Bio Co-Scientist
Bio Co-Scientist
In HyperLab's Bio Co-Scientist Lab space,
research, design, and structure validation
happen in a single flow.
Start your research with one line of chat.
01
Turn your idea
into design results
Generate multiple candidates from a single research idea, compare them side by side, and steer toward the direction you want
— cutting research time at every step.
02
Hundreds of structure predictions, faster
K-Fold runs hundreds of structure predictions at once, averaging 30 seconds per complex on an NVIDIA L40S GPU, saving researchers time and cost.
03
Search, predict, and analyze
— All in one place
Run your entire workflow through conversation: from literature search to structure prediction to interpreting results, all in one uninterrupted flow.
K-Fold — the model that predicts structure
HyperLab predicts structures with K-Fold*,
a next-generation bio foundation model co-developed by KAIST and HITS.
*K-Fold is a complex-folding model that predicts the 3D shapes biomolecules fold into. Working from sequence alone
— no experimental structures like X-ray crystallography required
— it predicts not just standalone protein structures but also complexes such as protein–small molecule and protein–nucleic acid, ready to use right away for binding-site analysis and binder or antibody design.
Broad research extensibility
Predicts binding structures based on the principles of physical and chemical interactions inherent to molecules.
Demonstrates outstanding performance in designing new modalities such as antibodies, TPD, and synthetic binders, where natural evolutionary information is scarce.
Principle-based prediction precision
Goes beyond simple sequence pattern recognition to take into account physical and chemical interactions occurring inside and outside proteins, as well as dynamic structural changes during the binding process, and predicts precise complex structures.
High-speed co-folding in seconds
With high-speed prediction averaging 30 seconds per complex on an NVIDIA L40S GPU, you can quickly narrow down candidates.
30s
Average inference time per complex (on an L40S GPU)
Multi-state
Prediction of various states of key targets such as kinases and GPCRs
Apo-to-holo
Precise prediction of challenging targets such as KRAS and Src based on unique principles
Your entire design workflow, in one flow
Antibody design
1
Pinpoint the target antigen's binding site
Find the antigen structure through research chat and identify where binding can occur.
2
Pull in a suitable antibody sequence
Retrieve the sequence that gives your design a starting point.
3
Design the CDR backbone with RFdiffusion
Build the backbone of the CDR loops to fit the binding site.
4
Design the sequence with ProteinMPNN
Generate the amino acid sequence to fill the designed backbone.
5
Predict the antigen–antibody complex with K-Fold
Predict how your designed antibody binds the target antigen, then review it in the 3D Viewer.
6
Screen candidates with ipTM confidence scores
Use antigen–antibody binding confidence as a metric to surface your most promising antibody candidates.
Every step happens in one place
— no shuffling sequences or results between tools.
Experience protein and antibody design research without the friction.
Every step happens in one place — no shuffling sequences or results between tools.Experience protein and antibody design research without the friction.
We’ve got the answers
How does it differ from AlphaFold3?
The approach and principles are different.
While existing co-folding models rely heavily on evolutionary statistics from multiple sequence alignment (MSA), K-Fold predicts binding structures based on the physical and structural features of the molecules themselves, without MSA.
In particular, unlike existing approaches that learn only the final static structure, it learns structural changes that occur during the binding process. As a result, it demonstrates robust generalization performance for next-generation drug modalities such as antibodies and TPD, where natural evolutionary information is scarce, and for protein targets such as kinases and GPCRs, where structural changes are important.
In addition, with no MSA search bottleneck and the use of optimized computational kernels, ultrafast prediction averaging 30 seconds per complex is possible. (on an L40S GPU)
Can the accuracy be trusted?
Reliability is verified through sophisticated benchmark evaluations and multifaceted confidence metrics.
K-Fold shows top-level accuracy not only on representative complex structure benchmarks, but also in predictions of antigen–antibody complexes, TPD ternary complexes, and structural changes in kinases and GPCRs.
In addition, it provides confidence metrics for complexes and binding interfaces, such as pLDDT, PAE, pTM, and ipTM, along with prediction results, allowing researchers to critically validate the prediction results.
What molecules and complexes can it predict?
Protein–protein and protein–small molecule complexes, nucleic acids (DNA/RNA), and
ions — and it predicts complexes that combine all of these in a single run.
What are K-Fold's input and output formats?
Inputs: protein, DNA, and RNA sequences, small-molecule SMILES strings, and ion specifications.
Outputs: CIF and PDB formats together, returning 5 conformers per structure prediction by default.
Do I need other tools to design with K-Fold?
No. Strategy research, sequence retrieval, backbone and sequence design, and structure prediction all run in one place, end to end — no preparing sequences by hand, no moving results between tools.
Is the prediction speed fast enough for use in actual research?
Predictions are completed in approximately 30 seconds per complex on average on an NVIDIA L40S GPU. (excluding delays due to infrastructure factors)
The MSA search step, which used to take tens of minutes, is absent, and accelerated computational kernels are also applied, so predictions are completed in an average of 30 seconds for typical complexes and around two minutes even for large complexes.
Thanks to this ultrafast prediction speed, it can be immediately used in virtual screening workflows that validate numerous candidate compounds in research settings.
Is my uploaded research data used to train the AI?
No. Your data and analysis results are never used for model training.
How is my data protected?
Production and development servers are physically separated. Access to production servers is secured with OTP and a VPC-based private network, with no external access.
Ready to revolutionize
Your drug discovery process?
Sign up with your email and start using HyperLab right now
—
AI drug discovery Platform
HITS Inc.
CEO : Woo Youn Kim
Address : 8F, 28, Teheran-ro 4-gil, Gangnam-gu, Seoul, Republic of Korea
© HITS Inc. All rights reserved.
1. Turn your idea into design results
Generate multiple candidates from a single research idea, compare them side by side, and steer toward the direction you want — cutting research time at every step.
2. Hundreds of structure predictions, faster
K-Fold runs hundreds of structure predictions at once, averaging 30 seconds per complex on an NVIDIA L40S GPU, saving researchers time and cost.
3. Search, predict, and analyze
— All in one place
Run your entire workflow through conversation: from literature search to structure prediction to interpreting results, all in one uninterrupted flow.
Bio Co-Scientist
Bio Co-Scientist
In HyperLab's Bio Co-Scientist Lab space,
research, design, and structure validation happen in a single flow.
Start your research with one line of chat.
Bio Co-Scientist
Bio Co-Scientist
In HyperLab's Bio Co-Scientist Lab space,
research, design, and structure validation happen in a single flow.
Start your research with one line of chat.
K-Fold Model
K-Fold — the model that predicts structure
HyperLab predicts structures with K-Fold*,
a next-generation bio foundation model
co-developed by KAIST and HITS.
*K-Fold is a complex-folding model that predicts the 3D shapes biomolecules fold into. Working from sequence alone
— no experimental structures like X-ray crystallography required
— it predicts not just standalone protein structures but also complexes such as protein–small molecule and protein–nucleic acid, ready to use right away for binding-site analysis and binder or antibody design.
1. Broad research extensibility
K-Fold's prediction algorithm draws on a protein's own physical and chemical properties instead of relying on evolutionary information — so it works even where that data is scarce, as with synthetic binders or molecular glues.
2. Proven accuracy
In predicting the 3D structures of biomolecular complexes
— proteins, small molecules, and nucleic acids
— K-Fold reaches accuracy close to AlphaFold3.
3. Faster results
Up to ~30× faster than AlphaFold3, so you can narrow down candidates quickly.
K-Fold
Workflow
Your entire design workflow,
in one flow
Antibody design
1
Pinpoint the target antigen's binding site
Find the antigen structure through research chat and identify where binding can occur.
2
Pull in a suitable antibody sequence
Retrieve the sequence that gives your design a starting point.
3
Design the CDR backbone with RFdiffusion
Build the backbone of the CDR loops to fit the binding site.
4
Design the sequence with ProteinMPNN
Generate the amino acid sequence to fill the designed backbone.
5
Predict the antigen–antibody complex with K-Fold
Predict how your designed antibody binds the target antigen, then review it in the 3D Viewer.
6
Screen candidates with ipTM confidence scores
Use antigen–antibody binding confidence as a metric to surface your most promising antibody candidates.
Protein binder design
1
Pinpoint the target protein's binding site
Find the target structure through research chat and identify where the binder should attach.
2
Design the backbone with RFdiffusion
Build the binder's backbone to fit the binding site.
3
Design the sequence with ProteinMPNN
Generate the amino acid sequence for the designed backbone.
4
Predict the target–binder complex with K-Fold
Predict how your designed binder binds the target, then review the complex in the 3D Viewer.
5
Screen candidates by confidence score
Evaluate the target–binder interface with ipTM to surface your most promising candidates.
Every step happens in one place
— no shuffling sequences or results between tools.
Experience protein and antibody design research without the friction.
Every step happens in one place
— no shuffling sequences or results between tools.
Experience protein and antibody design research without the friction.
We’ve got the answers
FAQ
How does it differ from AlphaFold3?
The approach and principles are different.
While existing co-folding models rely heavily on evolutionary statistics from multiple sequence alignment (MSA), K-Fold predicts binding structures based on the physical and structural features of the molecules themselves, without MSA.
In particular, unlike existing approaches that learn only the final static structure, it learns structural changes that occur during the binding process. As a result, it demonstrates robust generalization performance for next-generation drug modalities such as antibodies and TPD, where natural evolutionary information is scarce, and for protein targets such as kinases and GPCRs, where structural changes are important.
In addition, with no MSA search bottleneck and the use of optimized computational kernels, ultrafast prediction averaging 30 seconds per complex is possible. (on an L40S GPU)
Can the accuracy be trusted?
Reliability is verified through sophisticated benchmark evaluations and multifaceted confidence metrics.
K-Fold shows top-level accuracy not only on representative complex structure benchmarks, but also in predictions of antigen–antibody complexes, TPD ternary complexes, and structural changes in kinases and GPCRs.
In addition, it provides confidence metrics for complexes and binding interfaces, such as pLDDT, PAE, pTM, and ipTM, along with prediction results, allowing researchers to critically validate the prediction results.
What molecules and complexes can it predict?
Protein–protein and protein–small molecule complexes, nucleic acids (DNA/RNA), and
ions — and it predicts complexes that combine all of these in a single run.
What are K-Fold's input and output formats?
Inputs: protein, DNA, and RNA sequences, small-molecule SMILES strings, and ion specifications.
Outputs: CIF and PDB formats together, returning 5 conformers per structure prediction by default.
Do I need other tools to design with K-Fold?
No. Strategy research, sequence retrieval, backbone and sequence design, and structure prediction all run in one place, end to end — no preparing sequences by hand, no moving results between tools.
Is the prediction speed fast enough for use in actual research?
Predictions are completed in approximately 30 seconds per complex on average on an NVIDIA L40S GPU. (excluding delays due to infrastructure factors)
The MSA search step, which used to take tens of minutes, is absent, and accelerated computational kernels are also applied, so predictions are completed in an average of 30 seconds for typical complexes and around two minutes even for large complexes.
Thanks to this ultrafast prediction speed, it can be immediately used in virtual screening workflows that validate numerous candidate compounds in research settings.
Is my uploaded research data used to train the AI?
No. Your data and analysis results are never used for model training.
How is my data protected?
Production and development servers are physically separated. Access to production servers is secured with OTP and a VPC-based private network, with no external access.
Ready to revolutionize
Your drug discovery process?
Sign up with your email and start using HyperLab right now
HITS Inc.
CEO : Woo Youn Kim
Address : 8F, 28, Teheran-ro 4-gil, Gangnam-gu,
Seoul, Republic of Korea
© HITS Inc. All rights reserved.
Ready to revolutionize
Your drug discovery process?
Sign up with your email and start using HyperLab right now
HITS Inc.
CEO : Woo Youn Kim
Address : 8F, 28, Teheran-ro 4-gil, Gangnam-gu,
Seoul, Republic of Korea
© HITS Inc. All rights reserved.
Ready to revolutionize
Your drug discovery process?
Sign up with your email and start using HyperLab right now
Ready to revolutionize
Your drug discovery process?
Sign up with your email and start using HyperLab right now
—
AI drug discovery Platform
Small Molecule Discovery
Publications
HITS Inc.
CEO : Woo Youn Kim
Address : 8F, 28, Teheran-ro 4-gil, Gangnam-gu, Seoul, Republic of Korea
© HITS Inc. All rights reserved.
Bio Co-Scientist
Bio Co-Scientist
In HyperLab's Bio Co-Scientist Lab space,
research, design, and structure validation happen in a single flow.
Start your research with one line of chat.
1. Turn your idea into design results
Generate multiple candidates from a single research idea, compare them side by side, and steer toward the direction you want — cutting research time at every step.
2. Hundreds of structure predictions, faster
K-Fold runs hundreds of structure predictions at once, averaging 30 seconds per complex on an NVIDIA L40S GPU, saving researchers time and cost.
3. Search, predict, and analyze — All in one place
Run your entire workflow through conversation: from literature search to structure prediction to interpreting results, all in one uninterrupted flow.
K-Fold — the model that predicts structure
K-Fold Model
HyperLab predicts structures with K-Fold*, a next-generation bio foundation model
co-developed by KAIST and HITS.
*K-Fold is a complex-folding model that predicts the 3D shapes biomolecules fold into. Working from sequence alone
— no experimental structures like X-ray crystallography required
— it predicts not just standalone protein structures but also complexes such as protein–small molecule and protein–nucleic acid, ready to use right away for binding-site analysis and binder or antibody design.
1. Broad research extensibility
K-Fold's prediction algorithm draws on a protein's own physical and chemical properties instead of relying on evolutionary information — so it works even where that data is scarce, as with synthetic binders or molecular glues.
2. Proven accuracy
In predicting the 3D structures of biomolecular complexes
— proteins, small molecules, and nucleic acids
— K-Fold reaches accuracy close to AlphaFold3.
3. Faster results
Up to ~30× faster than AlphaFold3, so you can narrow down candidates quickly.
K-Fold
Your entire design workflow, in one flow
Workflow
Antibody design
1
Pinpoint the target antigen's binding site
Find the antigen structure through research chat and identify where binding can occur.
2
Pull in a suitable antibody sequence
Retrieve the sequence that gives your design a starting point.
3
Design the CDR backbone with RFdiffusion
Build the backbone of the CDR loops to fit the binding site.
4
Design the sequence with ProteinMPNN
Generate the amino acid sequence to fill the designed backbone.
5
Predict the antigen–antibody complex with K-Fold
Predict how your designed antibody binds the target antigen, then review it in the 3D Viewer.
6
Screen candidates with ipTM confidence scores
Use antigen–antibody binding confidence as a metric to surface your most promising antibody candidates.
Protein binder design
1
Pinpoint the target protein's binding site
Find the target structure through research chat and identify where the binder should attach.
2
Design the backbone with RFdiffusion
Build the binder's backbone to fit the binding site.
3
Design the sequence with ProteinMPNN
Generate the amino acid sequence for the designed backbone.
4
Predict the target–binder complex with K-Fold
Predict how your designed binder binds the target, then review the complex in the 3D Viewer.
5
Screen candidates by confidence score
Evaluate the target–binder interface with ipTM to surface your most promising candidates.
Every step happens in one place
— no shuffling sequences or results between tools.
Experience protein and antibody design research without the friction.
Every step happens in one place
— no shuffling sequences or results between tools.
Experience protein and antibody design research without the friction.
How does it differ from AlphaFold3?
The approach and principles are different.
While existing co-folding models rely heavily on evolutionary statistics from multiple sequence alignment (MSA), K-Fold predicts binding structures based on the physical and structural features of the molecules themselves, without MSA.
In particular, unlike existing approaches that learn only the final static structure, it learns structural changes that occur during the binding process. As a result, it demonstrates robust generalization performance for next-generation drug modalities such as antibodies and TPD, where natural evolutionary information is scarce, and for protein targets such as kinases and GPCRs, where structural changes are important.
In addition, with no MSA search bottleneck and the use of optimized computational kernels, ultrafast prediction averaging 30 seconds per complex is possible. (on an L40S GPU)
Can the accuracy be trusted?
Reliability is verified through sophisticated benchmark evaluations and multifaceted confidence metrics.
K-Fold shows top-level accuracy not only on representative complex structure benchmarks, but also in predictions of antigen–antibody complexes, TPD ternary complexes, and structural changes in kinases and GPCRs.
In addition, it provides confidence metrics for complexes and binding interfaces, such as pLDDT, PAE, pTM, and ipTM, along with prediction results, allowing researchers to critically validate the prediction results.
What molecules and complexes can it predict?
Protein–protein and protein–small molecule complexes, nucleic acids (DNA/RNA), and
ions — and it predicts complexes that combine all of these in a single run.
What are K-Fold's input and output formats?
Inputs: protein, DNA, and RNA sequences, small-molecule SMILES strings, and ion specifications.
Outputs: CIF and PDB formats together, returning 5 conformers per structure prediction by default.
Do I need other tools to design with K-Fold?
No. Strategy research, sequence retrieval, backbone and sequence design, and structure prediction all run in one place, end to end — no preparing sequences by hand, no moving results between tools.
Is the prediction speed fast enough for use in actual research?
Predictions are completed in approximately 30 seconds per complex on average on an NVIDIA L40S GPU. (excluding delays due to infrastructure factors)
The MSA search step, which used to take tens of minutes, is absent, and accelerated computational kernels are also applied, so predictions are completed in an average of 30 seconds for typical complexes and around two minutes even for large complexes.
Thanks to this ultrafast prediction speed, it can be immediately used in virtual screening workflows that validate numerous candidate compounds in research settings.
Is my uploaded research data used to train the AI?
No. Your data and analysis results are never used for model training.
How is my data protected?
Production and development servers are physically separated. Access to production servers is secured with OTP and a VPC-based private network, with no external access.
We’ve got the answers
FAQ
30×
Average inference time per complex (on an L40S GPU)
AF3-level
Prediction of various states of key targets such as kinases and GPCRs
No MSA
Precise prediction of challenging targets such as KRAS and Src based on unique principles