⚠️ Important caveat: this is a retrospective field report, not a controlled benchmark. Results are contributor-reported and case-specific.
The real shift may be from "Can an agent write the code?" to "Who proves it is scientifically correct—and who maintains it afterward?"
I asked Bohrium to compare the evidence across all 8 cases. See its answer in the images, and explore the paper further with Bohrium. 👇
#AgenticAI #AI4Science #ScientificComputing #ResearchSoftware #ComputationalBiology #Bioinformatics #OpenSource #Reproducibility
2
154
Today's #POTD | Kimi K3: OPEN FRONTIER INTELLIGENCE.🚨
Moonshot AI released Kimi K3—the first open 3T-class monster rivaling Claude Fable 5 & GPT-5.6 Sol.
Highlights from the technical report:
🔥 2.8T params (104B active MoE) with 2.5x scaling efficiency
🧠 1M-token context & native multimodal vision
💻 Autonomous coding powerhouse (built its own GPU compiler!)
The real takeaway? The intelligence bottleneck is gone. The new challenge is infrastructure. How do you efficiently run and build complex agentic workflows around a massive 2.8T model?
How do you digest dense technical reports like Kimi K3 in seconds? Read it on Bohrium. 🚀
#KimiK3 #OpenSourceAI #Bohrium #LLMs #TechInfrastructure #FrontierAI
2
198
Try @Bohrium's new agent — BioMaster! 🧬
Computational biology doesn't break in theory—it breaks in execution.
Stop just chatting about biology with generic AI. BioMaster is built to RUN real computational biology tasks.
Why it’s a game-changer:
🏆 Top Tier: Beats Biomni AI with 97.73% on STELLA & 83.58% on Biomni-Eval1.
🧠 Smart: 307 biological skills distilled from 10k+ papers.
🛠️ Actionable: Schedules GPUs and calls tools (like Scanpy & scGPT) for multi-omics & paper reproduction.
📊 Traceable: Hands you 100% inspectable code, logs, and ready-to-review reports.
Stop wasting days fixing environment dependencies. Let BioMaster run the pipeline. 👇
🔗 Try it here: bohrium.com/en/agent/biomast…
📄 Learn More: bohrium.com/en/blog/biomaste…
#Bohrium #BioMaster #Bioinformatics #AIforScience #DrugDiscovery #CompBio
Announcing DPA4: Pushing the Accuracy–Cost Frontier of Interatomic Potentials with EMFA SO(2) Convolution
The Large Atomic Model (LAM) race is shifting from "who is bigger" to "who is faster, smarter, and more accessible".
💡 What is DPA4?
OpenLAM Team (AISI), Peking University, DP Technology, and IAPCM are excited to officially announce DPA4—a next-generation architecture built for the Large Atomic Model era.
🏆 Two World #1s: Matbench Discovery (Materials) & SPICE-MACE-OFF (Small Molecules)
⚡ 300x Efficiency Boost: Reaches eSEN SOTA accuracy in ~1 GPU-day on 1x RTX 5090 (down from 300+ GPU days)().
📦 1/10th Parameters: Less than 10% the size of eSEN.
🚀 World First: First ML potential to natively support torch.compile training (2-3x faster).
SOTA accuracy no longer requires a supercomputing budget. Atomic modeling is now a daily tool. 👇
Explore DPA4 & Get Involved:
•📄Read the Paper: arxiv.org/html/2606.02419v1
•💻Open-Source Assets & Early Access(via DeePMD-kit v3.2.0+): github.com/deepmodeling/deep…
•Hugging Face: huggingface.co/OutisLi/DPA4-…
•🤝Joint Research & Inquiries:
[email protected]
#AIforScience #ComputationalChemistry #MaterialsScience #DeepModeling #DPA4 #RTX5090 #MachineLearning
3
182








































