AI in Drug Discovery

The question

If computational methods lift the conditional probability of a preclinical asset reaching Phase II by 300-500 basis points, which incumbents capture the resulting expected-value lift, and are they priced for it?

Base rates

The reference class is pharmaceutical productivity regimes. Industry preclinical-to-approval success rates have declined from roughly 10% in the 1990s to roughly 5-7% by the late 2010s, with cost-per-approved-drug roughly doubling per decade (Eroom's Law). The base rate for a technology claim reversing Eroom's Law is low -- high-throughput screening, combinatorial chemistry, and genomics-era target identification all produced smaller effect sizes than promoters forecast. Roughly the 20th percentile for technology-driven productivity claims in pharma.

However, the reference class for structure-prediction and generative-chemistry tools is narrower and more encouraging: in published retrospectives, campaigns using structure-based design plus ML-guided candidate ranking show hit-rate improvements at the lead-optimization stage of roughly 2-4x versus historical controls. Whether that translates to approval-rate lifts is the open question, and the answer will take 4-7 years of clinical read-outs to settle.

The imputed probability in consensus biopharma numbers is that computational methods add roughly zero to pipeline expected value. That is a reasonable prior given Eroom. It is likely wrong at the margin, and the margin is where the money is.

Why the consensus view is wrong (or incomplete)

The bear case says drug discovery is constrained by biology, not by chemistry, and better molecule selection does not help if the target is wrong. That is partially correct. The causal mechanism we think consensus misses is that computational methods change the economics of failing earlier. If screening moves 20% of failures from Phase I to preclinical, per-program cost falls by roughly 30-40% even if overall approval rates are unchanged. That flows directly to pipeline NPV without requiring any claim about biology.

The second mechanism is platform leverage at the tooling layer. Sequencing read volume and the associated assay consumables are inputs to every computational campaign; tools vendors capture a royalty on industry-wide activity regardless of which therapeutic bet wins.

Position construction

Three 20% anchors: GILD, ALNY, and REGN. All three have durable cash-generative franchises that fund pipeline regardless of the thesis, mid-cycle pipelines where computational methods apply directly, and valuations that do not require the thesis to work for the position to survive. This is the margin of safety in the book.

NTRA (~12%) is the diagnostics lever -- MRD and NIPT volume is the data substrate for translational ML. ILMN (~7%) is the sequencing oligopoly; EXAS (~5.6%) is screening adjacency; CRL (~4.2%) is the preclinical services layer that runs the experiments computational candidates ultimately need. MRNA (~5%) is the mRNA platform exposure with a cash cushion and a credible computational-design workflow for antigen selection.

The gene-editing tail -- CRSP (~3%), NTLA (~1.3%), BEAM (~1.3%) -- is deliberately small. These are binary clinical-outcome names, and the book's AI thesis does not require them to work. They are held as right-tail optionality on the same platform that benefits from better in silico target engagement. ABCL (~0.8%) is a token position in antibody discovery; small because the platform's commercial model is still unsettled.

Asymmetric payoff

Base case: GILD/ALNY/REGN compound modestly (high-single-digit to low-teens) while one or two mid-cap positions rerate on clinical read-outs aided by computational design. Book returns roughly 12-18% annualized. Bear case: Eroom's Law continues, the thesis fails, and the book returns 0-5% annualized on the cash-generative anchors while the tail goes to zero; overall -5 to +3%. Bull case: a credible approval-rate signal emerges and the tooling/diagnostic layer rerates; 30-45% annualized on a 2-3 year horizon.

At a 60% base, 25% bear, 15% bull, expected value is roughly +11 to +16% annualized. The book is structured so that the margin of safety sits in the anchors; the asymmetric payoff comes from the smaller positions, and the bear case is survivable.

Three things that would change our mind

1. A large-cap pharma pulling out of a headline AI-discovery partnership after two cycles with no candidate advancing to IND.
2. Illumina consumables revenue growth decelerating below 5% for two consecutive quarters, indicating underlying research activity is not expanding.
3. A clean Phase II failure from an asset widely branded as AI-discovered, with no credible mechanistic alternative explanation.

What we are explicitly NOT betting on

We are not buying pure-play AI-native drug discovery platforms with no marketed products or late-stage pipeline. The historical base rate for pre-revenue platform stories in biotech is poor, and the valuation levels required to hold them imply imputed probabilities of success materially above observed clinical base rates. The book expresses the same thesis through names where the downside is bounded by existing franchise cash flow. That is a deliberately more conservative expression of the idea, and we accept a lower right tail in exchange for a survivable left tail.