AI has now guided compounds into clinical trials — that much is confirmed. Insilico Medicine's INS018_055 for idiopathic pulmonary fibrosis, Exscientia's early compounds, and a handful of others carry a credible "AI-designed" label, and the protein structure prediction revolution triggered by AlphaFold is simply not reversible. But here is the sharpest thing a small biotech team needs to hear before acting on any of this: not one fully AI-discovered molecule has cleared Phase 3 clinical trials and reached patients as an approved drug, and the teams betting their entire value proposition on AI speed have repeatedly discovered that the bottleneck was never computational. The good news — and it is genuine — is that the accessible, open-source tier of AI drug discovery tools has crossed a threshold where a two-person computational team can run hit identification campaigns that required a 20-person informatics group in 2020, at a fraction of the cost.
Derek Lowe's Science.org assessment of where the field actually stands surfaced on Hacker News with 129 points and 64 comments — unusually substantive engagement that cut through the standard marketing noise from AI biotech vendors. For small biotech teams, CRO startups, and computational chemistry freelancers, the implications are immediate: where to invest tool spend, which capabilities to build in-house, and how to avoid overpromising to investors who have heard the word "AI" attached to drug discovery for five years and are beginning to ask harder questions.
What Is This Actually?
"AI drug discovery" collapses at least four distinct problems into a single phrase, and that conflation is where most confusion begins. They have different tool requirements, different readiness levels, and very different cost curves for small teams.
Target identification is the problem of figuring out which biological molecule — usually a protein — is causally involved in a disease and represents a viable point of intervention. AI here means processing genomic, proteomic, transcriptomic, and clinical datasets at scale to surface associations that would take human researchers years to establish. Companies like Recursion Pharmaceuticals and BenevolentAI have built entire platforms around this problem. For small teams, this is the stage that requires the most proprietary data infrastructure — the open-source tools exist, but the value comes from the data, not the algorithm.
Structure prediction is now synonymous with AlphaFold. DeepMind's AlphaFold2 solved a 50-year grand challenge in 2021: predicting the three-dimensional structure of proteins from amino acid sequence alone, with accuracy that stunned structural biologists who had spent careers on single-protein X-ray crystallography projects. AlphaFold3, released in 2024, extended the approach to predict how proteins interact with other molecules — including small molecule drugs, DNA, RNA, and other proteins simultaneously. The European Bioinformatics Institute now hosts 200+ million predicted structures, freely accessible via browser or API. This is the single undisputed transformation AI has brought to drug discovery. It is permanent.
Lead generation is where most of the company-building activity has lived, and most of the hype. Given a protein target structure, can AI generate candidate molecules that are likely to bind it tightly and selectively? Generative chemistry models — including diffusion-based approaches, graph neural networks, and large language models trained on chemical space — can now propose novel molecules with reasonable predicted binding profiles in hours rather than months of wet lab synthesis-and-test cycles. Insilico Medicine's INS018_055 was designed using a generative AI pipeline, validated in animal models, and advanced into clinical trials. As of mid-2026, it is in Phase 2. That is the landmark example. There are others, but they are numbered in the dozens, not the hundreds.
ADMET prediction (absorption, distribution, metabolism, excretion, toxicity) is the unglamorous discipline of predicting whether a promising compound will actually behave like a drug in a human body. Most drug candidates fail not because they miss the target but because they're toxic, poorly absorbed, metabolized too quickly, or impossible to formulate. Machine learning models trained on historical ADMET data have improved substantially over the past four years. They still struggle meaningfully with genuinely novel chemical scaffolds outside the training distribution — which is exactly the space where generative AI tends to produce its most interesting candidates.
The timeline matters here. AlphaFold2's 2021 publication was the pivotal inflection. 2022–2023 saw the first wave of AI-designed compounds entering Phase 1 trials. By 2024, the vendor market had begun to consolidate — Exscientia was acquired by Sanofi, several AI-first drug companies quietly revised their timeline projections. Now, in 2026, early Phase 2 clinical data is starting to exist. That data, more than any benchmark score on ChEMBL, will determine whether AI drug discovery's promise was real.
Why This Matters Right Now
The timing of Lowe's analysis is not accidental. The industry is at an inflection point in 2026 that resembles the "trough of disillusionment" in the hype cycle — but with the crucial difference that real clinical data is now arriving, making the question answerable rather than deferrable.
Twelve months ago, there was still enough positive momentum from Phase 1 readouts that the hard question could be avoided: does any of this actually produce better drugs faster? In 2026, Phase 2 results are coming in, and they are mixed. Some AI-designed compounds have shown encouraging efficacy signals. Others have failed in ways that suggest the model was optimizing for the wrong objective — binding affinity without the context of selectivity, or potency without metabolic stability. The gap between "AI suggested this molecule" and "AI understood why this molecule would work" is becoming uncomfortable to ignore.
Several structural changes have converged to make this specific moment worth paying attention to. The cost of GPU compute has dropped enough that protein structure prediction and basic generative chemistry are now within reach of teams with a modest cloud budget. Running ColabFold on a single A100 in 2026 costs a few dollars per structure — that was not true in 2022. AlphaFold3 expanded to protein-ligand interactions, meaning structure-based drug design is now something a computational biologist with Python skills can engage with meaningfully, rather than requiring a specialist crystallography facility.
The regulatory environment has also begun to clarify. FDA guidance on AI/ML-assisted drug development, released in stages between 2023 and 2025, has started defining what AI evidence packages sponsors need to include with IND submissions. This matters for small teams because it shapes what you actually need to document from day one — and retrofitting regulatory documentation for AI workflows is considerably more painful than building the habit upfront.
And the open-source ecosystem has genuinely tipped. DeepChem's continued development, Chai-1's open weights from Chai Discovery, and the maturation of molecular dynamics tools like OpenMM mean that capability that was Schrödinger-exclusive in 2020 is now available to anyone who can configure a Python environment. What changed is not that AI suddenly got dramatically better at finding drugs. What changed is that the cost of entry dropped, the regulatory path got clearer, and the clinical data started arriving — simultaneously.
Practical Implications for Small Teams
This is where the analysis has to get specific, because "AI drug discovery" still sounds like something only Pfizer and DeepMind work on. That is no longer the case.
The two-person biotech startup doing target-based discovery. A founder with a biology PhD and a co-founder who can code is a real configuration in biotech today. With ColabFold or the EBI AlphaFold database, they can have high-confidence structural models of their target protein in hours. AutoDock Vina and RDKit together enable virtual screening of 10,000 commercially available compounds for predicted binding affinity in a weekend on a decent workstation. Compounds that score well can be ordered from vendors like Enamine for $50–$200 per sample and tested in an assay. This is a compressed hit-identification workflow that would have required a CRO contract costing $50,000–$200,000 in 2018. The caveat — and it cannot be soft-pedaled — is that docking scores are notoriously unreliable predictors of actual binding. False-positive rates can exceed 80%. What small teams gain is a prioritization tool, not a confirmation tool.
The computational chemistry freelancer. A freelance computational chemist who has historically served as an outside expert for early-stage biotech clients is now running into a specific competitive problem: clients have downloaded ColabFold and want to know why they should pay for expertise they think the tool provides. The correct answer — and the real value proposition — is that the tools represent half the equation, and often the smaller half. Knowing which protein structure to trust (AlphaFold is excellent for globular, well-folded proteins and genuinely unreliable for intrinsically disordered regions that are often the most interesting drug targets), how to prepare a protein structure correctly for docking (protonation states, cofactors, cryptic allosteric pockets), and how to interpret docking results in the context of structure-activity relationships — that expertise comes from years of practice and literature immersion, not from installing a package. The freelancer who articulates this distinction clearly will be fine. The one who competes on raw tool access will be commoditized by the tools themselves.
The small CRO adding AI services. Contract research organizations with five to twenty people are the segment with the most immediate opportunity and the most significant reputational risk. Adding AI-assisted virtual screening to a service menu is genuinely achievable with a DeepChem installation and one computational chemist. The danger is the framing: clients who hear "AI screening" consistently expect it to replace wet lab work. It does not. The CROs succeeding in this space are positioning AI virtual screening as a way to narrow a 100,000-compound virtual library to 50 candidates for synthesis and testing — a genuine and substantial cost reduction for clients, not a black box that finds drugs. That positioning requires educating clients rather than selling to their AI expectations, which is harder but more durable.
The digital health company eyeing drug discovery. Several digital health companies that built symptom checkers, patient data platforms, or clinical trial software in the 2021–2023 era are looking at drug discovery as a strategic extension. The logic is that their patient data represents an asset for target identification AI. This is a real possibility but a fragile one. Patient data collected for one purpose is rarely usable for drug target discovery without extensive cleaning, IRB considerations, re-consenting under new protocols, and often HIPAA-level data residency constraints. What tripped up many teams we've watched is treating the data as immediately usable when it required 18 months of infrastructure work before it could feed any model. The teams that have done this successfully treated it as a multi-year investment in data infrastructure, not a product feature they could ship in a quarter.
The academic spin-out built on protein structure prediction. The number of academic startups founded on some variant of protein structure prediction, protein-protein interaction modeling, or antibody design has grown substantially since 2022. The competitive question these teams face in 2026: what is their moat when AlphaFold3 is free, Chai-1 is open source, and every large pharma has already integrated structure prediction into standard workflows? Our read is that the moat has to be in proprietary experimental validation throughput, wet lab integration speed, or specific chemistry expertise for a narrow indication — not in model architecture. The teams trying to compete on prediction accuracy alone against Google DeepMind and academia's combined output are on the wrong battlefield.
How to Respond and Act on This
The framework for small teams deciding how to engage with AI drug discovery tools in 2026 should be: start with open-source, earn your way to proprietary, and never let the computational stage become the bottleneck in a process that is still governed by biology.
Step 1: Audit what you actually need AI for. The four stages — target identification, structure prediction, lead generation, ADMET — have very different tool requirements and cost curves. Most early-stage small teams only need structure prediction and virtual screening. Start there and resist the temptation to build a comprehensive platform before you have wet lab results to iterate against.
Step 2: Build on the open-source foundation. ColabFold handles the vast majority of protein structure prediction needs at zero cost. RDKit is the standard Python toolkit for molecular manipulation and property calculation. AutoDock Vina, or its deep learning-augmented successor GNINA, handles docking. DeepChem provides pre-trained ML models for ADMET and activity prediction. These four together cover the core computational pipeline at essentially zero license cost. The learning curve is real, but the documentation and community support are now extensive.
Step 3: Evaluate Schrödinger's FEP+ at the lead optimization stage only. If your team has moved past hit identification and needs to prioritize among 10–20 promising compounds for synthesis, Schrödinger's free energy perturbation calculations are the industry standard and the computational method most predictive of experimental binding affinity. The cost — roughly $15,000/year for academic use, considerably more for commercial licenses — is only justified once you have an active synthesis campaign ahead of you. Using it for early-stage exploration is spending precision tooling budget on a question that does not yet need precision.
Step 4: Use Chai-1 for commercial protein-ligand predictions. AlphaFold3's web server carries licensing restrictions that limit commercial use at scale. Chai-1 from Chai Discovery offers comparable protein-ligand and multi-chain structure prediction with open weights and a more permissive license, making it the better choice for commercial teams that want to run predictions in bulk without per-query fees or usage monitoring.
Step 5: Document AI workflows for regulatory purposes from day one. If there is any expectation of eventually submitting an IND or working with an FDA-regulated process, document your AI inputs, model versions, data sources, and decision rationale beginning with the first computational screen. The FDA's emerging guidance expects sponsors to explain what AI was used, where, and why. Retrofitting that documentation is exponentially more painful than building the habit when the science is being done.
Step 6: Be precise with stakeholders about what computational hits mean. The most consistent damage we observe in small biotech AI projects is investor and board communications that treat computational hits as validated hits. A compound that docks well is a hypothesis worth testing. Nothing more. Every computational result needs wet lab validation. The AI compresses the search space; the biology remains the sole arbiter of whether the hypothesis is correct.
Tool Comparison for Small Teams
| Tool | Best for | Free plan | Starting price | Key differentiator |
|---|---|---|---|---|
| ColabFold | Protein structure prediction; runs on Google Colab | Yes | Free | Fastest accessible AlphaFold implementation; massive user base |
| Chai-1 | Protein-ligand and multi-chain structure prediction | Yes (open weights) | Free | Commercial-friendly license; strong AF3 alternative for small teams |
| Schrödinger Suite (FEP+) | High-accuracy binding affinity, lead optimization | No | ~$15,000/yr academic; higher commercial | Physics-based hybrid; industry gold standard for lead optimization |
| AutoDock Vina / GNINA | Virtual screening, molecular docking | Yes | Free | Open source; GNINA adds deep learning scoring; widely reproducible |
| BioNeMo (NVIDIA) | Foundation model inference at scale for protein/molecule tasks | Limited trial | ~$500/mo cloud credits | GPU-accelerated; enterprise NIM deployment; integrates with HPC |
| RDKit | Cheminformatics, molecular property calculation, library preparation | Yes | Free | The standard Python library; underpins almost every other tool in the stack |
| DeepChem | ML models for ADMET prediction and activity modeling | Yes | Free | Widest range of pre-trained models; active open-source community |
For most small teams beginning to build a computational chemistry capability, the ColabFold + AutoDock Vina + RDKit + DeepChem stack is the right first year. The only cost is time. Schrödinger becomes relevant when an active synthesis and testing campaign is on the table. BioNeMo is for teams with HPC infrastructure and a defined need to run foundation model inference at scale — typically ten or more people with an established compute budget.
What the HN Community Is Saying
The Hacker News discussion around Lowe's piece drew a quality of commentary you don't find in general tech threads, because the topic self-selects for computational biologists, medicinal chemists, and ML practitioners who have actual skin in the game. Three distinct camps emerged.
The largest is what you might call the disappointed pragmatist. These are people with chemistry or computational biology backgrounds who think AI drug discovery has been systematically oversold but acknowledge real progress. The recurring argument: AlphaFold is genuinely revolutionary and deserves to be said plainly; the rest of the pipeline represents a 10–20% improvement on expensive, unreliable baseline methods, not a paradigm shift. Several commenters with industry backgrounds noted that the companies best positioned are using AI to triage and prioritize, not to replace medicinal chemists — and that the "AI-designed drug" framing is misleading because experienced chemists are deeply involved at every stage, iterating on AI-generated ideas rather than accepting them wholesale.
A second camp argues it is simply too early to judge. Drug development timelines run 10–15 years. AI tools have only been seriously deployed in IND-stage programs since roughly 2019–2020. The absence of approved AI-discovered drugs in 2026 is consistent with the timeline you would expect even if AI genuinely halved discovery time. Several commenters pointed out that the question will only be answerable in 2028–2030, when the first cohort of AI-designed compounds either clears or fails Phase 3 at scale. This is a legitimate point that the skeptics sometimes dismiss with what looks like premature closure.
A smaller thread included practitioners actively building with these tools and reporting genuine productivity gains. Academic researchers using ColabFold to shortcut what used to be months of structural biology. Computational chemists at small biotechs reporting that virtual screening pipelines now surface candidates that pass initial assays at meaningfully higher hit rates than their historical averages on similar campaigns. The signal is real; the magnitude is still being debated.
The concern that ran through almost every substantive thread was data quality. AI drug discovery models are only as good as the data they were trained on, and that data is largely proprietary to large pharma, biased toward certain target classes, and rarely reproducible. Several commenters noted that published benchmark results in AI drug discovery papers are systematically optimistic due to data leakage and the field's positive-results publication bias. One specific comment worth quoting in spirit: "The docking score is a prior. The assay is the posterior. Never confuse them."
Risks and Things to Watch
Vendor lock-in is underappreciated. Several AI drug discovery platforms charge per compound screened, per model inference, or through opaque enterprise SaaS pricing. If your hit identification work is embedded in a proprietary platform and you later need to migrate — or the vendor changes pricing after a funding round — you may find your computational history not exportable in any useful form. Strong preference for open file formats (SDF, PDB, SMILES, CSV) and open tooling reduces this risk. Any vendor who cannot export your compound library in a standard format should be treated with skepticism.
The benchmark problem is serious. Almost every AI drug discovery vendor pitch cites performance on standard datasets like ChEMBL, DUD-E, or BindingDB. These benchmarks carry known issues: temporal splitting problems (models trained on data that implicitly includes future compounds), heavy bias toward well-validated target classes, and significant gaps in novel chemical space. A vendor who quotes benchmark numbers without discussing these limitations is either not engaging in good faith or doesn't understand the limitations of their own evaluation. Either is a red flag.
Cloud compute costs can surprise. Running protein structure predictions or molecular dynamics simulations at scale on AWS or GCP can generate substantial bills quickly. A team that runs several thousand AlphaFold predictions or a long MD simulation without monitoring can see cloud costs of several thousand dollars accumulate over a weekend. Set budget alerts and cost caps before starting any compute-intensive workflow. This is the AI equivalent of leaving a cloud GPU instance running over a holiday.
The "AI-first" label as fundraising strategy. In the current biotech funding environment, "AI-first" in a pitch deck has meaningfully raised valuations. This creates pressure to overclaim. Our take is that investors are getting more sophisticated about asking specific questions: which model, trained on what data, with what prospective validation? Teams that can answer those questions clearly are in a genuinely differentiated position. Teams that use "AI" as marketing terminology without underlying substance are facing increasingly skeptical term sheets.
Regulatory clarity is improving but incomplete. FDA guidance on AI-assisted drug development is evolving. Different review divisions may have meaningfully different expectations for AI evidence packages, and the guidance does not yet cover every scenario a small biotech might encounter. Teams planning IND submissions should work with regulatory consultants who have specific AI experience — general pharma regulatory experience is insufficient for handling the nuances of AI model documentation requirements.
The attrition rate has not changed. This is perhaps the most important risk to name. The 90%+ failure rate of drug candidates in clinical trials is driven by disease complexity, patient heterogeneity, and biological unpredictability — not primarily by inadequate chemistry. AI can improve the quality of candidates entering the pipeline, and there is early evidence it does so somewhat. It cannot change the fundamental difficulty of predicting what happens inside a human being.
Frequently Asked Questions
Has any AI-discovered drug actually been approved?
As of mid-2026, no drug with a credible "primarily AI-discovered" label has received FDA or EMA approval. Insilico Medicine's INS018_055 is the furthest advanced, currently in Phase 2 clinical trials for idiopathic pulmonary fibrosis. Exscientia's early compounds reached Phase 1 before the company's acquisition by Sanofi. Several other AI-assisted compounds are in various trial stages across oncology, immunology, and fibrosis indications. Given the 10–15 year timeline for drug development, the first wave of fully AI-designed drugs would realistically reach approval somewhere between 2027 and 2032, assuming ongoing clinical success — and that timing would represent genuinely compressed discovery timelines compared to historical averages.
Can a small team actually do meaningful drug discovery with free tools?
Yes, with specific and honest caveats. ColabFold, AutoDock Vina, RDKit, and DeepChem together form a legitimate computational drug discovery pipeline, and a team with one strong computational chemist and one biologist can run hit identification campaigns against well-characterized protein targets at very low cost. What the free tools don't provide is the precision of free energy perturbation calculations for binding affinity, reliable prediction on novel scaffolds far outside training data, or the institutional knowledge embedded in commercial platforms. For target classes well-represented in open training data, the free stack is genuinely competitive with what a mid-size pharma company's informatics team could do in 2019. For genuinely novel targets or unusual chemical space, the capability gaps widen considerably.
How is AlphaFold3 different from AlphaFold2, and does it matter for drug discovery?
AlphaFold2 predicted protein structure from sequence with revolutionary accuracy for single chains and homo-multimers. AlphaFold3 extended prediction to how proteins interact with other molecules — small molecule ligands, DNA, RNA, and other proteins simultaneously, in a single model. For drug discovery, this shift is significant because you can now obtain a predicted structure of your target protein with a candidate ligand bound, rather than only an apo (unbound) structure that you then have to dock ligands into separately. The accuracy on protein-ligand interactions is not yet good enough to replace experimental crystallography or cryo-EM for lead optimization, but it is good enough to meaningfully guide virtual screening and prioritize candidates for more expensive experimental validation.
What is the realistic cost of running an AI-assisted hit discovery campaign?
A rough breakdown for a small team using open-source tools: ColabFold structure predictions run $10–$50 in cloud GPU time. A virtual screening library from Enamine REAL Space or a similar commercial supplier is free to screen computationally. AutoDock Vina or GNINA screening of 50,000 compounds takes a few hours on a workstation or $20–$100 in cloud compute. RDKit-based property filtering is essentially free. Ordering the top 50 compounds for experimental testing costs $2,500–$10,000 depending on vendor and compound complexity. Running a biochemical binding assay costs $5,000–$30,000 depending on the assay format and whether it's outsourced. A complete computational hit identification to confirmed biological hit cycle can run $10,000–$40,000 for a team that handles its own assays, compared to $100,000–$500,000 for a traditional approach. That compression is the real value proposition, and it is substantial.
What about antibody drug discovery — does AI help there too?
AI may be having an even more pronounced effect on antibody design than on small molecule discovery. Tools like RFdiffusion and ProteinMPNN can design novel antibody sequences with predicted binding to a target antigen, and companies like Absci have built generative platforms around this problem. The structural relationship between antibody sequence and binding function is better characterized than for small molecules, making it more learnable by current models. The limiting factors remain consistent with small molecule discovery: every designed antibody still requires extensive wet lab characterization, humanization for therapeutic use, and developability testing for manufacturing. But the design-test cycle has compressed substantially, and academic teams have benefited from open tools here perhaps more than anywhere else in drug discovery.
Should a small team build internal computational capability or use a CRO with AI services?
The decision depends entirely on whether computational expertise is core to your scientific differentiation or a means to an end. If your competitive advantage is a novel biological hypothesis about a target, outsourcing the computational work to a specialist CRO and concentrating resources on the biology is often the more capital-efficient path. If your competitive advantage is fundamentally computational — a proprietary model, a unique dataset, a better prediction approach — then building in-house is necessary because your capability cannot be recreated by a CRO with standard tooling. The mistake is building internal computational capacity purely as a cost-saving measure when the team doesn't have the expertise to critically evaluate the outputs.
How do I evaluate an AI drug discovery vendor's claims?
Ask five specific questions. First: what dataset was the model trained on, and does it cover your specific target class? Second: what is the temporal split in your benchmarks — were test compounds synthesized after the training cutoff? Third: do you have prospective validation data, meaning compounds your model designed that were then synthesized and experimentally tested? Fourth: what happens to my data — do you train on customer compounds or targets? Fifth: how does your model perform on scaffolds significantly outside the training distribution? Vendors who answer these confidently and with specifics are operating in a different register from those who respond with generalized benchmark claims or redirect to case studies from other customers.
Final Verdict
The honest assessment in mid-2026 is that AI drug discovery has earned one unambiguous and permanent win (protein structure prediction), demonstrated credible but not yet definitively proven value in lead generation, and is still working through whether the complete system produces better clinical outcomes than traditional methods. That is not failure — it is exactly where a serious scientific field should be after five to six years of genuine production deployment.
For small biotech teams and computational chemistry freelancers, the practical implications are clear. The open-source tier is now genuinely good enough to run meaningful hit identification campaigns at a fraction of traditional costs. The expertise required to apply these tools well — understanding their failure modes, preparing inputs correctly, interpreting outputs in biological and chemical context — is real, not easily replaced by the tools' availability, and not yet commoditized. That is where the durable value lives for practitioners.
The teams that should act now are those with a defined protein target, some wet lab capacity to validate computational hits, and a computational chemist on staff or available as a contractor who can run and interpret results. The return on the ColabFold + AutoDock Vina + RDKit stack, in early-stage hit identification costs avoided, is almost certainly positive for any team that has a synthesis and testing budget to work with. The tools have crossed the threshold from specialist curiosity to standard practice.
The teams that should proceed carefully, or significantly adjust their positioning, are those planning to sell "AI drug discovery" as a service without deep medicinal chemistry expertise to back it up. The tools are accessible enough that clients will eventually distinguish between getting tool access and getting scientific judgment. That distinction is becoming clearer in the market every quarter.
The broader signal from Lowe's analysis and the HN discussion combined is that the field is entering its credibility test. The next two to three years of Phase 2 and Phase 3 clinical data will separate the tools and approaches that genuinely improved drug quality from those that moved faster to the same clinical failures. Small teams that have been using AI drug discovery tools honestly — to prioritize, to generate hypotheses, to compress search space — will likely emerge from this period with validated, reproducible workflows. Teams that have used AI primarily as a narrative for investors will face increasingly uncomfortable questions as clinical data accumulates.
One thing that seems beyond dispute: the structural biology revolution triggered by AlphaFold is not going away, and it has already permanently changed what a small team can accomplish. If your scientific work involves proteins — and nearly all serious drug discovery does — the free tools available today represent some of the most consequential scientific software released this decade. Engaging with them seriously is no longer an advanced option. It is the baseline.