This episode delves into how Nova (Subnet 68 on Bittensor) leverages decentralized AI and crypto-economic incentives to revolutionize the costly and slow process of drug discovery, turning miner competition into a powerful engine for identifying novel therapeutics and uncovering weaknesses in state-of-the-art predictive models.
Introduction to Nova & Metanova Labs
- Amanda, Pedro, and Miko from Metanova Labs introduce Nova (Subnet 68), a Bittensor subnet dedicated to accelerating drug discovery.
- Amanda highlights the convergence of AI, crypto, and pharma—three sectors with immense value creation potential. She emphasizes that merging them addresses individual limitations, positioning Nova not just as a tech venture but as a mission to accelerate medical innovation with profound real-world impact beyond token value.
The Crisis in Drug Development
- The current state of drug development is framed as a crisis. Amanda points out the staggering cost ($2.6 billion) and time (10 years) required to bring a single drug to market, coupled with a low (~10%) success rate.
- This high-risk, high-reward environment often leads to incrementalism (minor tweaks to existing drugs) rather than pursuing true novelty, hindering breakthrough discoveries.
AI's Potential and Current Limitations
- While AI has been explored for drug discovery since the 1960s, only recent advances in data processing have made it truly impactful.
- Amanda notes AI's potential to make drug discovery "better, faster, cheaper, more novel" by exploring uncharted regions of the "chemical universe."
- However, she argues AI hasn't fully delivered because development often occurs in isolated academic or private settings ("in a vacuum"), lacking the diverse, real-world data needed for robust predictive power.
- Incentives are misaligned, prioritizing proprietary models over collaborative progress.
Bittensor's Role: The Valley of Death & Nova's Approach
- The "valley of death" in early drug development, where candidates fail due to economic and technological hurdles rather than scientific ones, is presented as a key challenge Nova addresses.
- Amanda posits that Bittensor, through token emissions (via the Decentralized Token Offering - DTO), provides crucial early-stage funding.
- This allows Nova to pursue bolder, "moonshot" research compared to traditional routes.
- The decentralized, global talent pool and merit-based competition inherent in Bittensor are highlighted as key advantages, reducing fundraising burdens and fostering rapid optimization.
- Amanda states, "it really reduces the burden of having to raise all of that money to get into this field."
- Strategic Insight: Bittensor's model offers a potential alternative funding and R&D pathway for capital-intensive biotech research, bypassing traditional VC constraints and enabling riskier, potentially higher-impact projects.
Nova's Mechanics: How it Works
- Pedro explains Nova's operational mechanics. Miners utilize a massive dataset from Savvy (over a billion synthesizable molecules – a key practical constraint) to predict binding affinity between molecules and specific protein targets/anti-targets.
- Binding Affinity: A measure of how strongly a molecule interacts with or binds to a protein target. High affinity is often desired for a drug's primary target, while low affinity is sought for anti-targets (to avoid side effects).
- Validators then use a state-of-the-art model (currently "Psychic") to score these predictions and determine winners for each challenge. The outputs generated—chemical maps, molecule libraries, model fine-tuning data—feed back into the system, strengthening the subnet and informing Metanova Labs' own R&D and potential partnerships.
Early Progress and Incentive Mechanism Evolution
- Since launching on March 1st, Nova has run ~780 challenges, exploring 4,000 proteins and generating 17,000 unique molecule predictions.
- Pedro details the rapid iteration of their incentive mechanism, driven by miners efficiently finding "paths of least resistance." This evolution included:
- Moving from single targets to target/anti-target pairs, then to one target vs. multiple anti-targets.
- Implementing encrypted submissions to prevent copying.
- Expanding the protein dataset significantly (to over 2 million) to increase unpredictability.
- Requiring a minimum number of heavy atoms (atoms other than hydrogen, contributing more significantly to molecular weight and structure).
- Continuously tuning scoring parameters.
- Actionable Insight: The rapid evolution highlights the adversarial nature of decentralized networks. Investors should assess a subnet's ability to adapt its incentives quickly to maintain alignment with its core scientific or business goals against miner optimization strategies.
Miner Exploits & Model Weaknesses
- Pedro provides specific examples of how miners uncovered weaknesses in the Psychic prediction model, demonstrating the power of the decentralized network in stress-testing AI.
- Low Atom Count Issue: Miners initially found that molecules with very low atom counts often scored highly, representing a "low confidence" region or potential "hole" in the Psychic model's predictive capabilities, correlating with areas of high prediction variance. This wasn't reported in the original Psychic paper.
- Rigidity Issue (Rotatable Bonds): Even after addressing atom counts and implementing anti-targets, miners discovered another strategy: submitting molecules with high heavy atom counts but very few rotatable bonds (bonds allowing parts of a molecule to rotate freely, impacting flexibility and interaction potential). These rigid molecules, while underrepresented in the dataset, could win challenges without necessarily being good drug candidates, revealing another potential blind spot in the model, particularly concerning selectivity predictions.
- Key Takeaway: Miner behavior, often seen as adversarial, is repurposed here as a feature. It provides invaluable, real-time feedback on model limitations in under-explored chemical spaces, data often unsought or unpublished by model creators focused on positive results.
The Shannon Upgrade: Addressing Exploits & Promoting Diversity
- The upcoming "Shannon Upgrade" (scheduled for Friday, April 11th) directly responds to these findings. Pedro explains it incorporates:
- A fixed, weekly rotating target (for building specific chemical libraries) alongside randomly chosen anti-targets (proxies for off-target effects/toxicity).
- Miners submitting sets of molecules, not just one.
- Penalties for molecule repetition to discourage focusing on narrow chemical scaffolds.
- A diversity bonus calculated using Shannon entropy (a measure of unpredictability or diversity within a set of data, here applied to molecular representations) to incentivize broader exploration of the chemical space.
- Strategic Implication: This upgrade aims to harness miner efficiency not just for finding edge cases, but for systematically mapping the chemical space and identifying regions of high prediction variance (uncertainty), crucial for real-world drug application.
Mapping the Chemical Universe & The "Treat" Initiative
- Pedro notes that recent real-world data confirms the hypothesis that computationally mapping vast chemical spaces improves drug discovery success rates. Nova aims to create such maps, starting with the "Treat" (Targeted Reward Evaluation and Therapeutics) initiative.
- Amanda explains this focuses on targets related to reward and learning systems (affecting focus, eating, sleep, trauma recovery).
- The goal is to develop "metaprogramming drugs"—drugs enabling people to train themselves—linking to behavioral regulation's role in health and longevity. While sounding novel, this maps to traditional indications like ADHD or sexual dysfunction (e.g., Adderall, Viagra), aiming for better solutions.
Future Steps & Vision
- Amanda outlines Nova's roadmap:
- Launching the Shannon Upgrade and Treat challenge.
- Releasing new prediction models (benchmarked, retrospective data).
- Integrating more property predictions crucial for drug development, such as ADME/Tox (Absorption, Distribution, Metabolism, Excretion, Toxicity) and BBB (Blood-Brain Barrier permeability – how well a drug crosses from the bloodstream into the brain). Pedro specifically mentions their work on a BBB model relevant to the Treat initiative.
- Creating high-value chemical libraries.
- Real-world validation: Synthesizing promising molecules and conducting wet lab assays to bridge the digital-to-real-world gap. This validation is key to building trust and value.
- Advancing three distinct partnerships (Research University, Private Company, CRO - Contract Research Organization) to connect crypto and traditional pharma.
- Investor Note: The progression towards real-world validation and integration of complex properties like ADME/Tox and BBB permeability are critical milestones for assessing Nova's long-term viability and potential impact.
Deep Dive: The Validation System
- Miko probes the specifics of the validation system. Amanda and Pedro reiterate that miners find molecules maximizing predicted binding affinity for a target while minimizing it for anti-targets, using the Psychic model which predicts interaction favorability based on molecule (SMILES representation) and protein (amino acid sequence).
- Pedro emphasizes the importance of selectivity: finding molecules that hit the desired target protein but avoid off-targets linked to side effects. This is a major reason drugs fail in clinical trials.
The Importance of Selectivity & Scale
- The conversation underscores that drug discovery involves searching a vast chemical space (billions or trillions of molecules) for selective compounds. Pedro explains that finding diverse chemical starting points ("entry points") that meet selectivity criteria increases the chances of developing a successful drug.
- Miko highlights the computational scale: finding a molecule that, for instance, activates specific brain receptors (like GABA) without hitting others requires immense screening.
Speeding Up Drug Discovery
- Pedro confirms that Nova aims to drastically accelerate this screening process. Traditional in silico (computational) screening for billion-molecule datasets can take weeks or months.
- Speeding this up significantly, potentially 10x or more as suggested by Miko, allows for faster iteration cycles in drug development.
- Amanda adds that optimizing this early "top of the funnel" stage makes the entire downstream process more efficient.
The Value Proposition for Pharma
- The discussion touches upon the immense cost savings potential for pharmaceutical companies. By identifying promising candidates and filtering out likely failures much earlier computationally, Nova can reduce the ~$2.6 billion average cost, largely driven by late-stage failures.
- Amanda contrasts Nova's open approach with the proprietary models of companies raising hundreds of millions, arguing that Bittensor's incentive for data sharing can lead to broader exploration and more robust discoveries, benefiting the entire industry.
Miner Behavior Analysis (Exploits as Features)
- Miko asks whether miners are simply scaling compute or employing intelligent optimizations. Pedro suggests a mix: large-scale predictions combined with deep exploratory data analysis to find patterns and model weaknesses ("holes"). Amanda reiterates that miner behavior, initially seen as problematic, is now viewed as a feature for uncovering model uncertainties.
- The Shannon upgrade aims to channel this efficiency towards generating diverse, high-value outputs rather than repetitive exploitation of specific model weaknesses.
The Flywheel Effect: Improving Models with Miner Data
- Miko inquires about a feedback loop for model improvement. Pedro confirms this is planned. The data generated, particularly highlighting regions of high variance/low confidence (often in underrepresented chemical spaces), can be used for:
- Fine-tuning Psychic (or other models) using known data not in the original training set.
- Generating new lab results for promising candidates and feeding that real-world data back into model training.
- This iterative process aims to create more accurate models, especially in "drug-like" chemical space regions.
Composability & Future Model Integration
- Pedro confirms the subnet's modular design allows for incorporating additional models beyond Psychic. This could include fine-tuned versions for specific protein families (enzymes, receptors) or models predicting other crucial properties like BBB permeability or broader ADME/Tox profiles.
- Amanda emphasizes that this modularity allows Nova to optimize any underfunded open-source model, potentially creating a new wave of robust tools "trained on Bittensor."
Nova's Long-Term Vision & Monetization Strategy
- Amanda describes a multi-track approach: advancing Nova's own drug R&D while making the subnet a powerful tool for others. Monetization strategies evolve with validation:
- Early Stage: Partnerships (service model), leveraging Pedro's network.
- Mid Stage: Token-gated access to curated, validated chemical libraries.
- Long Term: Developing proprietary drug candidates based on the platform's discoveries.
- The core idea is to monetize validated outputs while building bridges between crypto and pharma.
Addressing Pharma Industry Problems
- Amanda argues Nova tackles core pharma issues: excessive time and cost, lack of novelty due to risk aversion (VCs demanding late-stage data), and herd mentality in research focus.
- By enabling earlier, cheaper, and potentially more novel discovery, Nova aims to realign R&D with actual needs.
- Pedro adds that even with open data, significant post-processing and analysis are required, allowing for value creation on top of the open-source outputs for specific partners.
Partnership Details (Characterization)
- Responding to a chat question, Amanda characterizes the three partnerships without naming them due to confidentiality: one with a research university, one with a private company, and one with a CRO (Contract Research Organization) lab.
- These diverse partnerships are crucial for bridging the crypto-pharma divide.
Conclusion
- Nova demonstrates decentralized AI's unique power to stress-test and refine drug discovery models by leveraging miner competition.
- Investors/researchers should track Nova's translation of these insights into validated drug candidates and valuable chemical libraries, representing a novel intersection of crypto incentives and biotech R&D with significant real-world potential.
