What this note covers
“AI drug discovery and research services” is not one technique. It is a stack of machine-learning methods aimed at shortening the preclinical path from biological target to drug candidate. The main technical families are:
- Structure-based virtual screening and binding-affinity prediction — often using three-dimensional atomic representations fed into convolutional neural networks.
- Generative molecular design — generating novel small molecules with desired property profiles, often with reinforcement learning or generative models.
- Protein-structure prediction — predicting three-dimensional protein structures from amino-acid sequences, which can enable computational hit discovery when no experimental structure exists.
- Ligand-based activity prediction — training models on assay data to predict antibacterial, antiviral, or other phenotypic activity.
- ADMET property prediction — predicting absorption, distribution, metabolism, excretion, and toxicity before synthesis.
This note stresses a hard distinction: a granted patent is not proof that a technology works commercially. It is evidence of recorded claims. A peer-reviewed paper is stronger evidence of reproducibility than a preprint, but even a top journal paper is a single lineage unless independently replicated.
Source evidence: granted patent
The strongest patent found for the structure-based virtual screening lineage is:
- US 10,474,915 B2, assigned to Atomwise Inc., granted 2019-11-12.
Tier 1 — granted US patent; accessed 2026-04-30 via [patents.google.com
This is a granted patent, not an application. Atomwise’s patent family covers methods that apply convolutional neural networks to molecular structure representations for binding prediction. It is relevant to the claim that AI can prioritize small molecules before they are physically screened.
Why the patent matters for moat/feasibility:
- It is evidence of recorded claims, not evidence of commercial performance.
- A competitor can read the claims to understand what Atomwise asserts it owns.
- It does not by itself show that the method generalizes to new targets, that false-positive rates are acceptable, or that any pharmaceutical customer paid for it.
What the patent does not prove:
- It does not prove that Atomwise’s in silico hits translate to cell-active, drug-like leads.
- It does not prove that the method is faster or cheaper than experimental high-throughput screening.
- It does not prove that any FDA-approved drug emerged from the method.
For the patent to support a moat claim, you would need additional evidence: claim construction, office actions, litigation, licensing agreements, or at least third-party implementation. Not present in this source set.
Source evidence: peer-reviewed papers
3.1 Insilico Medicine generative chemistry — Nature Biotechnology 2019
The most important feasibility paper in this note is:
- Zhavoronkov, A., et al. “Deep learning enables rapid identification of potent DDR1 kinase inhibitors.” Nature Biotechnology 37, 1038–1040 (2019). DOI: 10.1038/s41587-019-0224-x. Published 2019-09-02.
Tier 1 — peer-reviewed, DOI-verified; accessed 2026-04-30 via [doi.org
This paper reports a generative tensorial reinforcement learning approach used to design molecules against the DDR1 kinase. The numbers used in this note’s worked calculation come from this paper:
- 30,000 molecules generated by the model.
- 6 molecules synthesized.
- 4 molecules showed biochemical activity.
- 2 molecules were active in cell-based assays.
- The authors reported identifying leads in 46 days.
These numbers are a single lineage: one research group, one paper, one study. That does not make them false; it makes them insufficient for a broad “AI works” conclusion. The exact 4-of-6 hit rate is [SINGLE-SOURCE] for this note because no independent replication was retrieved.
3.2 AlphaFold2 — Nature 2021
A second independent, peer-reviewed paper supports protein-structure prediction as a credible input to AI drug discovery:
- Jumper, J., et al. “Highly accurate protein structure prediction with AlphaFold.” Nature 596, 583–589 (2021). DOI: 10.1038/s41586-021-03819-2. Published 2021-07-15.
Tier 1 — peer-reviewed, DOI-verified; accessed 2026-04-30 via [doi.org
AlphaFold2 does not identify drugs by itself. It predicts protein structures. Those structures can be used downstream for structure-based virtual screening, binding-site identification, and medicinal chemistry. The paper’s importance is that it provides independent evidence that deep learning can reduce a structural-biology bottleneck. The patent and the Insilico paper are different institutions and different technical lineages, so AlphaFold2 is useful as a cross-technique check on feasibility, though not on any specific commercial hit rate.
3.3 Stokes et al. antibiotic discovery — Cell 2020
A third peer-reviewed paper supports ligand-based deep learning:
- Stokes, J. M., et al. “A Deep Learning Approach to Antibiotic Discovery.” Cell 180, 688–702 (2020). DOI: 10.1016/j.cell.2020.01.021. Published 2020-02-20.
Tier 1 — peer-reviewed, DOI-verified; accessed 2026-04-30 via [doi.org
This paper used a directed-message-passing neural network trained on E. coli growth inhibition to identify the antibiotic candidate halicin. It is evidence that machine-learned activity models can find structurally novel active compounds. Like the Insilico paper, it is one study; it is not an independent replication of a product-level claim.
Note on preprints: No arXiv preprint is used in place of a journal version. All three papers above have journal versions with DOIs.
Cross-reference against company filings
Two public-company SEC filings provide the adoption/risk context.
4.1 Schrödinger, Inc. — Form 10-K
- Schrödinger, Inc. Form 10-K, filed via SEC EDGAR under CIK 0001490978.
Tier 2 — primary regulatory filing; accessed 2026-04-30 via [sec.gov
Schrödinger sells physics-based molecular modeling software and also operates internal and partnered drug-discovery programs. Its 10-K describes revenue from software licenses and drug discovery collaborations. The filing’s risk-factor language is much more conservative than promotional language about “revolutionizing drug discovery.” Typical filed risk language includes statements to the effect that the company may never develop a product, may not maintain or grow software revenue, and may not realize anticipated benefits from collaborations.
This is not a fraudulent discrepancy; it is standard SEC risk-factor language. But it is a real evidence gap between marketing-level claims and filed disclosures. Most “AI will cut drug discovery costs by X%” claims are not independently auditable in these filings.
4.2 Recursion Pharmaceuticals, Inc. — Form 10-K
- Recursion Pharmaceuticals, Inc. Form 10-K, filed via SEC EDGAR under CIK 0001601830.
Tier 2 — primary regulatory filing; accessed 2026-04-30 via [sec.gov
Recursion operates an AI-enabled drug discovery platform and internal pipeline. Its 10-K caveats that platform revenue may not scale, that its partnerships may not translate into milestone or royalty payments, and that clinical candidates may fail. These are standard risk factors, but they directly limit any claim that a platform is commercially validated merely because it exists.
Reported discrepancy: Promotional materials from multiple AI drug discovery companies emphasize speed and predictive power. SEC filings emphasize uncertainty, dependency on partners, and absence of marketed products. The source set does not contain an audited filed number proving a specific cost reduction from AI; such figures should be treated as needs primary source validation unless a company explicitly files them.
Worked calculation: generative hit-rate yield and sample-size fragility
This calculation uses the Insilico 2019 peer-reviewed data. It is not an economic model; it is a statistical exercise to show how fragile a “67% hit rate” claim is when the denominator is six synthesized compounds.
5.1 Inputs
From Insilico 2019 [DOI 10.1038/s41587-019-0224-x]:
- Generated molecules: N_generated = 30,000
- Synthesized molecules: N_syn = 6
- Biochemically active synthesized molecules: N_active = 4
5.2 Formula 1: observed synthesis-to-active hit rate
\[
p_{\text{syn}} = \frac{N_{\text{active}}}{N_{\text{syn}}} = \frac{4}{6} = 0.6667 = 66.7\%
\]
This is the often-quoted hit rate. But with only six observations, the confidence interval is extremely wide.
5.3 Formula 2: Wilson 95% confidence interval for the synthesis-to-active proportion
For a binomial proportion \( \hat{p} = 4/6 \), \( n = 6 \), \( z = 1.96 \):
\[
CI = \left[
\frac{\hat{p} + \frac{z^2}{2n}}{1+\frac{z^2}{n}}
\pm
\frac{z \sqrt{\frac{\hat{p}(1-\hat{p})}{n} + \frac{z^2}{4n^2}}}{1+\frac{z^2}{n}}
\right]
\]
Plugging in:
- \( \hat{p} = 0.6667 \)
- \( n = 6 \)
- \( z = 1.96 \)
- \( z^2 = 3.8416 \)
Denominator:
\[
1 + \frac{3.8416}{6} = 1.64027
\]
Center numerator:
\[
0.6667 + \frac{3.8416}{12} = 0.6667 + 0.32013 = 0.98683
\]
Center:
\[
0.98683 / 1.64027 = 0.6016
\]
Half-width numerator:
\[
1.96 \times \sqrt{\frac{0.6667 \times 0.3333}{6} + \frac{3.8416}{144}}
\]
\[
= 1.96 \times \sqrt{0.03704 + 0.02668} = 1.96 \times \sqrt{0.06372} = 1.96 \times 0.25243 = 0.49476
\]
Half-width:
\[
0.49476 / 1.64027 = 0.3016
\]
So the Wilson 95% CI is approximately:
\[
0.6016 - 0.3016 = 0.3000
\]
\[
0.6016 + 0.3016 = 0.9032
\]
Result: 30.0% to 90.3% at 95% confidence. This means the data are compatible with a true synthesis-to-active hit rate anywhere from roughly one-in-three to nine-in-ten. That is not a precise performance claim.
5.4 Formula 3: generated-to-active overall yield
\[
p_{\text{overall}} = \frac{N_{\text{active}}}{N_{\text{generated}}} = \frac{4}{30,000} = 0.0001333 = 0.0133\%
\]
5.5 Formula 4: molecules needed for 95% probability of at least one active
\[
n = \frac{\ln(1-P)}{\ln(1-p_{\text{overall}})}
\]
where \( P = 0.95 \) and \( p_{\text{overall}} = 0.0001333 \).
\[
n = \frac{\ln(0.05)}{\ln(0.9998667)} \approx \frac{-2.99573}{-0.00013331} \approx 22,470
\]
So under the Insilico 2019 yield, generating approximately 22,470 molecules would be needed to have a 95% chance of observing at least one biochemically active compound if the yield estimate is correct.
5.6 Sensitivity table
| Assumption | Active count | Synthesis hit rate | Overall yield | Molecules for 95% probability of ≥1 active |
|---|---|---|---|---|
| Insilico reported | 4 of 6 | 66.7% | 0.0133% | ~22,470 |
| If only 3 active | 3 of 6 | 50.0% | 0.0100% | ~29,960 |
| If only 2 active | 2 of 6 | 33.3% | 0.00667% | ~44,940 |
| If 5 active | 5 of 6 | 83.3% | 0.01667% | ~17,980 |
| If 6 active | 6 of 6 | 100.0% | 0.0200% | ~14,980 |
Interpretation:
- The “66.7% hit rate” collapses quickly under small-sample uncertainty.
- The difference between 4 and 2 active compounds more than doubles the number of generated molecules needed for 95% confidence.
- This is why a single study, however prominent, cannot bear a whole commercial thesis.
Parameter table
| Parameter | Value | Source | Tier | Measured/derived/assumed/inferred | Cross-verified? |
|---|---|---|---|---|---|
| Generated molecules in DDR1 study | 30,000 | DOI 10.1038/s41587-019-0224-x | Tier 1 | Reported/measured in study | No — single-source |
| Synthesized molecules in DDR1 study | 6 | DOI 10.1038/s41587-019-0224-x | Tier 1 | Reported/measured in study | No — single-source |
| Biochemically active synthesized compounds | 4 | DOI 10.1038/s41587-019-0224-x | Tier 1 | Reported/measured in study | No — single-source |
| Synthesis-to-active hit rate | 66.7% | Derived from above | Tier 1-derived | Derived | No |
| Wilson 95% CI lower bound | 30.0% | Calculated from DOI 10.1038/s41587-019-0224-x | Derived | Derived | No |
| Wilson 95% CI upper bound | 90.3% | Calculated from DOI 10.1038/s41587-019-0224-x | Derived | Derived | No |
| Overall generated-to-active yield | 0.0133% | Calculated from DOI 10.1038/s41587-019-0224-x | Derived | Derived | No |
| Molecules for 95% probability of ≥1 active | ~22,470 | Calculated from DOI 10.1038/s41587-019-0224-x | Derived | Derived | No |
| z-value for 95% CI | 1.96 | Standard normal distribution | Standard statistical constant | Assumed/standard | Standard — yes |
| Atomwise CNN-binding patent grant date | 2019-11-12 | US 10,474,915 B2 | Tier 1 | Recorded grant date | Yes — USPTO record |
| AlphaFold2 institution | DeepMind | DOI 10.1038/s41586-021-03819-2 | Tier 1 | Reported | Different institution from Insilico |
| Antibiotic discovery model | Stokes et al. 2020 | DOI 10.1016/j.cell.2020.01.021 | Tier 1 | Reported | Different institution from Insilico |
Load-bearing claim warning: The exact “4 of 6 active” and “66.7% hit rate” claims are [SINGLE-SOURCE]. They are not independently replicated in the retrieved sources. The Atomwise patent, AlphaFold2 paper, and Stokes paper are independent lineages, but they do not reproduce the Insilico numbers.
Unit economics / business model
Who monetizes this technology?
Three business models operate in AI drug discovery:
- Software license/subscription model. Companies sell access to computational chemistry or ADMET modeling software on a per-seat, annual, or enterprise basis. Schrödinger, Certara, and Simulations Plus operate near this model. Exact gross margins are not stated in this note; software margins are generally high, but needs primary source validation for each company.
Source for Schrödinger software revenue description: Schrödinger 10-K, CIK 0001490978 Tier 2; accessed 2026-04-30 via [sec.gov .
- Research services and partnership/milestone model. Platforms sign collaborations with pharmaceutical partners, receiving upfront research fees, milestone payments upon target validation or IND filing, and sometimes royalties. Recursion’s 10-K describes platform revenue and collaboration arrangements subject to milestone and royalty uncertainty.
Source: Recursion Pharmaceuticals 10-K, CIK 0001601830 Tier 2; accessed 2026-04-30 via [sec.gov .
- Internal pipeline/asset-creation model. Some companies use AI to create internal drug candidates, then out-license or sell them. This monetizes through downstream milestones, royalties, or M&A. This model is visible in Recursion and Schrödinger’s internal programs, but current revenue is not guaranteed.
Who captures the value?
- Platform owners capture near-term value through software fees and research payments.
- Pharmaceutical partners capture late-stage value if a candidate is validated, because they usually retain development and commercialization rights under collaboration agreements.
- GPU/cloud providers capture compute spend regardless of drug success.
No retrieved source supports a specific dollar margin; any margin figure should be treated as needs primary source validation.
Disconfirming Evidence
The AI drug discovery thesis would be materially weakened if:
- The Insilico 2019 4-of-6 hit rate does not replicate in independent, pre-registered studies. The current Wilson interval of 30.0%–90.3% already shows the original data are weak evidence on their own.
- The Atomwise patent is invalidated, narrowed, or never commercialized. A patent that cannot be enforced or licensed is a weak moat.
- No pharmaceutical partnership converts to a materially disclosed milestone or royalty payment. Securing pilots is not the same as earning milestone revenue; if collaboration revenue remains immaterial in subsequent 10-K filings, the platform claim is not commercially validated.
- AI-designed clinical candidates fail in Phase 2 or Phase 3 for efficacy. Early hit identification is not the same as human efficacy; a clinical failure that could be attributed to the AI-generated candidate would undermine the “better drugs” claim.
- Regulatory guidance changes to require traditional wet-lab validation on every in silico hit. This would raise costs and reduce the value of scaling purely computational screening.
Any one of these would not automatically falsify the entire field, but together they would refute the strongest version of the investment/moat thesis currently circulating.