Gaming vs. saving lives: the $2.5M drug discovery problem

The shape of the problem

Compute gets cheaper and faster every year — Moore's Law. Drug discovery gets slower and more expensive every year — the pharma industry calls this Eroom's Law, Moore's Law spelled backwards, because it runs in the opposite direction. Innophore, an Australian biotech company, sat squarely inside that gap: they needed to identify a specific enzyme, out of roughly 200 million known protein sequences, capable of catalyzing one particular drug reaction. Searched by hand in a wet lab, that's a multi-generational project. They needed an answer in weeks.

What the system does

Three models, three jobs: narrow 200 million candidates down with embeddings, predict what the survivors actually look like in 3D, then simulate whether they'd really bind the target molecule.

Treating proteins like language

The underlying trick is recognizing that biology already comes pre-tokenized: amino acids function like letters, chained together into sequences that fold into 3D structures the way sentences carry meaning. ESM-2 — described in the source as "the BERT of proteins" — is a transformer trained on 200M sequences from the UniRef database, and it turns each one into a high-dimensional embedding that captures functional properties, not just sequence similarity.

200,000,000 candidate sequences
        │  ESM-2 embedding
        ▼
   embedding space, nearest-neighbour search
        │  filter to functional-similarity threshold
        ▼
   thousands of promising candidates (hours, not years)

Filtering in embedding space is what compresses "200 years of wet-lab search" into hours — the model doesn't need to reason about each sequence individually, it just needs the ones that land near the right region of the space it already learned.

Structure prediction at scale

Sequence similarity alone isn't enough — a candidate has to actually fold into a shape that fits the reaction. AlphaFold2 and OpenFold predict 3D atomic coordinates from the 1D sequence using attention over multiple sequence alignments, and NVIDIA's optimizations (MMseqs2 for alignment, running on H100s) made that step 5x faster.

In a 2024 dataset-construction pass, Innophore and NVIDIA ran three separate folding models across 42,042 distinct human proteins from UniProt's reference proteome, then trimmed anything with a confidence score (pLDDT) below 70 or floating chains more than 10Å from the protein core:

Running three models instead of one isn't redundancy for its own sake — it gives structural diversity for training, alternative binding sites the single-model prediction might miss, and a consensus check across independent predictions.

Validating with diffusion, not brute force

The naive way to check whether a candidate enzyme actually binds the target molecule is to brute-force rotate and translate it in simulated space until something fits. DiffDock replaces that with a diffusion generative model that iteratively refines a random starting position into a plausible binding pose directly, then ranks candidates by predicted binding affinity.

Quality gates, not blind trust in the model

Every structure prediction ships with a confidence score, and the pipeline is built to respect it rather than average over it. pLDDT (per-residue confidence) runs from below 70, meaning "use cautiously," up to 90+, meaning high confidence — and low-confidence regions get trimmed before anything downstream relies on them. Structure quality overall is benchmarked against experimental PDB structures using LDDT/TM-Score on the CAMEO and CASP14 benchmark sets, and even the language model itself is validated on perplexity (10.45 at 8M parameters vs. 6.37 at 15B — bigger isn't automatically better, it's measured).

What this actually costs

The source article puts infrastructure ROI at roughly 9,000x, with compute costs breaking even the same day relative to a single wet-lab search cycle.

The pandemic stress test

The clearest real-world proof point predates this specific enzyme search: during COVID-19, Innophore used the same approach to scan 15 million SARS-CoV-2 variants, cutting compute time from roughly 2,000 hours to 15 minutes — fast enough to predict Omicron's binding properties before the variant became dominant.

The result

Two weeks, under $100,000, against a wet-lab baseline north of $50M and 12+ months. The enzyme they found has a sequence completely different from existing patented proteins, but is structurally identical to what the reaction required — the kind of candidate that's very hard to find by searching for sequence similarity alone, and exactly what an embedding search over structure-aware representations is good at surfacing.

As the source piece puts it: AI didn't replace the biologists here. It let them test 200 million hypotheses in two weeks instead of rejecting most of them without ever examining them.