Bittensor Subnet 68: Nova

The subnet’s architecture consists of three primary components​:

Miners – These are nodes that generate and propose candidate molecules. Each miner runs specialized search algorithms over a vast chemical library (the SAVI-2020 database of ~1.75 billion synthesizable compounds​) to find molecules predicted to bind strongly to the current target protein. Miners can use machine learning models (e.g. QSAR or deep learning) and heuristic methods (like active learning or substructure-based search) to intelligently navigate chemical space​. Notably, miners continuously refine their search using feedback from the network’s oracle model (described below), focusing on higher-scoring regions of chemical space rather than brute-forcing all possibilities​. Each miner can only hold one “active” molecule submission at a time, and is incentivized to update it only when they discover a better candidate​.

Validators – These are nodes that evaluate the quality of miners’ submissions using a predefined incentive mechanism. In NOVA, the validators all run a Deterministic Oracle model (called PSICHIC) to score each submitted molecule’s binding affinity to the target protein​. Essentially, validators act as judges: at the end of every block (each block ≈12 seconds in Bittensor), they compute a binding affinity score for each miner’s current molecule using the oracle and collectively rank the miners​. Because PSICHIC is a fixed, open-source model that gives deterministic predictions for a given protein–molecule pair, all validators will obtain the same score for the same input, ensuring objective and transparent evaluation across the network​. This removes subjectivity – the oracle’s score serves as a ground-truth standard that validators enforce uniformly. The top-scoring submission is identified each block, and validators submit those results to the blockchain (via Bittensor’s consensus mechanism) to determine reward distribution​.

Deterministic Oracle (PSICHIC) – PSICHIC is the pretrained predictive model at the heart of NOVA’s evaluation process. It is a state-of-the-art protein–ligand binding affinity model that takes a protein’s sequence (or structure) and a molecule (e.g. SMILES representation) and outputs a binding affinity score​. Because PSICHIC’s outputs are reproducible and deterministic (given a fixed model version and input)​, it provides a reliable yardstick for comparing candidates. In NOVA’s “competition,” PSICHIC essentially plays the role of an automated in silico screening assay – it provides the scoring function (analogous to a loss function in ML terms) that all miners are trying to maximize​. By anchoring the incentive mechanism to an open and fixed oracle, NOVA ensures a level playing field: the only goal is to find molecules that score higher by PSICHIC’s metric, and everyone knows the metric in advance.

Network process: NOVA operates in iterative challenge rounds targeting specific drug discovery problems. In NOVA v1, each challenge is defined by a single protein target – miners must find a molecule from the database with the highest predicted affinity to that target​. All miners start searching as soon as a new challenge (target) is announced, and they continuously submit molecules over the round (which spans roughly 360 blocks, ≈1 hour in the current design​). Importantly, submissions are updated on the fly: if a miner discovers a better candidate, it replaces their previous one (ensuring that at any time, each miner only occupies one “slot” in the competition)​. Every block, validators evaluate all current submissions using PSICHIC and determine that block’s top scorer​. This creates a fast-paced, block-by-block tournament: “a high-speed race for breakthroughs.” The miner whose molecule yields the highest affinity score (if higher than all peers that block) is declared the winner of that block’s sub-challenge​. Ties are broken by whoever submitted the winning molecule first​.

All evaluated scores and winning molecules are recorded, and a leaderboard of results is made visible to the community in real time​. This transparency lets all participants learn which molecular strategies are working best, encouraging adaptive competition – miners can observe what kinds of molecules are scoring well and adjust their search accordingly in subsequent submissions​. At the end of the round (e.g. after 360 blocks), the challenge concludes and a new protein target can be selected for the next cycle​. Over time, NOVA plans to increase the complexity of these challenges – moving from single-target affinity optimization to multi-objective searches that consider multiple protein targets or additional drug-like properties (such as toxicity, metabolic stability, etc.)​. This will make the subnet’s task more closely resemble real drug discovery, which requires balancing many factors for a viable drug candidate.

Role within the Bittensor Ecosystem

NOVA is one of many specialized subnets in the broader Bittensor ecosystem. Each subnet in Bittensor is essentially an independent “market” for a particular AI service or commodity, running on the shared Bittensor blockchain (often called Subtensor)​. While some subnets focus on language modeling, data curation, or other AI tasks, Subnet 68 (NOVA) is uniquely dedicated to decentralized drug discovery​. For context, other notable subnets include SN9 (Pretrain) for large-scale language model pre-training, SN13 (Data Universe) for data collection, and SN25 (Folding) for protein folding simulations​. NOVA complements this ecosystem by tackling the domain of medicinal chemistry and DeSci (decentralized science) — it brings a biotech use-case to Bittensor’s AI network​.

Because all subnets share the TAO token and core blockchain, participants can seamlessly move between subnets or even contribute to multiple ones in parallel. In NOVA’s case, this opens the door for cross-subnet collaboration. For example, in the future NOVA could integrate insights from other subnets: a language-model subnet might help generate novel molecular structures, a protein-folding subnet (like SN25) could provide protein 3D structures to improve scoring, or a data subnet could supply biochemical data to refine models. The NOVA whitepaper explicitly notes the potential synergy of NOVA with subnets like SN9, SN13, and SN25, creating a “synergistic ecosystem” of AI services that feed into each other​. This modular design is a strength of Bittensor – each subnet can specialize but also benefit from others’ outputs, all under a unified economic system.

Operationally, NOVA uses Bittensor’s infrastructure for security and consensus. Bittensor employs a consensus algorithm (Yuma) whereby validators’ reported performance scores determine how new tokens (“emissions”) are distributed to miners​. Each subnet creator (in this case, Metanova Labs for NOVA) defines the incentive mechanism that validators must apply​. NOVA’s incentive mechanism is the PSICHIC affinity score – validators simply use that to grade miners​. Thus, NOVA slots into Bittensor’s framework: miners and validators join by paying a TAO registration fee to get a seat (UID) in Subnet 68​, and then run NOVA’s open-source code (miners run search algorithms, validators run the oracle scoring code). The subnet can accommodate up to 256 active participants (max 192 miners + 64 validators, per Bittensor’s design limits)​, ensuring a healthy amount of decentralization while preventing the challenge from being flooded with low-quality contributions. Bittensor’s base layer handles block production, token transfers, staking, and registration, so NOVA’s creators could focus on the domain-specific logic (drug screening and reward rules). In summary, NOVA operates as a plug-in module on Bittensor – it leverages the network’s global decentralized compute power and token incentives to tackle a very specialized AI problem (virtual drug discovery) that would be impossible to crowdsource at this scale without blockchain coordination.

Unique Features and Innovations

NOVA introduces several innovative elements that set it apart from traditional drug discovery approaches and even other AI networks:

Decentralized & Democratized R&D: NOVA is arguably the first open, permissionless drug discovery engine. Anyone with the requisite expertise and compute can become a drug-hunting miner or a validator, regardless of institutional affiliation. This model crowdsources ideas and approaches from around the world, breaking the siloed nature of pharma R&D​. It also means progress in NOVA is transparent – every winning molecule and score is visible on-chain, creating an open repository of potential drug candidates over time​. This transparency and community access to results is a departure from the secretive, proprietary research model.

Gamified “Search Economy”: By structuring drug discovery as a competitive game with frequent rewards, NOVA incentivizes rapid innovation. The “high-speed race” format (12-second rounds, continuous leaderboard updates) turns what is normally a painstaking lab process into something dynamic and engaging. Researchers are encouraged to try bold or diverse strategies since only the top result matters per round – exploration of chemical space is rewarded. This game-theoretic approach, powered by crypto incentives, is novel in scientific research. It aligns individual profit-seeking with a collective goal (finding better drugs)​.

Objective Oracle & Fair Evaluation: The use of a fixed, state-of-the-art ML model (PSICHIC) as a deterministic oracle is a unique innovation. It provides an objective, reproducible benchmark for all contributions​. All miners effectively optimize the same known function, and all validators enforce that function. This avoids human bias or subjective judging of results – something quite important in science competitions. It’s akin to a leaderboard challenge in Kaggle or other ML competitions, but running continuously on-chain with automatic scoring. The choice of PSICHIC (which is high-performing in binding affinity prediction​) ensures the evaluations are scientifically meaningful. Over time, NOVA can upgrade or fine-tune the oracle (with community governance) as better models become available, always maintaining a single source of truth for scoring.

Massive Search Space & Dataset: NOVA leverages the SAVI-2020 library (1.75 billion synthesizable molecules) as its search domain​. This is orders of magnitude larger than what any single lab could screen experimentally. By encoding this huge chemical space into a distributed search problem, NOVA pushes the boundary of what virtual screening can do. Moreover, every molecule in SAVI-2020 comes with known synthetic routes (it’s “make-able” chemistry)​. This means any top hits NOVA finds are not just theoretical – they can be synthesized in real life, greatly increasing the impact of the computational results. The focus on synthesizability and drug-likeness (SAVI compounds are built from known reactions) gives NOVA an edge in producing drug candidates that are realistic, not just random molecules​. This is a deliberate innovation to bridge computation and wet-lab applicability.

Integration of DeSci and Web3: NOVA is at the intersection of decentralized science and blockchain. By building on Bittensor, it integrates with a wider Web3 ecosystem – for instance, results could be NFT-encoded or used in decentralized biotech funding platforms down the line. The project has garnered interest from both the AI community and the DeSci community (followers include prominent DeSci figures) because it demonstrates a working model of crypto-driven scientific research. The entire pipeline from data (open dataset on HuggingFace​
X.COM) to model (PSICHIC on HuggingFace​) to code (open-source on GitHub​) to results (blockchain explorer) is open and transparent. This kind of end-to-end openness is an innovative paradigm for drug discovery, which historically has been proprietary.

Future Real-World Impact: A key planned innovation is to close the loop with real experiments. NOVA’s roadmap includes taking the highest-scoring molecules from the network (the “virtual hits”) and synthesizing/testing them in actual laboratories (wet lab validation)​. The results of those biological assays would then be fed back into the network – for example, to retrain or recalibrate the oracle model (making PSICHIC more accurate in future rounds)​. This creates a powerful virtuous cycle between decentralized computation and centralized experimentation. If successful, NOVA could not only propose drug candidates but also verify them, leading to a pipeline of potentially patentable or development-ready compounds emerging from a decentralized community. Such an integration of on-chain and off-chain research is highly unique. Additionally, NOVA envisions tokenizing libraries of top molecules – essentially creating crypto assets representing promising compounds, which could be traded or funded for further development​. This would bring liquidity and market dynamics to early-stage drug assets, a radical innovation in how drug discovery projects are valued and advanced.