đ Sneak peek: our upcoming generative model for molecular design (preprint coming soon!)
At Klyne, weâre building one of the most advanced platforms for hit-to-lead optimization. Our generative AI engine is designed to:
đ Enhance binding affinity
đ Optimize ADMET properties
đ Maintain synthesizability & stability
What makes it powerful is its flexibility: the model can integrate with any scoring frameworkâfrom docking and rapid simulations to full free energy perturbation (FEP)âto guide design decisions with precision.
đŹ Example: starting from Imatinib (a well-known BCRâABL inhibitor), we generated novel derivatives predicted to engage the same target. These were not cherry-picked for performanceâjust molecules that looked interesting from an early run.
Weâre excited to share more details soonâthis work is being carried out in collaboration with @robpollice at the University of Groningen.
Stay tuned for the preprint!
Have proteinâligand cofolding methods moved beyond memorization?
It is well recognized in AI research that many models excel at memorizing patterns from training data, but struggle to generalize to truly novel cases. This limitation is especially consequential in drug discovery, where predicting new proteinâligand interactions is essential for de novo design.
A recent study presents the first comprehensive benchmarking of four leading cofolding tools on Runs Nâ Posesâa dataset carefully curated to reduce overlap with existing training sets. The findings highlight clear limitations:
đĄ Model predictions remain highly correlated with known examples, indicating reliance on memorization.
đĄGeneralization to unseen complexesâparticularly drug-like moleculesâremains weak.
đĄPerformance improves primarily for prevalent molecules with extensive prior representation (e.g., cofactors, nucleotide analogs).
These results emphasize that while deep learning has advanced protein structure prediction, cofolding methods for proteinâligand interactions have not yet achieved the level of generalization required for innovative drug discovery.
The authors underscore the need for:
đImproved model architectures capable of learning transferable principles
đRobust data augmentation strategies
đDiverse, high-quality structural datasets shared across the community
đRigorous benchmarking standards that reflect real-world use cases
Together, these steps will be critical for moving beyond pattern recognition toward genuine predictive innovation in computational drug discovery.
Read the full pre-print here: https://t.co/BJA280VO37
Why Simplicity Still Wins in Molecular Machine Learning: Highlights from the Largest Embedding Benchmark Yet
Rull article here: https://t.co/nWoXYu4fI1
In the rapidly evolving field of molecular machine learning, choosing the right features for representation learning remains one of the biggest challenges â and recent research reveals it is still more art than science.
A new study, âBenchmarking Pretrained Molecular Embedding Models For Molecular Representation Learningâ (Praski, Adamczyk, & Czech, arXiv:2508.06199), delivers the largest head-to-head comparison to date. The team rigorously tested 25 pretrained molecular embedding models against real-world chemical tasks, spanning everything from ADMET property prediction to virtual screening, using 25 diverse benchmark datasets. To ensure a fair comparison, all models leveraged a âfrozenâ embedding setup â meaning no model could benefit from task-specific retraining or fine-tuning.
The findings were strikingly counterintuitive. Almost every neural networkâbased model failed to outperform the classic ECFP molecular fingerprint â a handcrafted feature engineering approach widely used in cheminformatics. Only CLAMP, a model based on molecular fingerprints, achieved consistent and statistically significant improvement.
âNearly all neural models show negligible or no improvement over the baseline ECFP molecular fingerprint. Only the CLAMP model (âŠ) performs statistically significantly better than the alternatives.â
This study raises important concerns about the way molecular ML models are evaluated and calls for greater rigor in benchmarking. The lesson is clear: In molecular machine learning, sophisticated architectures and pretraining arenât enough. Deep domain knowledge and thoughtful feature design â as embodied by traditional fingerprints â still play a decisive role.
Takeaway for practitioners and researchers: Donât underestimate the power of well-crafted chemical fingerprints; sometimes, the simplest approach delivers the strongest results in molecular ML, even as deep learning models continue to proliferate.
â Mateusz Praski, Jakub Adamczyk, Wojciech Czech, âBenchmarking Pretrained Molecular Embedding Models For Molecular Representation Learning,â arXiv:2508.06199
KlyneAI: Democratizing HighâPerformance AI Drug Discovery
In the past two months, @KlyneAI has welcomed over 200 new followers to our brand journey and community. If you're among the newcomers - or just curious about what we do - hereâs a look under the hood as we accelerate earlyâstage drug discovery with AI at dramatically lower cost.
Breaking the Bottleneck in Early Drug Discovery
Traditional drug discovery is prohibitively costly ($1â2âŻbillion) and timeâintensive (15+ years). Startups, academic labs, and emerging biotech ventures often lack access to the infrastructure or budgets needed for ultraâlarge screening, affinity estimation, and optimization.
Thatâs where @KlyneAI comes in.
Founded in late 2023 with roots at Stanford and energy from JLABS, our leadership team blends deep expertise in machine learning, computational chemistry, and biotech entrepreneurship.
We set out to make drug discovery affordable and scalable: helping small teams go from hundreds of thousands of dollars to tens of thousands, an orderâofâmagnitude reduction in cost.
How We Deliver Value
1. UltraâLarge Virtual Screening + Active Learning
Screen billions of compounds with active learning to reduce compute by ~90%. Use KlyneDock, an AI-enhanced docking algorithm that cuts false positives by ~50%.
2. Precise, Affordable Affinity Estimation
Our HYRDA technology delivers free-energyâlike affinity accuracy at just ~5% of traditional FEP cost. KlyneâFEP achieves almost equal precision to industry standards at ~95% lower cost.
3. Generative AI for HitâtoâLead Optimization
Fineâtune analog designs with multiâparameter optimization: targeting binding affinity, ADMET, selectivity, synthetic accessibility, and patentability.
4. Flexible Business Model & HubâandâSpoke Service
No recurring subscriptions or library licenses - work is done on a pay-as-you-go basis, accessible for lean companies.
We also provide a white-glove service, adapting workflows to project needs and handling technical complexity for you.
Enabling Affordability at Scale - Our Key Partnerships
A critical driver of cost-efficiency was sourcing scalable GPU compute. Leveraging SaladCloud's consumer GPU network, we reduced compute costs by more than 50%, allowing us to run hundreds of thousands of molecular dynamics simulations without traditional cloud costs.
In collaboration with SaladCloud, we have successfully explored trillions of compounds, pushing the limits of accessible virtual drug screening.
Weâre evolving the drug discovery landscape with tools that democratize access and make early-stage discovery a reality for lean, mission-driven teams.
From screening to lead optimization, we're covering the full pipeline - tailored, smart, and cost-effective.
If you're interested in collaborating, exploring a pilot project, or learning more about our tech roadmap, reach out at [email protected].
To read more âȘ https://t.co/XFr3KxIEGw
Welcome to the future of biotech.
Enhancing Ligand Binding Predictions: The Power of Molecular Dynamics in Drug Discovery
We get really excited by advancements in MD ranking here at @KlyneAI because our mission is to deliver highly accurate, affordable virtual screening and early-stage lead optimization for clients who need reliable compound prioritization without massive compute budgets.
A landmark study by Steinbrecher et al. showcases how integrating molecular dynamics (MD) with Free Energy Perturbation (FEP) and the OPLS3 force field significantly improves the accuracy of ligand binding predictions across diverse pharmaceutical targets.
Unlike traditional docking, this MD-informed FEP approach captures protein-ligand flexibility, solvent effects, and entropic contributions, leading to up to 6x enrichment of true positives and correlations up to R = 0.89 with experimental data.
We observe something similar in HYDRA, an internal tool in development at Klyne, where we use short MD simulations to rapidly estimate binding affinity.
Early results show promising alignment with more resource-intensive methods - helping us push the boundaries of speed and accuracy in virtual screening.
In essence, the study validates that considering the dynamic nature of molecular interactions - not just static poses - leads to better decision-making in lead optimization and compound prioritization.
By leveraging molecular dynamics to more faithfully capture protein-ligand flexibility and interaction nuance, we drastically reduce false positives and meaningfully improve hit rates, empowering our partners to make smarter, lower-risk R&D decisions and accelerate discovery with a fraction of the resources required by traditional methods.
đ Read the full article: https://t.co/SFRA3qERtz
đ Learn how KlyneAI accelerates early-stage drug discovery at low cost with AI: https://t.co/XFr3KxIEGw
Machine Learning Meets Higher-Order Network Biology
Weâre excited about a new paper in Physical Review Research, authored by Charo del Genio, introducing a novel spectral method for community detection in hypergraphs âan approach designed to capture higher-order interactions that go beyond traditional pairwise connections.
In biological systems, interactions often occur among groups of genes, proteins, or ligands. Hypergraphs provide a natural way to represent this combinatorial complexity, where Charo del Genio develops the âhypermodularityâ framework to identify functional modules in complex networks.
Implications for drug discovery:
Traditional network models often miss emergent behaviors that arise from collective interactions.
This method enables:
âąIdentification of multi-target modules
âąBetter understanding of co-regulated pathways
âąNew strategies for polypharmacology and target deconvolution
Itâs a meaningful step toward embracing the systems-level complexity of human biology.
Full paper: https://t.co/TmjzuztLmM
#DrugDiscovery #NetworkBiology #KlyneAI #SystemsPharmacology #Hypergraphs #AIinBiotech #ComputationalBiology #Polypharmacology
đŹ AI + Medicinal Chemistry: A Smarter Path to Drug Discovery
Full paper here:
https://t.co/LX41Bbnklk
Excited to spotlight new research from Ron Drorâs lab at Stanford University: MedSAGE, a purpose-built generative AI framework for de novo small-molecule design.
Unlike traditional models that construct molecules atom-by-atom or as SMILES strings, MedSAGE builds molecules from medicinal chemistry-relevant fragments, embedding chemical and 3D geometric data into an interpretable latent space.
â Why it matters:
MedSAGE can perform multi-parameter optimizationâbalancing potency, selectivity, and synthesizabilityâsomething most previous models struggled with. It outputs drug-like, synthetically accessible compounds with high predicted affinity for relevant protein targets.
đ In benchmarks across 25 protein targets, MedSAGE:
â Produced compounds with predicted affinities matching or exceeding known drugs
â Outperformed large-scale virtual screening by 100x efficiency
This research is an important signal: generative AI in drug design is maturing, moving from novelty toward real-world utility.
Experimental validation and data quality remain key challengesâbut MedSAGE shows how embedding medicinal chemistry principles directly into model design can overcome some of these barriers.
đĄ It's not just about generating moleculesâit's about generating viable drug candidates.
đ Read the full paper here:
https://t.co/LX41Bbnklk
Machine Learning Meets Non-Covalent Interactions
We're proud to share that our own @akshat_ai co-authored a major review recently published in @ACSChemRev, exploring how machine learning (ML) is revolutionizing the study of non-covalent interactions (NCIs).
NCIs, like hydrogen bonds, van der Waals forces, and Ï-Ï stacking, are the âinvisible glueâ holding together everything from drug molecules to the proteins they bind. But understanding and predicting these subtle forces is tough, and traditional computational approaches can be slow and expensive.
Machine learning changes the game.
Here's how we leverage these concepts at KlyneAI:
- Accurate Forcefields: ML helps us build next-generation forcefields that capture the nuances of NCIs far better than traditional models. This translates into more reliable, physics-aware simulations.
- Binding Affinity Prediction: By integrating ML models trained on high-quality experimental and quantum chemical data, we can predict protein-ligand binding affinity much faster and more accuratelyâcritical for drug discovery, where every kcal/mol counts.
- Inverse Design & Screening: The ability to âdesign backwardâ (inverse design) by specifying desired binding properties and letting ML generate candidates supercharges our virtual screening pipelines.
Takeaway: Machine learning unlocks truly predictive, data-driven modeling of molecular interactionsâa core capability we use every day at KlyneAI to accelerate drug design.
As the paper highlights, this is an emerging, rapidly evolving fieldâand weâre excited to be pushing its boundaries!
đ Read the full paper:
https://t.co/QM8VzmbzwK
đŹ Curious about how KlyneAI is building cheaper, smarter tools for in silico drug discovery? Letâs connect!
Evaluating AlphaFold3 Predictions of GPCRâLigand Interactions
Paper: https://t.co/xEwjMLI8YU
A new study benchmarking AlphaFold3, the latest AI model from Google DeepMind, reveals both its promise and current limitations for drug discovery.
While AlphaFold3 delivers impressive accuracy in predicting the overall structure of G protein-coupled receptors (GPCRs), a crucial class of drug targets, it often struggles with the precise placement of small molecule ligands within these proteins. This inaccuracy is especially pronounced for novel ligands and complex binding modes, such as allosteric modulators, which are increasingly important in pharmaceutical research.
The analysis shows that AlphaFold3âs predictions are most reliable when similar structures are already present in existing databases, highlighting a challenge in generalizing to new, previously unseen interactions. Additionally, the modelâs internal confidence scores do not consistently reflect the real-world accuracy of its ligand placements, limiting their utility as a measure of prediction quality.
These limitations closely echo what we observed in our recent Boltz-2 case study: strong classification performance, but decreased reliability in regression and nuanced prediction for novel ligands in underexplored chemical space.
AlphaFold3 still requires validation and human expertise to ensure reliability in structure-based drug design and integrating AI predictions with laboratory data remains essential for accelerating and de-risking drug discovery pipelines.
Boltz-2 on Novel Ligands for a Known Target
At KlyneAI, we believe model performance should be evaluated under the toughest conditions - not just retrospective benchmarks.
In this case study, we challenged Boltz-2 with a set of novel ligands against a well-characterized PDB target: a protein with extensive structural data but no close ligand analogs in our training set.
We curated a focused benchmark:
âąÂ 2 confirmed binders (nM-range, same assay)
âą12 confirmed non-binders
The results (see plot) reflect both the promise and limitations of AI in uncharted chemical space.
âą Boltz-2Â correctly flagged the binders (probability > 0.5)
âą But its ICâ â predictions were off by an order of magnitude, and
âą A few non-binders also fell into the âhigh binding probabilityâ range.
This highlights the need for multiple outputs - Affinity Probability, predicted ICâ â, and beyond to form a more nuanced view of performance. Relying on a single score, especially in novel regions, can be misleading.
We're continuing to stress-test our models in realistic discovery settingsâand iterating rapidly.
#AI4DrugDiscovery #DeepLearning #MolecularDesign #Biotech #KlyneAI
đMoonshot spotlight: AI-powered age reversal is moving from science fiction to reality - and we're proud to be part of the breakthrough.
@davidasinclair recently shared an incredible update on his longevity research in his latest interview with @PeterDiamandis: "Imagine in 10 years you you just take a pill for 4 weeks and you get younger"
What's remarkable? This isn't distant speculation. "We're doing experiments in a matter of a month that would take hundreds of thousands of years to to do... trillions of molecules coming through screening virtually"
At https://t.co/AkKsKo88g8, we're thrilled to have contributed to this moonshot by delivering exactly what we do best: making high-performance virtual screening accessible and cost-effective.
Our collaboration with Dr. Sinclair's lab showcases the power of AI in drug discovery:
- Completed 7.2 billion enzyme docking calculations across multiple targets to identify age-reversal molecules
- Klyne's platform delivered a 90% efficiency gain over conventional methods
- Enabled massive molecular screening at a fraction of typical costs - Accelerated the timeline from gene therapy (costing $300K-$2M) to potential oral pills at ~$100/month
The breakthrough? "Using AI we're now at a point where we've got molecules that really would only cost $100 to make or less... for a month's course" - transforming expensive gene therapies into accessible treatments.
This is exactly why we built Klyne: to be as high performance and cost-efficient as possible for virtual screening. Whether it's age reversal research like Dr. Sinclair's breakthrough, or other drug discovery projects, we're here to make enterprise-level capabilities available to labs that previously couldn't access them.
From "$10 million just to make the first batch" to potentially "pennies on the dollar every time someone takes a pill" - this is the kind of transformation AI enables in drug discovery.
The future? Dr. Sinclair predicts we're moving from simply slowing aging to "the ability to truly reset the body reset all of the cells in the body to be young again" At Klyne, we're not just observers of this revolution - we're helping make it happen, one virtual screen at a time.
What breakthrough projects are you working on? How could AI-powered drug discovery accelerate your research?
Listen to the full interview: https://t.co/tDsh16JcBV
#LongevityScience #AI #DrugDiscovery #AgingResearch #VirtualScreening #Innovation
đ§Ź New Preprint Spotlight đ§Ź
A simple AI ensemble that delivers powerful results for antibody design.
https://t.co/hcU5Cv1XLN
We get really excited when we see work around generative design for protein sequences - something we do here at @KlyneAI, along with designing small molecules with our AI tool.
In this new preprint, the team at the Institute for Protein Design (UW) paired ProteinMPNN with AbLang - unlocking major improvements in antibody CDR design.
By combining ProteinMPNN with AbLang, this team:
- Outperformed ProteinMPNN alone in silico
- Generated >10x more HER2-binding trastuzumab variants
- Produced CDR sequences that are far more natural and functionally relevant
Whatâs exciting? No retraining needed, just smart model integration.
This work highlights how domain-specific AI models can meaningfully advance therapeutic antibody discovery.
New research alert: A team from the University of Geneva used MD simulations to uncover a cryptic pocket in KRAS Q61H â totally invisible in crystal structures. Exciting implications for those relying on static models in drug discovery.
At KlyneAI, we're actively exploring how MD-based reranking can surface better hits, especially in hard targets like this!
đCurious to hear how others are tackling cryptic site discovery - whatâs been working for you?
Link to paper: https://t.co/WHHWfI2sAo
đ Exciting to see our collaboration with Cerebrum DAO featured on their latest podcast with our COO Luis Rios and @bmagierski!
At Klyne, we were thrilled to support this initiative with our short simulation technique, enabling fast and accurate binding affinity calculations across huge compound libraries.
In a field where timelines and costs matter, this approach is proving to be a game-changer for cost-effective drug discovery. Proud to be accelerating progress on neuroprotective agents in neurodegeneration. đ§ đ
đ Check out the full podcast: https://t.co/USsjM0t1o8
It was our pleasure to speak to Dr. Luis Rios on our podcast "Meet the Scientist" together with @bmagierski talking about @FissionBioDeSci and the work that we are doing together to tackle neurodegeneration with mito-protective agents.
We are also collaborating with @KlyneAI, who did an AI-enabled ultra-large-scale virtual screen for Fission, helping us reach our goals.
Watch the full interview here: https://t.co/oC9bzDB8K4
Check out anti-aging mystro @bryan_johnson highlighting a paper co-authored by our CTO @akshat_ai đ
Generative tools are starting to deliver real-world drug leads â and this is just the beginning.
The future of drug discovery is getting faster, smarter, and more accessible.
More cures for all. Letâs go!!! đâš
This is cool. The first real-world demonstration of a quantum-enhanced AI workflow producing experimentally validated drug leads for a notoriously "undruggable" cancer protein.
Offering a template for integrating quantum computing as a molecule generator in a hybrid model with classical AI models for groundbreaking drug discovery.
What they did: a research team from @InSilicoMedsand and @UofT leveraged a potent hybrid quantum-classical AI system to design novel small molecules capable of inhibiting mutant KRAS, a notoriously "undruggable" cancer protein.
KRAS is one of the most frequently mutated oncogenes across human cancers, particularly prevalent in lung, colorectal, and pancreatic tumors. These mutations hyperactivate the KRAS protein, leading to uncontrolled cell growth. For decades, KRAS was deemed "undruggable" due to its smooth surface and lack of deep binding pockets, making it incredibly challenging for small molecules to bind effectively and inhibit its function.
Given that KRAS mutations are responsible for approximately 25% of human cancers, often associated with poor prognoses and limited efficacy of existing drugs, it presents an ideal target for AI-driven drug discovery.
Congratulations to @biogerontology for this breakthrough.
Here are more details for those of you curious about AI drug discovery
1/ Quantum-Classical AI setup and workflow. Data and training.
A custom training set of 1.1 million molecules, including known KRAS binders, was compiled and used to train both the classical and quantum systems using Virtual Flow docking screening.
+ Generation of molecules: Hybrid Generative model
Quantum AI: Generator, Classical AI: discriminator.
A Quantum Circuit Born Machine (QCBM) ran on a quantum chip to generate new molecular structures.
A classical AI system (Insilicoâs Chemistry42) evaluated those molecules, filtering for likely KRAS interaction
+ Filtering and identifying best drug candidates:
Using Chemistry42 a million QCBM-generated molecules were sampled, filtered according to their drug-likeness including absorption metabolism and toxicity, and then ranked based on their docking scores.
+ Chemical synthesis and real world validation
15 candidates were chosen, chemically synthesized, and tested in the labâŻ.
2/ The main outcome
Two of the 15 molecules successfully inhibited multiple KRAS variants in live cells. Real experimental hits based on quantum AI generation (QCBM) coupled with classical AI (Chemistry42) screening and validation.
3/ Advantage of using Quantum Computing for drug discovery
Quantum systems are theoretically better at sampling from complex, high-dimensional distributions, which could help find unusual and novel molecules that classical systems might miss. But today, this is still experimental. No clear advantage has been proven yet.
4/ This study used a âreal but limitedâ quantum computer
The study used real, hardware-based quantum processor (IBM Quantumâs superconducting qubits), not just a quantum simulator.
However, the scale and depth of the quantum computation were limited, and the scope of application was limited to the molecule generation step, while all the evaluation and optimization steps were performed by Chemistry42, a classical generative AI model.
5/ Significance
+ First experimental validation of quantum-AI drug generation
This is the first study to use a quantum generative model (QCBM) integrated with classical AI to design small molecules that were synthesized and shown to inhibit a real, disease-relevant target (KRAS) in vitro.
+ Demonstrates hybrid quantumâclassical synergy
The study showcases a practical workflow, where quantum computing contributes to early-stage ideation, while classical AI tools like Chemistry42 handle evaluation and optimization.
+ Progress against the âundruggableâ KRAS target
KRAS has long been considered one of the most challenging oncogenic proteins to target. Generating novel KRAS inhibitors using this approach highlights the potential of AIâquantum tools to tackle previously intractable targets.
+ Opens the door to expanding the chemical space
Quantum models can sample from non-intuitive, high-dimensional chemical spaces, potentially leading to new classes of molecules that classical methods might overlook.
+ Lays groundwork for scalable quantum drug discovery
While quantum advantage has not yet been demonstrated, this study is a foundational step showing that quantum hardware can be functionally integrated into real-world discovery pipelines.
6/ Limitations
+ Limited scale and maturity of quantum hardware
The quantum component (QCBM) was run on noisy, small-scale quantum hardware, limiting circuit depth and complexity. The study does not demonstrate quantum advantage over classical generative methods.
+ Narrow experimental validation
Only 15 molecules were synthesized, and just 2 showed modest in vitro activity against KRAS. There was no in vivo testing, pharmacokinetics, or toxicity data â so therapeutic potential remains speculative.
+ Single-target, proof-of-concept scope
The approach was tested on only one protein family (KRAS). Itâs unclear how well this quantumâclassical pipeline generalizes to other targets.
WoW! Achievement unlocked -- @bryan_johnson just tweeted about our work! Glad you liked the paper!!!
Check out @KlyneAI we are trying to create accessible tools like these for any/everyone. FASTER DRUG DISCOVERY FOR ALL!! đ
Check out our latest review on modeling non-covalent interactions â a fundamental aspect of how we understand and predict proteinâligand binding at Klyne.
This work underpins much of our ability to design better molecules through accurate interaction modeling.
đ Read the full review here:
https://t.co/AwYZl5b4Ap đ đ
đš New paper out in Chemical Reviews!
"Studying Noncovalent Interactions in Molecular Systems with Machine Learning"
đhttps://t.co/qwQ6uSFXAC
Thrilled to have worked with Serhii Tretiakov and the brilliant team in @robpolliceâs group on this!
Also proud to share this as the first publication featuring @KlyneAI â more exciting work on the way!
If you live for 5 yrs, you may live for another 200 yrs.
AI is at work. Thanks to @KlyneAI
An astonishing 7.2 billion enzyme docking calculations across multiple targets to identify age-reversal molecules using AI. This is the future of medicine.
Definitely listen to our CTO đĄ â we offer cost-effective solutions for tough biotech challenges. Working on early-stage drug discovery? We specialize in hit identification and hit-to-lead optimization. Reach out: [email protected] đŻ
We're thrilled to be working with @davidasinclair! At @KlyneAI, we're focused on tackling some of the toughest, most impactful diseases out there. Got a biotech problem you're working on? Let's talk â [email protected] đ