And that’s fine. Biology is complex. Progress doesn’t need to be elegant to move the field forward.

Progress in RNA structure will require:
• more thinking
• more diverse data types
• specialised models, not generic LLMs
• wet-lab feedback
• acceptance that solutions will be messy and not universally generalisable

This tells us something important: unresolved RNA sequences near known templates may be the next practical targets.
In RNA, the “general-purpose AI breakthrough” is not here yet. 8/8

Another example: Stanford RNA folding competitions (Rhiju Das) - www.kaggle.com/competitions...
AI models still lag far behind the level of protein prediction. The best scores rely on old-school templates. Human intuition still matters. 7/8

The recent Arc Institute challenge made this painfully clear: defining evaluation metrics is hard. In some cases, trivial data transformations—and even random data—can score astonishingly high.

Great AI performance ≠ biological meaning. 6/8

Experiments must be in the loop. Automation helps, but the real bottleneck is choosing the right optimisation targets and defining what “good” looks like.

AlphaFold succeeded because protein structures could be validated experimentally. Not all biological tasks have that luxury. 5/8

The second direction: post-AlphaFold biology. Architectures built specifically for molecules, not NLP models.

To replicate AlphaFold’s impact, we need:
• high-quality experimental benchmarks
• sufficient data volume
• new architectures
• tight integration with wet-lab validation 5/8

Those harder problems require better questions, richer data, and meaningful evaluation metrics. Biology still doesn’t provide clean, well-defined AI benchmarks, such as NLP or vision. 4/8

Breakthroughs will come here, but this is a space dominated by the biggest players: Anthropic, Google, OpenAI. Useful—but not enough to crack the hardest biological problems. 3/8

First: the virtual bioinformatician—agentic AI systems that automate routine analyses and even propose new hypotheses by combining LLMs with sequence and structure models. 2/8

Mile Sikic @msikic.bsky.social presents "AI for genomes—Rethinking de novo assembly" genome.cshlp.org/content/35/4/839
They devised a bidirectional message-passing procedure in GNN for the problem of genome assembly

📅 Deadline: 30 Nov 2025
📧 Send your CV, research statement, & brief PhD topic to → mile_sikic@a-star.edu.sg

Let’s redefine how AI understands life itself. 🌱

#AIforScience #PhDopportunity #Singapore #Genomics

5/5

💰 What ARAP offers:
– S$3,600 / month stipend
– ✈️ Airfare & settling-in allowance
– 🏡 Housing, 🩺 insurance, 💻 IT & 🎤 conference support

Work at the intersection of AI × biology × medicine with a vibrant international team.

#AI #Genomics #PhDlife

4/5

🎯 We’re looking for:
– PhD students in AI / ML / algorithms
– Strong coding & scientific curiosity
– A home supervisor open to collaboration

Shorter research visits are also possible — limited lab funding available.

#AIForgood #Bioinformatics

3/5

🌍 Through the A*STAR Research Attachment Programme (ARAP) — spend 1–2 years at A*STAR labs, co-supervised with your home university.

🔗 a-star.edu.sg/Scholarships...

#Research #Singapore
2/5

Biology is complex, and better problem formulation and the right data is the path to success.

The question of what data is right remains open. 2/2