AI and new medicines: what turns a designed candidate into value? $ABCL, $ABSI, $RXRX, $SDGR

01The useful question is what has become more certain

Artificial intelligence can help a drug researcher decide which experiment to run, which protein to investigate, or which molecule to make. Those contributions can be valuable well before a medicine reaches a patient. They become investable evidence, however, only when the claim is connected to an observable result and to an economic right. A promising prediction, a successful laboratory assay, a clinical response and a royalty payment answer four different questions.

AbCellera, Absci, Recursion and Schrödinger make that distinction unusually concrete. Their activities include antibody discovery, generative protein engineering, phenotypic biology and computational chemistry. They also use different ownership arrangements. A company may retain a clinical candidate, license an existing molecule, conduct research for a partner, or contribute programs to a separate biotechnology company. The scientific contribution and the shareholder's claim on the eventual proceeds need to be examined together.

The relevant unit of analysis is therefore the specific program at a specific decision point. Ask what uncertainty has been reduced since the previous announcement. Has an experiment confirmed that manipulating a target changes disease biology? Has a molecule achieved a useful exposure in people? Has a randomized trial separated its effect from placebo? Has a partner accepted responsibility for expensive development? A favorable answer to one question does not automatically settle the others.

This approach also allows positive developments to retain their proper meaning. A discovery milestone can be a real achievement without being clinical proof. A promising patient study can justify further investment without establishing regulatory approval. A financing can improve the ability to complete a trial while increasing the number of claims on future value. The following analysis follows those distinctions through identifiable programs and contracts, using information available on September 26, 2026.

02Four different jobs can carry the same AI label

The first job is learning relationships in biology. A model might connect a gene perturbation, a cellular image and a disease state, then suggest a target worth testing. The output is a hypothesis about how the system behaves. Its value depends on whether the relationship survives interventions that the model has not already seen. Recursion's phenotypic approach belongs partly in this category; predicting useful biology is different from designing the eventual chemical structure.

The second job is finding and selecting biological molecules. Antibodies generated by an immune system provide a large starting population with different binding properties. Screening, computation and engineering can narrow that population toward candidates with useful activity and practical development characteristics. AbCellera describes such an integrated process. The participation of machine learning does not mean that every antibody was invented from scratch by a generative model.

The third job is directly proposing or modifying molecular sequences. Absci's generative approach to antibodies raises questions about which properties the model optimized, how those properties were measured, and whether improvements persisted during experimental testing. Generating a plausible sequence is only the beginning: an antibody also needs appropriate specificity, stability, production characteristics and exposure. Improving one property can leave another unchanged or make it worse.

The fourth job combines molecular simulations with statistical methods to guide chemistry. Schrödinger uses physics-based calculations alongside AI in discovery and optimization. Here, the practical question may concern which compound binds more effectively, which chemical modification preserves activity, or which synthesis should be prioritized. Across all four jobs, the description of the model should lead to a description of the experiment. The commercial label alone does not reveal what has actually been demonstrated.

03An evidence chain is more useful than a technology ranking

A useful evidence chain starts with a computational prediction and then moves through increasingly demanding settings. Laboratory experiments test whether the predicted property exists under controlled conditions. Cellular and animal work can investigate mechanism, exposure and biological consequences. Human studies then examine whether an acceptable dose can be given and whether the treatment produces a benefit in the intended population. Each transition introduces new uncertainties that were absent from the previous setting.

The chain is not a universal scorecard. An antibody for menopausal symptoms, an oncology small molecule and an early immunology program need different experiments. Their study populations, treatment durations and acceptable risks differ. Ranking their companies by the number printed after the word phase would discard much of the information that matters. A later-stage study can still leave a crucial commercial or safety question unresolved, while an earlier experiment can decisively reject a weak hypothesis.

The important distinction is between progress in a program and validation of an entire discovery platform. A successful trial supports the molecule, regimen and patient population that were tested. It does not measure what would have happened if another discovery method had been used. Demonstrating a platform-wide advantage requires a broader record, including unsuccessful programs, costs, timelines and appropriate comparisons rather than a collection of selected successes.

For investors, this suggests reading an announcement twice. First identify the result on its own terms. Then identify the inference being made from it. Evidence that a compound reaches the bloodstream supports an exposure claim. Evidence that patients improve relative to a concurrent control supports a clinical claim. Evidence that a partner pays for further research supports a commercial commitment. Moving from one category to another requires additional observations, not more enthusiastic language.

04Training data matter because experiments define the target

A model learns the patterns represented in its training data, including the limitations of the measurements. In drug discovery, the label being predicted may be binding, cellular activity, expression, toxicity or another property. The practical quality of that label matters as much as the model's architecture. A consistent, relevant assay can create useful information; a noisy or poorly matched assay can teach a sophisticated model to optimize the wrong thing.

The 2026 study by Saunders, Comyn and colleagues in mAbs, Efficient inference of non-polyreactive antibody variants dependent on local fine-tuning, illustrates a precise version of this issue. The authors examined antibody polyreactivity, using experimental measurements to adapt protein models. Their reported improvement after local fine-tuning did not transfer in the same way to a separate, out-of-distribution set of clinical antibodies. The paper's abstract supports a conclusion about a defined laboratory task, not patient benefit.

That distinction is financially relevant. A platform may become very good at improving closely related candidates within a program. Such an advantage can reduce wasted experiments even if it does not generalize equally well across unrelated targets. Conversely, a headline about a large dataset says little unless the dataset covers the problems the next program will face. The number of observations cannot substitute for their relevance.

A practical assessment therefore asks how a test set was separated from training, whether the examples resemble the proposed use, and whether predictions were followed by prospective experiments. It also asks which failures were retained in the learning process. Negative results can be informative when they are measured consistently and incorporated into the next decision. The useful feedback loop is between predictions and reliable experiments; simply accumulating more model outputs does not create the same evidence.

05Better molecular properties still require a complete drug profile

Drug design is a problem of simultaneous constraints. Strong binding may be useful, but a molecule also needs to act in the relevant tissue at a practical exposure. An antibody that binds well can still present manufacturing or specificity problems. A small molecule can show attractive cellular activity but have an unsuitable metabolic profile. Computational methods can help navigate these tradeoffs, while experimental development determines which tradeoffs are acceptable.

An informative external example is the primary study by Ren and colleagues, A small-molecule TNIK inhibitor targets fibrosis in preclinical and clinical models, published online in 2024 in Nature Biotechnology. The authors described AI-assisted target identification and compound generation, followed by preclinical testing and a randomized phase 1 study in healthy participants. Those are distinct pieces of evidence. The phase 1 component assessed safety, tolerability and pharmacokinetics rather than proving benefit in patients with fibrosis.

The broader lesson is about attribution. An accelerated path to a nominated candidate can be operationally meaningful, but the starting point and endpoint of the timing claim must be explicit. Time from a first hit to candidate selection excludes the work needed to build the platform and the subsequent clinical program. A comparison with another company's historical timeline may involve different targets, existing knowledge, budgets and development standards.

For the four companies considered here, an efficient discovery process would be most valuable if it repeatedly produced candidates capable of answering important clinical questions. That possibility is worth examining without treating speed as a substitute for quality. The relevant package includes the molecule's properties, reproducibility, development plan, manufacturing practicality and ownership. A faster route to an experiment is useful when the experiment helps make a better decision about that package.

06AbCellera begins with biological diversity and ends with development choices

AbCellera's current platform description starts with antibodies derived from immune cells and follows them through discovery, selection, engineering and development. Its earlier Gilead collaboration announcement explicitly described a technology stack that included single-cell methods, computation, machine learning and automation. This supports describing a data-intensive antibody discovery process. It does not justify attributing every current candidate to a named generative model.

The business significance is that a useful antibody is more than a sequence that recognizes a target. Researchers must choose among candidates with different functional behavior and development characteristics. The same target may be approached through different epitopes, binding strengths and molecular formats. Selection can influence whether a program later faces difficult formulation, production or dosing decisions. An integrated process aims to identify those issues earlier, although the degree of success must be assessed program by program.

That makes AbCellera a different case from a company selling access to a general-purpose AI system. The relevant outputs are therapeutic candidates, development capabilities and contracts around those candidates. Internal programs expose shareholders more directly to clinical outcomes and future development spending. Partnered programs can distribute costs and responsibilities while retaining a narrower downstream economic interest. Both models can coexist, and their financial statements capture different portions of the activity.

The AbCellera Stock Hub provides a continuing reference for its pipeline and scheduled updates. In this comparison, the central question is how the scientific selection process connects to an actual clinical decision. ABCL635 supplies an especially useful example because the discussion has moved beyond a candidate's predicted properties into a randomized study involving the people the treatment is intended to help.

07ABCL635 brings the discussion to a controlled patient experiment

ABCL635 is an investigational antibody directed at the neurokinin 3 receptor, being developed for vasomotor symptoms associated with menopause. In its August 10 announcement, AbCellera reported positive topline results from the phase 2 portion of its phase 1/2 study. Ninety-two postmenopausal women were randomized to a single 600 mg subcutaneous dose or placebo. The company reported that the primary assessments of symptom frequency and severity at four weeks were met.

A concurrent placebo group makes a substantial difference to interpretation. Symptoms can fluctuate, and enrolling in a trial can change how they are observed and recorded. Randomization provides a structured comparison under the same protocol rather than asking readers to compare a treated group with an unrelated historical population. The result addresses the tested antibody, dose, follow-up and patient criteria; it does not establish superiority to every available treatment.

The next level of assessment concerns the full clinical profile. How consistent was improvement across participants? How long did the effect persist? What adverse events occurred, and how were they distributed? What does a longer treatment period reveal about tolerability? These questions affect whether the observed effect can become a useful treatment option. A positive endpoint is important, but the regimen still has to fit the needs of patients and clinicians.

The AI connection should remain proportionate. A successful patient experiment can strengthen confidence in the candidate emerging from AbCellera's discovery process. It does not isolate the contribution of computation from immune-cell discovery, laboratory screening, engineering, clinical design and the underlying target biology. Treating the result as proof of the whole technology category would obscure the actual achievement: a specific investigational antibody has generated controlled evidence that warrants further scrutiny and development.

08The October presentation can add detail without changing the study's identity

On September 16, AbCellera announced an oral presentation at the IMS World Congress on Menopause, scheduled for October 1. As of this article's research date, that presentation remains ahead. Its value lies in the additional information it can provide about the phase 2 results, rather than in the mere fact that a conference slot exists. A scheduled scientific presentation is not a new trial result already in hand.

Detailed data may help readers connect an average effect with the distribution of individual outcomes. A treatment can produce a favorable mean while leaving important questions about consistency, duration and discontinuation. The assessment also needs to distinguish prespecified endpoints from exploratory analyses. Exploratory observations can guide the next study, but they do not carry exactly the same evidentiary weight as a planned comparison that the trial was designed to test.

An antibody's dosing interval is another practical issue. Less frequent administration can be attractive if benefit lasts and the safety profile supports it. Yet a longer-lasting drug also remains in the body after dosing stops. The convenience proposition therefore needs to be considered alongside reversibility, monitoring and patient selection. Duration of exposure and duration of clinical benefit are related questions, but they are not interchangeable measurements.

For the business, richer phase 2 evidence can inform the scope and cost of the next development step. The requirements for later trials, supply, manufacturing validation and regulatory interaction remain separate workstreams. Stronger evidence may improve the case for funding those activities; it does not eliminate them. The most informative follow-up to the presentation will explain both what the data establish and how those findings shape a concrete development plan.

09AbCellera's partnerships show how costs and upside can be divided

The July 29 Vertex agreement provides a clear contractual example. AbCellera leads discovery and early development of multispecific T-cell engagers, while Vertex funds research and development and receives development and commercialization rights. The announced economics include $28 million in total upfront payments and potential milestone and royalty payments. These terms concern the collaboration programs, not ABCL635.

The June 17 Jazz collaboration uses an option structure. Jazz can obtain exclusive worldwide rights to individual programs after exercising its option and paying the associated fee. Upfront amounts are staged across the initial programs and a further program expected to start within twelve months. The large potential milestone figures require future events and cannot be treated as cash already earned.

These structures illustrate why a pipeline chart alone is an incomplete financial map. Two molecules at a similar scientific stage may give shareholders very different exposure. One may require the company to fund all subsequent trials. Another may provide research funding and a contingent royalty while a partner bears much of the development cost. A third may remain subject to an option that the partner can choose not to exercise.

The economic tradeoff is not automatically favorable in one direction. Retaining rights preserves more possible upside but also concentrates future spending and execution risk. Partnering can reduce that burden while limiting what the originator receives if the product succeeds. The useful question is whether the retained economics are proportionate to the obligations that remain. Contract structure helps explain why a research payment, an option exercise and a clinical result can each matter without meaning the same thing.

10Absci's ABS-201 makes the design claim testable

Absci describes ABS-201 as a generatively designed antibody targeting the prolactin receptor. The program is being investigated in androgenetic alopecia, with development also planned in endometriosis. The two indications involve different clinical questions. A shared molecular target does not make the trial endpoints, patient populations or treatment expectations interchangeable. The mechanism is a scientific hypothesis that must be translated into an acceptable human regimen and a measurable benefit in each setting.

For an antibody design program, the attractive proposition is that computational work can help assemble a useful combination of properties rather than optimizing only binding. A candidate must recognize the intended target, avoid problematic interactions, be produced reliably and maintain appropriate exposure. These properties can be investigated before a large efficacy trial, allowing some weak candidates to be rejected earlier. None of them independently establishes that altering the target will improve the disease.

This distinction separates molecular execution from biological validation. A well-engineered antibody can fail because the target is less important in human disease than expected. Conversely, useful target biology can be difficult to exploit if the molecule's exposure or safety is unsuitable. Evidence should therefore be organized around both questions: whether the candidate behaves as designed and whether that behavior produces a meaningful clinical effect.

ABS-201 gives investors a concrete sequence to follow. Early clinical data can inform dose and interval. Subsequent controlled observations can test benefit. A larger program can then examine durability and a broader safety experience. The purpose of this sequence is to progressively constrain uncertainty. Describing the antibody as AI-designed identifies part of its origin; it does not change the clinical burden of showing that the resulting medicine helps patients at an acceptable risk.

11HEADLINE's first observations concern exposure and tolerability

In its August 11 second-quarter update, Absci described interim HEADLINE observations using aggregated data while treatment assignment remained blinded. The reported information concerned safety, tolerability and pharmacokinetics, including an estimated prolonged half-life. The company expected initial proof-of-concept information in the second half of 2026 and fuller twenty-six-week data in early 2027. These expectations should not be rewritten as completed efficacy results.

A long half-life can support investigation of a less frequent dosing schedule. It does not show that hair growth improves, determine the magnitude of any improvement or establish the best dose. Clinical activity depends on the relationship between exposure, target engagement and the relevant tissue response. A convenient schedule is commercially useful only if the effect and safety profile support the way it would actually be prescribed.

Blinded aggregate observations also have an interpretive boundary. Without separating treatment groups, readers cannot use those figures as if they were a randomized efficacy comparison. They can still inform the conduct of the study, and a safety monitoring process can operate while participants and investigators remain blinded. The key is to describe the released dataset accurately rather than filling in a treatment effect that has not been disclosed.

When efficacy observations arrive, the analysis should move to the prespecified measurement, the placebo comparison, baseline characteristics and treatment duration. Hair outcomes can require time, consistent measurement and careful handling of missing data. An early signal can justify further work while remaining sensitive to small numbers and follow-up. The next useful milestone is therefore a sufficiently described clinical dataset, not another repetition of the antibody's computational origin or its exposure profile.

12STORYLINE's revised timing changes the financing calendar

The September 24 STORYLINE announcement describes a planned randomized, double-blind, placebo-controlled phase 2 study in endometriosis. The preliminary design anticipates approximately 150 participants across roughly ten or more countries, with dose selection informed by HEADLINE. Initiation is now expected around the middle of 2027, subject to protocol finalization and regulatory clearance. The study has not already delivered patient efficacy data.

The timing matters because the August business update had referred to an anticipated fourth-quarter 2026 start. The later announcement is the current planning reference. A revised start date moves the expected interval before enrollment, treatment and follow-up can generate evidence. It also affects how investors should map operational spending and possible financing needs. The available statements do not justify inventing a reason for the change or assigning it a precise cost.

The clinical design raises its own questions. Endometriosis-related pain must be assessed in an appropriate population with measures that can capture meaningful improvement. A placebo-controlled design helps separate treatment effect from variation over time and study participation. Dose-ranging can investigate whether more exposure improves benefit, increases adverse effects, or simply adds little. Those are prospective questions, not conclusions embedded in the proposed protocol.

For valuation, the implication is a shift in when uncertainty may be resolved, rather than an automatic declaration that the molecule has become better or worse. A later study can still be informative; a longer interval can still have a financing cost. Keeping these dimensions separate allows a disciplined reading of the update. The appropriate response is to revise the calendar and examine the evidence as it develops, without confusing a planning change with a clinical outcome.

13Lilly's investment and ABS-201 rights answer different questions

Absci's second-quarter release reported a completed $100 million equity financing that included a $40 million strategic investment from Lilly. It explicitly stated that the investment did not confer program rights on Lilly. That distinction is central to the economic analysis. Purchasing shares gives an investor exposure to the company; acquiring a license gives contractual rights to a program. The participation of a pharmaceutical company does not, by itself, convert one arrangement into the other.

Equity capital can fund experiments and extend the time available to make decisions, but it also increases the pool of securities sharing future value. The relevant comparison is not simply cash before and after financing. It is the additional development capacity obtained, the ownership sold and the remaining clinical uncertainty. A well-timed raise can improve strategic flexibility while still diluting existing holders. Both statements can be true at once.

The reported June 30 balance of cash, equivalents and marketable securities was $201.1 million. Management estimated runway into the second half of 2028 under its plans. A runway statement is a forecast of resources and spending, not a guarantee that every contemplated study can be completed under every scenario. Trial size, development choices, manufacturing commitments and additional programs can change the path.

Retained ownership of ABS-201 gives Absci meaningful exposure to future outcomes, but ownership should always be read with the development bill beside it. Advancing a candidate in more than one indication can increase opportunity and expense simultaneously. The investor's task is to judge whether each next experiment is capable of resolving a commercially important uncertainty. A large partner's name is useful context; the actual contract and financing terms determine what it means.

14REC-4881 is a case of biological discovery around a licensed molecule

REC-4881 is a MEK1/2 inhibitor being studied in familial adenomatous polyposis, or FAP. Recursion's original licensing announcement and its January 2026 filing identify the molecule's origin in a Takeda license. Recursion's contribution involved phenotypic insights connecting the mechanism with disease biology. It would be inaccurate to describe the original chemical structure as a newly generated Recursion AI molecule.

This provenance does not diminish the potential importance of identifying a useful therapeutic application. Drug discovery can create value by finding a new use for an existing compound, provided the biology and clinical results support it. The relevant scientific question is whether modulation of the pathway improves the disease in the tested setting. The relevant economic question also includes the obligations associated with the in-license, rather than assuming unencumbered ownership.

The TUPELO program supplies the clinical context. FAP is a different development problem from treating an advanced solid tumor, even when a molecular pathway may overlap. The patient population, disease course, acceptable treatment burden and useful endpoints need to match the intended use. A mechanistic label such as MEK inhibition does not make results from unrelated indications interchangeable or settle the risk-benefit assessment in a chronic hereditary condition.

For investors, the provenance distinction prevents a common attribution error. A clinical benefit could validate a disease hypothesis and strengthen the candidate's value while providing only indirect evidence about a broader discovery platform. The program can succeed or fail on its actual clinical merits. Understanding which part of the process Recursion contributed makes the achievement more interpretable and helps separate molecule ownership, biological insight and platform capability.

15REC-7735 presents a different test of the discovery process

REC-7735 is a separate program: an AI-designed inhibitor directed at the PI3Kα H1047R mutant. Recursion's second-quarter update reported IND clearance and an expected start of the phase 1/2 ZINNIA study in the second half of 2026. IND clearance permits the clinical investigation to proceed under the applicable framework; it is not drug approval or evidence of patient benefit. A planned start is also distinct from a confirmed first patient dosed.

Here the design question concerns whether a chosen chemical profile can translate into a usable therapeutic window. Mutant selectivity can be a meaningful design objective, but its clinical relevance depends on exposure, activity and adverse effects at the administered dose. Laboratory selectivity ratios cannot by themselves identify the dose that will work in people. The initial clinical program must investigate those relationships rather than merely repeat the preclinical description.

This makes REC-7735 a different kind of platform test from REC-4881. One program involves designing a candidate with desired molecular properties; the other involves developing a licensed compound around a biological hypothesis. Combining them into a single count of AI-generated medicines would erase that difference. Keeping them separate allows later results to be attributed to the work that actually preceded them.

It also clarifies the financial horizon. An early clinical oncology program typically generates information in stages, with dose exploration preceding a stronger assessment of efficacy in defined populations. The timing and size of those steps affect spending, potential partnerships and the need for additional capital. Candidate nomination or IND clearance can justify entering that process; neither removes the subsequent cost of learning whether the intended selectivity produces a clinically useful result.

16The Genentech option is a discovery milestone with commercial substance

Recursion's August update announced that Genentech exercised a validated target option in the neuroscience collaboration. The company described a sequence beginning with computational predictions from neuronal phenomic maps and proceeding through experimental validation. The target entered an early small-molecule discovery program. That is a meaningful progression from hypothesis toward a funded development path, while remaining far earlier than evidence of benefit in patients.

The distinction between target validation and drug validation is essential. Demonstrating that altering a target changes relevant biology does not establish that a practical medicine can alter it safely in humans. Chemistry, tissue exposure, specificity and tolerability still need to be solved. Even a well-supported target may be difficult to drug, and the eventual candidate can introduce properties that were absent from the original biological experiment.

A partner's option exercise adds a separate commercial observation. It shows a decision under a contract to advance work beyond the preceding stage. That decision is more concrete than a general statement of enthusiasm, but it still reflects the partner's judgment under uncertainty. Investors should distinguish money already paid, contractual commitments and milestone ceilings that depend on future success. The total possible value of a collaboration is not a present receivable.

This case also provides a useful standard for future announcements. The strongest updates specify what the initial prediction was, what experiments tested it and what decision followed. The more clearly those pieces connect, the easier it is to assess whether the platform is producing useful hypotheses. The clinical distance that remains should be visible alongside that progress. Scientific progress and commercial recognition can be real without being compressed into the much stronger claim that a medicine has already been validated.

17Data access and model rights are assets with boundaries

On September 21, Recursion and Tempus announced an extension of their data relationship and a separate license for Recursion's RNA foundation model. The arrangements concern access to data and use of a model. They should not be interpreted as a clinical readout or automatically assigned to the economics of REC-4881 or REC-7735. Different contracts can support the same research organization while governing distinct assets.

A dataset can be valuable because it connects measurements with clinical context that is difficult to reproduce. Its usefulness depends on quality, permitted uses and relevance to the problem being studied. Access alone does not establish exclusive ownership, and a model trained on licensed information may operate under specific contractual boundaries. Investors should examine those boundaries before equating a large data relationship with a permanent competitive advantage.

The financial direction also matters. A company paying for access to external information is making a research investment. A company licensing its own model may receive a different form of consideration under separate terms. Combining both sides of a relationship into a single partnership-value number can obscure who pays whom and for what. Headline amounts need to be assigned to the correct obligation before they enter a cash-flow model.

This is especially relevant when technological capabilities improve rapidly. A model can become more useful through better data and experimentation, but competitors may also improve. Durable value may reside in a combination of rights, proprietary measurements, laboratory execution and accumulated experience rather than a model alone. The September agreement is therefore best understood as part of Recursion's research infrastructure and contractual position, with its clinical consequences still dependent on the programs using that infrastructure.

18Schrödinger connects computational chemistry with actual drug programs

Schrödinger is a hybrid business with software operations and a drug-discovery portfolio. Its relevance to biotechnology rests on the molecules it helps discover and the rights retained around them. The software business has its own adoption, licensing and accounting questions. Those questions should remain visible without displacing the therapeutic programs at the center of this analysis. Selling a research tool and owning part of a medicine represent different paths to revenue.

The company's August business update describes an approach combining physics-based simulation with AI. It also identifies a collaboration with Simcere in which Schrödinger leads drug design and optimization during joint research, while Simcere leads subsequent preclinical and clinical development. Potential milestone and royalty economics illustrate how computational contribution can be connected to a therapeutic asset rather than only a software subscription.

The scientific value proposition is practical: better predictions can help prioritize compounds and reduce unproductive cycles of synthesis and testing. This is a testable claim at the level of a chemistry campaign. It is not the same as demonstrating that a target is clinically relevant or that a trial will succeed. A program can benefit from improved chemistry and still fail because the biological hypothesis does not hold in patients.

The business model therefore needs two separate assessments. One concerns demand for the platform as a research product. The other concerns the development and economic outcomes of specific drug programs. Strong software adoption does not automatically increase the probability that a particular candidate will work. A valuable drug collaboration can produce substantial proceeds without turning the software segment into a pharmaceutical sales business. Maintaining those distinctions is the starting point for reading Schrödinger's financial results.

19Tectora shows how discovery value can move into a new company

On September 9, Schrödinger announced the Tectora licensing and collaboration arrangement. The new biotechnology company received the early immunology and inflammation programs SDGR-4594 and SDGR-8139, with financing from NEA and RA Capital. Schrödinger received equity and retained potential milestone and royalty economics. The announced $55 million Series A funds Tectora; it is not $55 million of Schrödinger revenue.

This arrangement changes where the next development decisions and financing burden sit. A separate company can assemble a focused team and capital around selected programs. The originating platform participates through its contractual interests rather than necessarily carrying the entire clinical development organization. That can be a rational way to turn discovery output into a more specialized development effort, provided the programs attract sufficient resources and execute successfully.

The retained equity and contractual payments also carry different risks. Equity value depends on the new company's financing, dilution, progress and eventual liquidity opportunities. Milestones require specified events. Royalties require a successful product, sales and the applicable contractual conditions. Adding their undiscounted potential together would not produce a sensible current value. The timing and dependencies need to remain attached to each component.

Tectora is still an early-stage drug-development proposition, not a patient efficacy result. Its significance is that identifiable programs have moved into a funded structure with defined categories of retained economics. Future evidence must show whether those programs can become clinical candidates and then useful therapies. For Schrödinger shareholders, the relevant questions include the quality of those assets, the capital they require and the proportion of eventual success that returns to the listed company.

20Schrödinger's revenue composition explains why headline growth is incomplete

For the quarter ended June 30, 2026, Schrödinger reported $58.889 million of revenue. The components were $32.544 million from software products and services, $22.986 million from drug discovery and $3.359 million of contribution revenue. The financial chart shows those reported categories, with percentages calculated from their common total. It does not estimate the amount attributable to AI or the value of any drug candidate.

The distinction matters because the categories arise from different activities. Software revenue reflects licensing and services under their accounting terms. Drug-discovery revenue can include collaboration milestones, whose timing may be uneven. Contribution revenue is reported separately. A shift in the mix can change quarterly growth without indicating the same change in recurring customer demand or in clinical progress.

The company identified a $10 million collaboration milestone associated with the Ajax transaction within the quarter's drug-discovery revenue. Such an event can be economically meaningful while remaining unsuitable for simple annualization. An investor projecting every quarter from the latest total would need to establish why an equivalent event should recur. The appropriate forecast separates contractually supported recurring activity from event-dependent payments.

The same discipline applies below revenue. Gains on investments, operating results and cash proceeds are different measures. A company can recognize income related to an ownership interest while continuing to spend heavily on its operating activities. Understanding that combination is particularly important for a company-creation model. The most useful reading connects each line to the activity that generated it and asks whether that activity can repeat, rather than treating all reported growth as sales of a successful medicine.

21AbCellera's resources illustrate why liquidity definitions matter

AbCellera's June 30 balance sheet reported $120.065 million of cash and equivalents and $420.039 million of marketable securities, totaling $540.104 million. A further $25 million was restricted cash. The financial chart separates these items. Government funding available under applicable arrangements is another category and should not be silently added to unrestricted cash as if every dollar were immediately available for any purpose.

The distinction is relevant to development planning. Cash can fund operations, securities may be converted into cash under their terms, and restricted balances have limitations. Reimbursement-based support can depend on qualifying expenditures and administrative processes. A company may reasonably describe a broader funding capacity, but investors need the underlying components before using the number to estimate resilience. The most expansive liquidity headline is not automatically the correct numerator for runway.

These are also historical balances rather than a September cash update. On August 13, AbCellera priced an offering expected to raise approximately $200 million gross, using common shares and pre-funded warrants. The pricing announcement expected closing on August 14 and identified clinical pipeline development among the uses of proceeds. Gross pricing terms should not be converted into an exact present net-cash balance without the closing, expenses and intervening spending.

This financing illustrates how a favorable clinical development can be followed by capital raising. The company can seek resources to pursue a larger opportunity while holders absorb additional ownership claims. The financing does not establish the clinical outcome, and the clinical result does not eliminate the financing effect. Both belong in the same analysis, with dates and definitions that prevent a historical chart from being mistaken for today's spendable balance.

22Cash runway is a plan, while development creates choices

The four business models convert money into evidence differently. An internally owned program requires the company to decide how much development to fund itself. A partner-funded collaboration can support research while transferring later responsibilities. An in-licensed asset may carry future payment obligations. A program contributed to a new company can shift direct spending elsewhere while leaving exposure through equity and contingent rights. Comparable cash balances would not make these obligations comparable.

Recursion's August update reported $556.8 million of cash, equivalents and restricted cash at June 30, with management expecting resources to support its plan into early 2028. Absci's forecast extended into the second half of 2028. Those dates summarize management assumptions about scope, timing and spending. They are not guarantees that every program reaches a decisive result before another financing is needed.

A more useful exercise is to map the decisions that consume capital. Will a company expand a trial after preliminary activity? Will it add an indication, build manufacturing capacity or retain commercialization rights? Will a partner exercise an option or decline to proceed? Favorable evidence can increase the amount of worthwhile investment available, so a growing development budget is not inherently a negative signal. It does, however, require a financing path.

The investor should also distinguish resource sufficiency from resource efficiency. A large balance can finance an unproductive program for a long time; a smaller company can obtain useful evidence through a focused study. The strongest position combines a meaningful question, a design capable of answering it and enough funding to act on the answer. AI can influence the efficiency of some research steps, but it does not remove the need to choose which clinical and commercial uncertainties deserve the next dollar.

23The next events should be judged by the information they add

For AbCellera, the October 1 ABCL635 presentation can deepen understanding of an already announced phase 2 result. For Absci, the important distinction is between expected HEADLINE proof-of-concept information and the later planned start of STORYLINE. For Recursion, additional REC-4881 data scheduled for November 2 and the planned ZINNIA initiation concern different molecules and different development questions. For Schrödinger, advancement of the Tectora programs would have to establish progress beyond the financing and licensing structure.

A constructive scenario would involve increasingly specific evidence: a clearer clinical effect, a dose with a workable safety profile, a reproducible experimental result or a partner decision tied to defined progress. Financial strength would come from funding those steps on terms that preserve a sensible share of the opportunity. No single announcement needs to solve every problem, but the sequence should make the next decision better informed.

A weaker scenario can take several forms. An attractive exposure profile may fail to translate into benefit. A target can remain difficult to drug despite convincing laboratory biology. A partner can reduce investment for reasons that affect the program's timing. A company can retain substantial rights but lack the resources to exploit them without significant financing. These outcomes operate through different mechanisms and should not be collapsed into a generic verdict about AI.

The most useful calendar therefore records what each event can establish, not merely its date. A conference presentation, first patient dosed, data release, option exercise and financing close each changes a different part of the picture. Matching the expected information with the unresolved question helps investors recognize real progress and avoid interpreting a routine operational milestone as a result that has not yet been produced.

24A durable framework follows the experiment and the contract

Reading AI-enabled biotechnology becomes easier when each program is described in one connected sentence: this method produced this candidate or hypothesis; this experiment tested this property; this company holds these rights and must fund these remaining activities. If any part is missing, the headline is incomplete. The framework accommodates genuine scientific progress without requiring every success to validate an entire technology category.

ABCL635 brings a controlled patient result into an antibody discovery story. ABS-201 creates a sequence from engineered properties and early human exposure toward indication-specific efficacy studies. REC-4881 and REC-7735 demonstrate why licensed compounds and newly designed candidates need separate attribution. Tectora shows that useful discovery output can be developed through a new company, leaving the originator with a combination of ownership and contractual claims.

The financial reading follows the same logic. Distinguish operating revenue from financing, liquid assets from conditional funding, and future milestone ceilings from earned payments. Identify which entity receives the money and which entity pays for the next experiment. Preserve the dates of balances and forecasts. These details determine whether a scientific achievement can create value for the shareholders of the company being examined.

The documents below provide the clinical, contractual and financial anchors. Company announcements are the primary sources for the current program updates; planned events remain conditional. The scientific papers contribute narrower methodological evidence, with the discussion based on their published abstracts where full text was not examined. The central conclusion is practical: the strongest AI drug-discovery story is one in which the predicted advantage survives the relevant experiment and the resulting value can be traced through a clear economic right.

Primary documents and scientific references

• AbCellera · ABCL635 phase 2, August 10, 2026
• AbCellera · IMS presentation announcement, September 16
• AbCellera · Q2 accounts, August 5
• AbCellera · offering pricing, August 13
• AbCellera · current platform description
• AbCellera–Gilead · computational discovery stack, 2021
• AbCellera–Vertex · agreement, July 29, 2026
• AbCellera–Jazz · agreement, June 17
• Absci · HEADLINE, financing and Q2 accounts, August 11
• Absci · updated STORYLINE plan, September 24
• Recursion · Q2 development update, August 5
• Recursion · REC-4881 license, May 2020
• Recursion · January 12, 2026 filing
• Recursion–Tempus · data and model licenses, September 21
• Schrödinger · Q2 accounts and collaborations, August 5
• Schrödinger–Tectora · agreement, September 9
• Saunders, Comyn et al. · mAbs, 2026 · DOI 10.1080/19420862.2026.2692763
• Ren et al. · Nature Biotechnology, online 2024, volume 43, pages 63–75

Research cut-off: September 26, 2026. Company releases carried by Finviz are attributed to the issuing company. The two scientific papers were checked through their published abstracts; their findings do not establish efficacy of unrelated candidates.